I sawthis thread yesterday and I wanted to chime in this idea I had, but I waited to long to act on it and now it's archived. So I guess I'll have to make a new thread.

It's clear that lots of new editors struggle making good content with AI assistance, and something has to be done.WP:G15 is already a good start, but I think restrictions can be extended further.Extended confirmation on Wikipedia is already somewhat of a benchmark to qualify editors to edit contentious articles, and I think the same criteria would do well to stop the worst AI slop from infecting mainspace. As for how this would be implemented, I'm not sure - a policy would allow human intervention, but a bot designed like ClueBot NG might automate the process if someone knows how to build one.Koopinator (talk)10:50, 18 October 2025 (UTC)[reply]

I do t see a practical way to enforce that. I also dont think that peoples skill level with AI can transfer to an assessment of their skill level in wikipedia. —TheDJ (talk •contribs)11:31, 18 October 2025 (UTC)[reply]

Regarding enforcement, I would suggest:

1. Looking at whatever process ClueBot uses to detect and evaluate new edits, and add a "extended confirmed/non-ec" clause.

1.1. I will admit I'm not entirely sure of how this would work on a technical level, which is why I posted this idea in the idea lab.

2. Look to word frequency as inUser:Gnomingstuff/AI experiment to distinguish AI from non-AI edits.Koopinator (talk)15:32, 18 October 2025 (UTC)[reply]

please don't use this in any kind of blocking enforcement capacity, it is not remotely ready for anything like thatGnomingstuff (talk)17:41, 20 October 2025 (UTC)[reply]

A person'swillingness to use AI on Wikipedia is an immediate and absoluteWP:NOTHERE, in my opinion.TooManyFingers (talk)05:50, 4 November 2025 (UTC)[reply]

Too sweeping an opinion in my opinion. First you would have to be talking about specifically using unsupervised AI to write articles. Secondly I think it would be "insistance" rather than "willingness". And thirdly it could well be aWP:CIR or user education issue rather than a NOTHERE one. All the best:RichFarmbrough18:03, 6 November 2025 (UTC).[reply]

Do you have any evidence that extended confirmed users create any better edits with AI than users who are not extended confirmed?Phil Bridger (talk)14:33, 18 October 2025 (UTC)[reply]

I would say it's a reasonable inference. Here's what I can say:

• We can expect that extended-confirmed users are more likely to be familiar with Wikipedia's policies and guidelines, by virtue of having been here longer.
  • This expectation is, in part, substantiated by thatmost vandalism comes from IP editors.
• Some anecdotal evidence:
  • [1] LLM edit with no sources, survived for almost 2 months. Was created by an editor who was neither confirmed nor extended confirmed.
  • [2] Personal project by yours truly, AI assistance was used, careful review of text-source integrity of every sentence as I constructed the page in my sandbox over the course of 59 days before airing it.
• I admit none of this is hard evidence.

I do feel LLM has its place on the site (otherwise I wouldn't have used ChatGPT assistance in constructing a page), but if it's allowed, the barrier for usage really should be heightened. Wikipedia's content translation tool is also restricted to extended-confirmed users.

Koopinator (talk)15:25, 18 October 2025 (UTC)[reply]

The issue is raising the bar to prevent bots from editing Wikipedia using LLMs.LDW5432 (talk)19:57, 27 October 2025 (UTC)[reply]

LLM detection for text is very hard and has far, far too many false positives, especially for non-native speakers and certain wavelengths of autism.Aaron Liu (talk)16:41, 18 October 2025 (UTC)[reply]

^ This is my experience. Also, a lot of edits are too brief for the already-dodgy AI "detectors" to be reliable for.

@Koopinator, you've made around 2,000 mainspace edits in the last ~2 years. Here's a complete list of all your edits that the visual editor could detect as being more than a handful of words added.[3] It's 78 edits (4% of your edits) – less than once a week on average. And I'd guess that half of your content additions are too short to have any chance of using an anti-AI tool on, so the anti-AI tool would check your edits two or three times a month. Why build something, if it could only be useful so rarely?WhatamIdoing (talk)00:58, 19 October 2025 (UTC)[reply]

Well, how would that tool's frequency scale across the entire Wikipedia community? I'd imagine it'd be used at least a little bit more often then. (or, I imagine, multiple orders of magnitude)Koopinator (talk)05:55, 19 October 2025 (UTC)[reply]

For brand-new editors, it might capture somethingon the order of half of mainspace edits. High-volume editors are much more likely to edit without adding any content, so it'd be much less useful for that group.WhatamIdoing (talk)19:54, 23 October 2025 (UTC)[reply]

We could at least use a flagging system for vandalism review.LDW5432 (talk)14:05, 6 November 2025 (UTC)[reply]

It should be possible to detect low hanging fruit AI text, based on certain common features. Raw AI inference cut and pasted from a chat bot is going to be easier to detect. I agree that the type of user doing this probably has no reputation at stake, doesn't care very much, more likely to be newbie and/or a non-native speaker from another Wiki. I don't know about policy, but a bot that sends a talk page notice, or flags the edit summary with a "[possible ai]" tag. No one is already working on this? --GreenC17:10, 18 October 2025 (UTC)[reply]

mw:Edit check/Tone Check uses aSmall language model to detect promotionalism. (See tagged edits.) I'd guess that it would be possible to add an AI detector to that, though the volume involved would mean the WMF would need to host their own or pay for a corporate license and address the privacy problems.

mw:Edit check/Paste Check is probably more efficient, though, as anyone copying from a chatbot is going to be pasting it into the article, and detecting a big paste is easier than checking the words that were pasted in.WhatamIdoing (talk)01:04, 19 October 2025 (UTC)[reply]

I think AI edits should be mandatory for everyone to disclose, both in articles and talk pages. There could be a box where you check it if your content comes from AI or is mostly AI, similar to how you can check minor edits.Bogazicili (talk)18:40, 21 October 2025 (UTC)[reply]

Having a UI element like that would work towards legitimizing LLM use in creating text for Wikipedia.Merko (talk)00:41, 22 October 2025 (UTC)[reply]

I agree: Either it will allow the material to be posted and thus legitimize LLM use, or it won't allow the material to be posted and cause people to tell lies so they can get it posted.WhatamIdoing (talk)02:18, 22 October 2025 (UTC)[reply]

Do we currently have a policy on LLM usage? This one seems failedWikipedia:Large language model policy

My position is that if it's not banned, it should be declared.Bogazicili (talk)10:45, 23 October 2025 (UTC)[reply]

I thought the failed policy proposal was supposed to require people to declare it.WhatamIdoing (talk)20:00, 23 October 2025 (UTC)[reply]

Almost 2 years ago.Merko (talk)22:09, 23 October 2025 (UTC)[reply]

LLM-generated content is a cancer on Wikipedia, and it will only get worse. "AI detectors" have many false positives, as do checks made by editors themselves, but just because we can't reliably detect something today doesn't mean we shouldn't implement a policy against it. I support mandating the disclosure of LLM-generated contributions by all users. We don't treatWP:GNG differently on articles created by extended-confirmed users or others, we shouldn't do it here either.Merko (talk)22:21, 21 October 2025 (UTC)[reply]

If you think original content generated by a program is a negative to that extent, then I don't think requiring disclosure is the appropriate approach, since that would only be a prelude to removal. We should skip straight to requiring editors not to use programs to generate original content.isaacl (talk)04:38, 22 October 2025 (UTC)[reply]

Wikipedia should first address LLM content from anonymous IPs.LDW5432 (talk)19:56, 27 October 2025 (UTC)[reply]

IP editing actually isn't that much of a problem here -- in my experience almost all AI text I find came from someone with a registered account. Off the top of my head I'd say less than 10% of it comes from IPs.

This may change with temporary accounts in a few days though, who knows.Gnomingstuff (talk)20:56, 30 October 2025 (UTC)[reply]

I came here to propose pretty much the same thing (policy, not bot). Having a blanket rule would be hugely helpful in dealing with editors, since it can get very tedious explaining why each AI edit they claim to have checked is in fact problematic. I might even go so far as to propose a separate user right (or pseudo-right?) called something like LLM user, for editors who can demonstrate they are sufficiently competent with content policies and have a legitimate use case. I don't think such a right should convey any actual abilities, but users found to be using LLMs without it could then be much more easily censured and guided towards other forms of editing. Applying exactly the same system but tying it to extended confirmation seems like it minimizes potential rule creep, but it's a blunter filter which might not be as effective, since I'm sure there are plenty of extended confirmed users who lack the requisite understanding of policy.lp0 on fire ()21:03, 10 November 2025 (UTC)[reply]

That is probably a good idea, but I don't see any way to enforce it automatically and also do it well, as it would not be good if someone got flagged for using AI when they did not, and Wikipedia is so large it would happen a lot. I believe that AI should be used extremely rarely on Wikipedia, as it is known to hallucinate mis-information and drag on and on about things that don't matter (see: Grokapedia, or search up AI hallucinations). It has many chances to cause things to go awry, and should not be made main-stream as a way to enhance/speed up editing. I suggest it is done by humans. If a new user joins Wikipedia and is flagged or seen on talk pages, maybe give there edits a look, just to make sure there doing good. Some ways to spot AI writing is looking for constant pairs of 3's (like, LOTS, basically every sentence), un-usual use of Em dashes,(looks like a bigger hyphen, — Vs. -) as they are not on a normal keyboard and either take a copy and paste or a very unique keyboard shortcut to type, repeating info or full paragraphs that don't really say/mean anything. A lot of these are hard to give examples for and you just have to see them for the first time to start noticing. Overall, I agree that there should be restrictions on AI edits.Oak lod (talk)15:49, 20 November 2025 (UTC)[reply]

I strongly support the suggestion and would even go as far as suggesting a new flag. The AI as a tool is similar toWP:AWB: in unskilled or malicious hands it can do a lot of damage in a short amount of time. Correspondingly, use of AWB is not allowed for drive-by accounts. Similar logic applies to AI, IMHO. For the avoidance of doubt, I think that proper use of AI improves articles, so I think that we should regulate the use of AI, and not prohibit it. Fear of outright hallucination is overblown, as far I can tell: as long as the input wasexplicitly restricted to correct sources (either a foreign-language Wikipedia article or manually-selectedWP:RS), there were no hallucinations. Note that texts of RS you are planning to use for the article should be fed to the engine first in their entirety, as for some reason the AI engines are really shy when it comes to actually fetching information off the Web (I suspect there are legal reasons in play here), so if you just point to the sources, AI will start generating ideas of its own, not summarizing theWP:RS as it should.Викидим (talk)00:14, 24 November 2025 (UTC)[reply]

What if we make a box that allows people to flag their own edits as AI-assisted, and a warning that lets people know that fully AI-generated content will be taken down in accordance with a policy and partially AI-assisted content must be marked so that humans can review it or it will be taken down if not marked. (if there's not a policy to ban unreviewed AI text already, make one). Then, we make a bot like Cluebot to detect AI slop and revert it and leave a warning, but we have it set to be very cautious so it minimizes false positives. I think this would solve the problem and it neatly combines all the ideas I saw above.RBarr-12@wiki:~/user/talk/contribs$ 20:07, 2 December 2025 (UTC)[reply]

That's probably the best solution. Good idea.

Oak lod (talk)20:14, 2 December 2025 (UTC)[reply]

It won't work.

First,Wikipedia:Nobody reads the directions. Then, if someone does manage to see the checkbox, they'll check it ...and check back, and if their edit has been reverted, they will never check it again. We have evidence of this in the in-editor image uploading tools. If people believe it's reasonable to upload a corporate logo (or some other common type of image), then they'll tick whatever box you require. Sure, I own the copyright to the McDonald's logo. Sure, I wrote all that myself. Sure, I'll give my first born toRumpelstiltskin. Whatever is necessary to do the task, people will claim they've done.WhatamIdoing (talk)02:02, 5 December 2025 (UTC)[reply]

Man. I guess the simplest, easiest, and first solution you think of really isnever the best solution.

Oak lod (talk)15:54, 5 December 2025 (UTC)[reply]

Good point. Maybe simplify to just the bot checking for AI content, warning editors. Basically, a Cluebot clone for AI detection.RBarr-12@wiki:~/user/talk/contribs$ 17:32, 5 December 2025 (UTC)[reply]

The problem is that "AI content" is nebulous and hard to define, any automated tagging will either include false positives or so many false negatives as to be useless (or both). Any edits flagged for being AI will include some that have no problems at all, some that have issues that can be trivially fixed, and some that have actually serious issues. These issues will be a mix of all-types making it harder to fix (e.g. it will mix non-existent references due to minor errors in with hallucinated references, text that includes prompts for the user and other problems).Thryduulf (talk)18:45, 5 December 2025 (UTC)[reply]

Honestly, the best way to find and fix issues regarding AI content would most likely to just have no specific bot for catching these edits. Might be best to just use the hundreds of thousands of editors already looking for errors in pages instead, as the best detector of not human content is a human.Oak lod (talk)15:52, 8 December 2025 (UTC)[reply]

Which is why when someone finds an error they can't (for whatever reason) fix it themselves there and then, they should be encouraged to tag it with what the specific problem is (failed verification, inappropriate tone, etc) rather than a generic AI tag. Being specific about what the problem is means others don't have to spend time figuring it out (what's obvious to one person isn't necessarily obvious to someone else) and those editors who are looking to fix problems of a given type know that such a problem exists in that article.Thryduulf (talk)16:18, 8 December 2025 (UTC)[reply]

The problem is that there are too many people willing to use AI, and AI can compose long stretches of almost-plausible ideas. There's more inaccuracies being produced than edits people are willing to spend their time on reverting said inaccuracies instead of on their new draft paper. We can't control the people making the edits until after they've made those edits, so we either need more editors or some assistance to make the editors we do have more efficient somehow. I don't know what that assistance would look like though.RBarr-12@wiki:~/user/talk/contribs$ 17:44, 8 December 2025 (UTC)[reply]

I'm not understanding how that follows on from my comment?Thryduulf (talk)18:01, 8 December 2025 (UTC)[reply]

Accidental reply to your comment, meant to reply to Oak Iod's comment above.RBarr-12@wiki:~/user/talk/contribs$ 18:52, 8 December 2025 (UTC)[reply]

We could use those specific tags to train AI detecting models. The edits with those tags could be validated if plausible or just straight added to the training data. After the model has been trained enough and proves to be accurate, we could do a test run on Wikipedia, with the bot only tagging edits and not reverting them. People could then remove the tag to indicate a false positive and the amounts of tags from the bot that were removed and the ones that stayed could be counted up. If there are too many false positives then the bot is scrapped or re-trained. This could be repeated as many times as found necessary. The bot could also be built upon already existing bots. This would be a little far fetched, as the issue is not that large, Wikipedia might not be built for a system like that, it could be very expensive, and Wikipedia hasn't done anything like this before as far as I am aware. This also might bring up similar issues to other solutions.

I also think that RBarr-12 intended to reply to my comment.Oak lod (talk)18:14, 8 December 2025 (UTC)[reply]

IDK about the technical feasibility of scanning all edits with a bot, but the policy side of this is justWP:LLMDISCLOSE. --LWG^talk20:42, 2 December 2025 (UTC)[reply]

Here's yet another AI edit that survived for years:Special:Diff/1145967029. I only noticed this one when some user decided to add an{{AI-generated}} tag to it, in 2025, and the edit was made in 2023. The user, by the way, had 144 edits and a history of creating AI-looking pages likethis one.Somepinkdude (talk)14:47, 13 December 2025 (UTC)[reply]

I feel as though it should be the case that for those who have minimal experience in editing, like me, should not be allowed to utilize AI to make edits. This is something that needs to be considered: AI cannot help you unless you already know what you are doing. I can't have AI write code for me if I don't have the skills necessary to interoperate, give the correct input into the AI, know when the AI is lying. Also sources must be considered. Most of the content on here is facts, that must be backed up by sources. It should be mandatory that research is done, and that facts never originate from AI, rather AI should only be allowed to be used as a phrasing tool. Because I get it, its quite annoying to write super long sections, you have to go look stuff up, then cite, then you go back to writing for a sentence and then it loops, and by the end you are just left with an incomprehensible, dense as hell paragraph that is useless because people can't understand it. But if the facts that the writer already collected are then told be phrased by an AI assistant, then that should be allowed, it would aid in comprehensibility.CatLove989 (talk)16:59, 14 December 2025 (UTC)[reply]

When reading articles about geographic locations in desktop mode, I am slightly annoyed if the coordinates are not available in a convenient and predictable spot near the article title. This forces me to hunt for the coordinates in the infobox or article body. It also means that the article will not be correctly geotagged.

For some examples of articles that have this issue, due to using{{coord}} with|display=inline alone, seeYerevan,Matera,Duluth, Minnesota,San Luis Potosí (city), andShivneri Fort. Also note, for example, that Shivneri Fort will not show up when viewingSpecial:Nearby#/coord/19.199,73.8595.

Conversely, when browsing on mobile, coordinates added using|display=title alone aren't visible at all. For some examples of articles with this issue, seeIslandmagee,Ostia (Rome), andMatthias Church.

To avoid both of these problems, I would tentatively propose that|display=inline,title should be preferred in most^* articles about settlements or geographic features. It seems that it would be possible to use a bot or semi-automated script to enforce this rule.

Perhaps my proposal is already the accepted approach and the articles above have just unintentionally deviated from it, but I'm not sure.MOS:COORDS doesn't really seem to address this issue and I couldn't find any other relevant guideline. This issue has probably been discussed before; links to past threads would be appreciated.

*There are obviously cases where|display=inline is appropriate. For example, the articleExtreme points of the United Kingdom discusses several different points and it would be wrong to geotag the entire topic to any specific one. There are likely other edge cases I haven't thought of. I'm only referring to how to format the "main coordinates" in articles about uniquely identifiable locations: villages, mountains, buildings, etc.~2025-32085-07 (talk)23:36, 9 November 2025 (UTC)[reply]

Hello. In my opinion, the title is a goofy spot for coords and we should list them only in the infobox alongside all the related metadata about a place. It's a weird historical artifact and anachronism that the coords get such special placement and their special page placement has been a constant headache for years with different views and different skins, as you note. Is there a reason coords are so special that they can't be put in the infobox? The coords seem as relevant toPittsburgh as its population. --MZMcBride (talk)20:47, 10 November 2025 (UTC)[reply]

Coordinates are still somewhat “special” in that they link to an external tool. However I personally don’t think that’s reason enough to separate them.  novovtalkedits00:02, 12 November 2025 (UTC)[reply]

They don't require this, we make a choice (we can also show them with the built in maps, but it's difficult to change something that has been around for as long as this. They are mostly special, in that they have to directly relate to the primary topic of the page and the page has to detail a specific spot that is not too large or otherwise vague. —TheDJ (talk •contribs)11:33, 13 November 2025 (UTC)[reply]

I would argue that a city's coordinates are a more defining property than its population. Population numbers change over time, coordinates generally don't. As for what's of greater value to readers, IDK.

Personally speaking, I find myself clicking coordinate links very frequently. The ability to view a location on a map is immensely useful. Even for articles that include a locator map image or embedded "Wikimedia Map", I find GeoHack useful because of the links it provides to external services.

Something else I'll mention, but which probably deserves its own discussion, is that WikiMiniAtlas now seems redundant to Wikimedia Maps. WikiMiniAtlas was great for its time but its design now feels outdated. The aesthetic recalls the early days of Web 2.0, there's no support forpinch to zoom, etc. The one area where WikiMiniAtlas shines is that it does provide links to other nearby articles. I'll admit that's a pretty major feature, arguably even the main feature.

(Also, is it just my imagination or is WMA's projection extremely distorted? WMA always seems to be stretched out along the east-west axis. Compare Iceland onWMA vs.OSM.)~2025-32085-07 (talk)07:10, 19 November 2025 (UTC)[reply]

Coordinates dochange over time if you give it enough time. 😀Anomie⚔12:57, 19 November 2025 (UTC)[reply]

also wondering myself how people even find coordinates. I had to remove some from a page recently for being totally wrong. ←Metallurgist (talk)04:51, 13 November 2025 (UTC)[reply]

I've also occasionally come across incorrect coordinates in Wikipedia articles. At least in the cases I've seen, the mixups sometimes arise when multiple nearby localities have similar names.~2025-32085-07 (talk)07:35, 19 November 2025 (UTC)[reply]

I've pointed this out on a few talk pages, but generally when it comes to coordinates, maps, and stuff like that all Wikipedia MOS goes out the window. Having coordinates without a source is original research. Having a locator map without a source for the boundaries is original research. There is almost no quality control, and rather rather then removing inaccurate or poorly sourced maps/geographic information, people argue they should be left until someone offers a better one. Really a huge issue, as a cartographer I'm a bit appalled.GeogSage ^{(⚔Chat?⚔)}07:49, 19 November 2025 (UTC)[reply]

Wikipedia:No original research defines original research as material for which the real word doesn't have a source saying that, which is importantly different from material for which the Wikipedia article doesn't cite an article.WhatamIdoing (talk)18:41, 19 November 2025 (UTC)[reply]

Having a boundary file without a citation is like a direct quote without attribution. There are several maps where the boundaries are user generated, or appear to be, and people grab coordinates for places from a map but don't have a source verifying that those are the actual coordinates. Going onto Google Earth, grabbing a bunch of points, and making a map that says those points are the locations of _______ is OR. Boundaries are often challenged by official organizations, stating "This is where the border for ____ is" without stating where we got that information would not be acceptable in text.GeogSage ^{(⚔Chat?⚔)}02:55, 26 November 2025 (UTC)[reply]

I'm not disagreeing that it's bad to have these uncited. I'm disagreeing that an accurate copy of the boundary (i.e., one that matches what's in at least one published reliable source) is specifically the kind of bad that we calloriginal research. It's all kinds of badexcept that one (unless it's completely made up, of course).WhatamIdoing (talk)02:49, 5 December 2025 (UTC)[reply]

Even if it doesn't match, it is notnecessarily original research. For example if I draw the boundary of Example Island'sEEZ extending 200 miles (320 km) from the low water mark, but it turns out the boundary is actually 200 nautical miles (370 km; 230 mi), that is wrong but it is not necessarilyoriginal research. Itmight be but it is equally possible that I consulted an unreliable source, or an outdated source. It's also possible my source was reliable but I made a mistake with my drawing.Thryduulf (talk)03:29, 5 December 2025 (UTC)[reply]

Recently I have learned that there is an International Mentoring Day on 17 January. The UK and the US also have national commemorations to celebrate mentoring and thank mentors of all sorts (i.e. in corporate mentoring programmes; adult-led youth groups; and teaching). In the UK, this is 27 October; in the US, the entire month of January.

With this in mind, I would like to propose that Wikipedia:

• Start an annual commemoration on January 17 of this coming year with notification about the day somewhat in advance, and encouragement to all editors to take a few minutes to thank their mentors whether current or past, as well as those who offer guidance as Teahouse, Help Desk, and Village Pump staff;

• Share stories about how mentoring helped; and

• Offer "Did You Know?" tidbits around and on January 17 about how the commemorations came about in the UK and the US.

As we are a little over 9 weeks away from January 17, there would be adequate time to plan for its commemoration on Wikipedia if the decision is taken to carry this idea forward.~2025-33078-41 (talk)17:52, 12 November 2025 (UTC)[reply]

The problem with days of X is that anyone can declare any day the day of X and these things die after a year or two when a few people forget about them.

Also I haven't really seen much active mentoring on Wikipedia, but that can be my fault because it is not the kinda thing I would notice.Polygnotus (talk)03:42, 20 November 2025 (UTC)[reply]

There really is anInternational Mentoring Day on 17 January. It was started as an extension of the US National Mentoring Month (held throughout the month of January), but is now encouraged worldwide.

Because mentorship is an important part of Wikipedia for many editors, it just seems like promoting the day would be a wonderful way to honor those who serve in this way.

Do you have any idea where else in the world of Wikipedia that this suggestion could be raised with greater likelihood of taking it further?~2025-36716-26 (talk)10:13, 27 November 2025 (UTC)[reply]

No clue, sorry.Polygnotus (talk)10:32, 27 November 2025 (UTC)[reply]

I think I have just found what seems a good step to move forward with this idea: to make a "Central Notice banner request."~2025-37075-42 (talk)16:54, 28 November 2025 (UTC)[reply]

Central Notice banners are rarely used and for fully fleshed out ideas with consensus behind them that have been implemented already.

So far you reached one person, and they were not enthusiastic about the idea.

Is there a reason you would like to push this, which could include but is not limited to being involved with the people/an organization who/which decided to give that day that label or who/which joined the initiative?Polygnotus (talk)17:07, 28 November 2025 (UTC)[reply]

The only reason I would like to "push this," Polygnotus, is because of the wonderful guidance I've received from my own mentor, as well as many other knowledgeable editors who staff Wikipedia help venues ... and the immense appreciation I've come to feel for volunteering their time and effort.

No, I'm not at all involved with any of the people or organizations who created or joined the International Mentoring Day initiative. It was only at some point this year that I even heard of such a day.~2025-39632-68 (talk)11:59, 9 December 2025 (UTC)[reply]

Maybe try theTeahouse?Polygnotus (talk)12:00, 9 December 2025 (UTC)[reply]

One of my hobbyhorses here is cleaning up promotional articles, particularly of BLPs. One tell-tale sign I see frequently is an overstuffed 'Awards and recognition' or 'Awards' section, full of prizes no one has ever heard of given out by obscure webmagazines or societies. However, similar sections are often created or added to by good-faith editors, and sometimes BLPsshould mention genuinely notable awards. As far as I know, there's no clear policy on these sorts of things beyond our general policies on avoiding puffery, overdetail, and trivia. This has occasionally led to editing conflicts.

I've been trying to think through a policy which could help us deal with these issues systematically. I think there are two key thing that might help:

• Awards granted to BLPs should be mentioned only if the award is itself notable (such as aNobel Prize or aIET Faraday Medal)
• Except in exceptional circumstances, we should not allow standalone 'Awards and recognition' sections (similarly to how we like to avoid 'Criticism' sections). Mention of awards received should be distributed throughout the text in a sensible way, typically chronologically.

I do worry that for academics, there exist non-notable awards that are nevertheless relevant to summarizing someone's career - these things matter in academia but a lot of the prizes are pretty obscure. We might also consider mentioning awards given by notable organizations if those awards are mentioned in the org's article. Any thoughts on these suggestions? Improvements? —Ganesha811 (talk)00:16, 20 November 2025 (UTC)[reply]

I think if an award received has received coverage in a secondary source, then that's another good reason to include the award in the Wikipedia article, regardless of whether or not that particular award received is notable. Say Sally Willis receives the Jon Brandt Award for Excellence in Journalism and the Jon Brandt award is not a notable award, but in a profile of Sally Willis,The New York Times lists that award amongst her accolades, I think that would be a good reason to include the award. Or perhaps Sally Willis lives inAthens, Ohio and local pressThe Athens Recorder runs a story on Sally Willis receiving this non-notable award because Sally Willis is the most notable person from Athens and everyone there is super proud of her accomplishments. I think that would be another good reason to include an award in an article. I think a good start to cutting out awards is to exclude the non-notable ones that are only mentioned on the recipient's CV / other personal website and sources from the body that bestows the award (e.g. website, award ceremony documents, etc).Katzrockso (talk)00:27, 20 November 2025 (UTC)[reply]

We could make lists of awards we consider worth mentioning, like RSN. We can also make a list of fake awards that should definitely be removed. I started one over atUser:Polygnotus/vanity. There are at least some awards that are notable and have an article, but are not worth mentioning (for exampleSuperbrands). Another complication with requiring articles is that you can require a standalone article about the specific award, or an article about the organisation behind it. Awards and recognition' sections can make sense in cases likeQuentin Tarantino who won like 4 trillion awards. See alsoList of awards and nominations received by Quentin Tarantino. Maybe an article should only be allowed to have a dedicated section for awards if you reach a certain threshold, like 10+ notable ones or if they have their own article.Polygnotus (talk)03:38, 20 November 2025 (UTC)[reply]

 Comment: Way to much policy creep. Many of the major awards in my discipline barely have a presence on Wikipedia. I've gone through the effort to get some content for the bigger ones, but unless someone interested in the topic also thinks to make a Wikipedia page for it, they will slide through the cracks. If an outside source states the award was given, and the source is reliable, why would we default to excluding it from the article?GeogSage ^{(⚔Chat?⚔)}07:07, 20 November 2025 (UTC)[reply]

@GeogSage I agree that if a truly reliable and independently written source thinks its worth mentioning then it is most likely worth including. The problem is that a lot of these claims do not have a reliable source attached, and often not even a source at all.Polygnotus (talk)07:19, 20 November 2025 (UTC)[reply]

Wikipedia:Biographies_of_living_persons#Reliable_sources: "Wikipedia's sourcing policy, Verifiability, says that all quotations and any material challenged or likely to be challenged must be attributed to a reliable, published source using an inline citation; material not meeting this standard may be removed." You could always tag^{[citation needed][according to whom?][additional citation(s) needed][promotional source?]} if you doubt it. I write a few biographies for academics, and I try to include an award section if applicable. Generally, getting the citation isn't hard if you know they got the award, the most extensive I've done was forWaldo R. Tobler so I'll use him as an example. Some, like the Andrew McNally Award, 1986, might not have made the transition to the digital realm but are mentioned in sources discussing Tobler. In another biography I'm working on right now (not of a living person), the award was won in 1947, and I'm not even sure the awarding organization is still around. It is noted in multiple peer-reviewed publications discussing the subject though. I feel like if you see an award that isn't sourced, you can try to find it online. If you can't find a source, you can tag it or delete it with an edit summary. I don't think we need to get more complicated then that about what counts for inclusion.GeogSage ^{(⚔Chat?⚔)}07:36, 20 November 2025 (UTC)[reply]

I know for film articles, to avoid overstuffing, we only include awards that have articles here. I see no reason why the same guideline couldn't be reasonably applied to BLPs. If one feels an award is notable enough to merit inclusion but it lacks an article, they can certainly undertake the effort to write the article at that point.DonIago (talk)07:23, 20 November 2025 (UTC)[reply]

Not a lot of the big academic awards have Wikipedia pages. The biggest award in American Geography is the Anderson medal of honor, and it is mentioned on theAmerican Association of Geographers page briefly. If we limited it to only awards on the AAG page, most of the ones the AAG issues couldn't be included.GeogSage ^{(⚔Chat?⚔)}07:39, 20 November 2025 (UTC)[reply]

@GeogSage I think a section in a larger article, or a standalone article, is both fine. I redirectedAnderson medal andAnderson Medal to the appropriate section.Polygnotus (talk)07:49, 20 November 2025 (UTC)[reply]

That is an example of the biggest award in the discipline. A better might be aUniversity Consortium for Geographic Information Science Education Award, or fellowship. Those would be a pretty big deal career wise, but the pages for those topics are abysmal. These are referenced in literature on the subjects, why would we need a Wikipedia page to mention them as well? If that is the case, the pages can be made.GeogSage ^{(⚔Chat?⚔)}08:00, 20 November 2025 (UTC)[reply]

@GeogSage I added that one as well. I agree that Wikipedia's coverage of academic awards is... not perfect. But I don't think you have to worry about us deleting awards from articles about hardworking scientists. I can't speak for Ganesha811 of course but I think they are more interested in getting rid of fake and dubious awards on promotional articles. So I think the focus is more on CEOs not academics. Although I agree that if policy is written it is a good idea to take pre-internet and academic awards into account, and treat them very differently than, for example, the Best in Biz awards you can just buy for a couple hundred dollar.Polygnotus (talk)08:10, 20 November 2025 (UTC)[reply]

My rule of thumb is that an award etc should have a decent cite, preferably secondary, but if the award or at least the org behind it has a WP-article, a primary one may be acceptable, say Grammy etc.

I think awards without WP-articles can be ok to include,if there is a decent secondary cite who bothered to notice. WP doesn't know all.Gråbergs Gråa Sång (talk)09:29, 20 November 2025 (UTC)[reply]

These sections are also common in sports articles (e.g.Michael Phelps#Honors and awards andCathy Freeman#Awards (once I fixed it), and to pick some local examples that I've worked on, [[Bill Roycroft#Recognition andJohn Maclean (sportsperson)#Recognition. Ditto for music, likeLuciano Pavarotti#Awards and honors,Blondie (band)#Awards and nominations, andJoan Armatrading#Honours. I agree with@GeogSage: that trying to police this area is guideline creep and could cause unintended consequences; some of the content in sections like this would disrupt the flow of pages if it was mentioned elsewhere.Graham87 (talk)10:47, 20 November 2025 (UTC)[reply]

In general, I think "Recognition" is a decent heading for this stuff. It can cover knighthoods, Grammys and "30 under 30" Time magazine lists etc. If I start an article, I always go with prose, not table, but that is a personal preference.Gråbergs Gråa Sång (talk)11:01, 20 November 2025 (UTC)[reply]

I agree that musicians, athletes and actors/actresses seem like a decent exception, in that they should probably have standalone sections called 'Recognition', 'Awards', or similar, especially if they've won major awards. But I note that the Phelps page, for instance, does seem to generally follow Proposed Rule #1 - that all the awards seem to have their own Wikipedia page, and for good reason. Pavarotti, too, has many notable awards. But does it really matter to anyone, anywhere, that he received an "Eisenhower Medallion"? Does anyone know what that is? Or that Blondie got the 2022 BBC Longshots Audience Award?

@Polygnotus is right to infer that I'm mostly concerned about businesspeople/politicians and junky "online" awards, not academics and athletes. That's where I most frequently see problems. I wonder if we could shape a policy that applies only to those BLPs. I don't think that merely requiring a secondary, "independent", source would do much, because of the proliferation of junk/slop websites that copy press releases, publish paid notices without disclosure, —Ganesha811 (talk)12:09, 20 November 2025 (UTC)[reply]

Googles AI suggests two possible medals:

People to People International (PTPI) "Eisenhower Medallion": This is the highest award given by the organizationPeople to People International, founded by President Eisenhower in 1956 to foster global peace and understanding. Notable recipients include Mother Teresa and Congressman Emanuel Cleaver, II.

American Nuclear Society (ANS) "Dwight D. Eisenhower Medal": Established in 2014, this award recognizes outstanding leadership in public policy for nuclear science and technology, or significant contributions to nuclear nonproliferation. It is presented bi-annually and honors excellence worthy of international recognition.Gråbergs Gråa Sång (talk)12:28, 20 November 2025 (UTC)[reply]

A source that is a copy of a press release isn't independent and just clarify that the secondary source is non-promotional and it's fine.Katzrockso (talk)12:30, 20 November 2025 (UTC)[reply]

On secondary source for "prize" without WP-article, context matters.Gråbergs Gråa Sång (talk)12:32, 20 November 2025 (UTC)[reply]

It seems no extra policy is needed to avoid award-cruft although it is clearly a major issue on many pages. Secondly, many people may have a long list of awards that are notable according to our secondary sourcing and due weight policies – hence a separate section is often appropriate – whether in prose, list or table form.

That said, it would certainly be helpful to write one or multiple competing essays interpreting how our policies apply to awards. I'm happy to provide feedback on such essays. If during drafting of such an essay it turns out that our policies are in fact deficient, an RfC can be started to upgrade the essay to apolicy supplement.Joe vom Titan (talk)12:23, 29 November 2025 (UTC)[reply]

(A policy supplement is alsoWP:JUSTANESSAY.)WhatamIdoing (talk)03:09, 5 December 2025 (UTC)[reply]

Three years ago, an editor got consensus to create a bot to blank all stale IP talk pages.Wikipedia:Village pump (proposals)/Archive 190#RfC: Bot to blank old IP talkpages The main reason for this was thatStale warnings and other messages will confuse legitimate new editors editing from that IP seeing it apparently directed at them

Fast forward to 2025, and we have temporary accounts; new editors willnever be directed toward talk page IPs. So we don't need to worry about scaring them off.

Given that, I would like to see what the community's attitude is toward this problem now.

Personally, this post was made because I'm trying to track downa Mississippi IP editor who inserted copyright violations into articles about American TV soaps, so I can remove the copyvios. Having their talkpages easily accessible, for searching and whatnot, would be very helpful. Speaking more generally in terms of my CCI work, non-obscured accessible talk pages allow me to more easily link to previous warnings, track copyright violations that were spotted at the times, and track older socks[4][5][6][7], especially if they were duck blocked at the time but not recorded at SPI. I also only have 24 hours in each day; time spent going back to previous revisions is time I'm not spending removing problematic content.GreenLipstickLesbian💌🧸09:35, 23 November 2025 (UTC)[reply]

I support stopping the bot. It has served its purpose.Toadspike[Talk]09:42, 23 November 2025 (UTC)[reply]

I do too.Thryduulf (talk)11:00, 23 November 2025 (UTC)[reply]

+1~/Bunnypranav:<ping>12:25, 23 November 2025 (UTC)[reply]

I'd support stopping this. I looked quickly but maybe is faster (I'm not sure the best way to find this) to just ask if any non-blocked bot is currently performing this task?Skynxnex (talk)12:33, 23 November 2025 (UTC)[reply]

The task was inherited byUser:VulpesBot (run sporadically byDr vulpes, but they've said they plan to run it again I believe?) but I know some editors do large AWB runs to indiscriminately blank the old IP talk pages.GreenLipstickLesbian💌🧸20:34, 23 November 2025 (UTC)[reply]

Ah, thanks. Still agree we should stop blanking them at this point. (And earlier maybe would have been better.)Skynxnex (talk)21:36, 23 November 2025 (UTC)[reply]

• Just to clarify, are we talking about stopping the bot with respect totemporary accounts? Because the bot is set to only blank pages for IPs who have not edited in over five years, there are still tens of thousands of IP talk pages identifying IP addresses. If you look at, for example,User talk pages that link to "Blueberry", there are dozens of them just on that list.BD2412T18:50, 23 November 2025 (UTC)[reply]

  No, it is for IP talk pages only, per what I understood from GLL's example above.~/Bunnypranav:<ping>18:53, 23 November 2025 (UTC)[reply]

  No, it's stopping it for the talk pages of IP's. There are benefits to not blanking these IP talk pages (detailed in GLL's first post), and given that no new editors will be assigned these talk pages in the future there remain almost no benefits to blanking them.

  Whether talk pages of temporary accounts should be blanked after the account expires is not something I can recall seeing anywhere and is not part of this proposal, but given that they will not be reused I can't immediately see any benefits to doing so.Thryduulf (talk)19:41, 23 November 2025 (UTC)[reply]

  I agree with Thryduulf that I see no benefit to blanking them. I do see potentially harm, however, for much the same reason. I often use the What Links Here tool to investigate, and if TA talkpages get blanked, then just like with old IPs, I am no longer able to do that.GreenLipstickLesbian💌🧸20:42, 23 November 2025 (UTC)[reply]

  I would think your use of "What Links Here" is hampered by an excess of links to IP talk pages from which no edits have come in many years, even decades. Wikipedia's purpose is not to serve as a permanent host for long-irrelevant IP talk page messages. That should be even less so when the IP talk pages no longer reflect any current account usage due to the changeover.BD2412T20:57, 23 November 2025 (UTC)[reply]

  Interesting enough, it is not - generally if there's enough links to IP talk pages to become unusable, then there's enough links to registered account talkpages to be unusable. Removing IP talk pages just hampers my ability to look for historic disruption on lower trafficked pages, and also stops me from being able to use the search tool as effectively.GreenLipstickLesbian💌🧸21:03, 23 November 2025 (UTC)[reply]

  To be perfectly clear, thetypical ancient IP talk page message has been where the IP did something like randomly add "poop" to an article once or twice in, say, 2012, got reverted with a warning, and no other edits ever came from that IP address (although I grant that most of those have already been blanked). I think we can refine the model to maintain pages where there is a possibility of copyvio involvement or the like, but I am at least dubious about the long term value of maintaining those pages.BD2412T21:47, 23 November 2025 (UTC)[reply]

  A lot of these old accounts don't always get reverted for copyvio, they get reverted with anti-spam, anti-unsourced content, page hijacking, and really pretty every warning under the sun. Knowing at a glance that an account was editing disruptively in a topic area is still very useful. SeeUser talk:70.49.196.202 orUser talk:62.28.161.202 for examples - I just reverted a bot blanking on the first, and the other was saved because the IP got notified of an AfD late last year. Both of these editors have still open CCIs which either have been or will need to be expanded to include IP edits.

  If somebody sees an IP where the IP only made one vandal edit, got warned, and would rather blank the talkpage than fix whatever lint error they found, they could still do so manually.GreenLipstickLesbian💌🧸22:04, 23 November 2025 (UTC)[reply]

  @BD2412 VulpesBot is exclusion compliant so you can just stick{{nobots}} onUser talk:70.49.196.202 if you want.Polygnotus (talk)00:00, 24 November 2025 (UTC)[reply]

  That was for me. I do a lot of IP talk page blanking outside of VulpesBot's strictures.BD2412T00:02, 24 November 2025 (UTC)[reply]

  I agree that there's no need to hide the content of these pages, and since temp accounts only last for 90 days (under the current configuration), there's no need to ever blank those.WhatamIdoing (talk)21:18, 23 November 2025 (UTC)[reply]

Please seeWP:BOTAPPEAL for instructions on how to start a discussion about reexamination of approved bot tasks. –Jonesey95 (talk)00:15, 5 December 2025 (UTC)[reply]

Note there's probably not an approval to review in this case.Wikipedia:Bots/Requests for approval/VulpesBot was approved as a one-time run ("will return six months after run is complete to request a rerun", which didn't happen), while the operator ofWikipedia:Bots/Requests for approval/MalnadachBot 13 is sock-blocked. Also note that establishing thatconsensus has changed would be a necessary part of a review, so a Village pump discussion would still be useful to establish that.Anomie⚔00:28, 5 December 2025 (UTC)[reply]

(copied and expanded fromWikipedia:Village pump (proposals)): Multiple editors above have said that they see no benefit in blanking IP talk pages. Here's a counterpoint. Most of them are not harmful, but I recently foundUser talk:144.160.98.31 on a report of Linter errors. Its only edits in the lasttwelve years had been seven edits by bots to perform various cleanup tasks, and when I visited, there were still 18 Linter errors on the page, meaning that someone was going to edit that page in the future to clean it up. I replaced its content with{{blanked IP talk}}. If someone had done that years ago, those seven bot edits would have been unnecessary. It made me wonder if there was any point in maintaining any of the IP editor talk pages, since there are (in my understanding) no more IP editors. Can we just blank them all, or at least blank the ones that have errors so that they don't clog up error reports? Is it really useful to maintain a live page with IP editor communication messages that are more than five years old? Editors investigating a particular IP can easily look at the pre-blanked page in the history. –Jonesey95 (talk)22:04, 4 December 2025 (UTC)[reply]

And lest anyone think that the page linked above is an edge case, here's a link tothousands of IP User talk pages with Linter errors. –Jonesey95 (talk)22:43, 4 December 2025 (UTC)[reply]

Jonesey95, why do you fix linter errors on those pages?GreenLipstickLesbian💌🧸22:45, 4 December 2025 (UTC)[reply]

For the same reasons that the MediaWiki developers tagged them. Seemw:Help:Extension:Linter § Why and what to fix for details. Note that stale IP User talk pages are not just an attractive nuisance due to Linter errors. They can also contain templates that are being deleted, categories that are being moved, code that has become obsolete, and other required maintenance needs that cause bots or humans to visit them. –Jonesey95 (talk)22:49, 4 December 2025 (UTC)[reply]

Correct me if I'm wrong, but it looks like a lot like it's to keep pages readable as support for various tags changes. (also, sorry, we should have edit conflicted when I made my post)GreenLipstickLesbian💌🧸22:53, 4 December 2025 (UTC)[reply]

Yes, or to display what the original editor intended without mis-rendering their or anyone else's contributions to the page. –Jonesey95 (talk)22:58, 4 December 2025 (UTC)[reply]

Do you think blanking the page makes it more readable?GreenLipstickLesbian💌🧸23:00, 4 December 2025 (UTC)[reply]

Blanking the page means fewer unnecessary bot and human edits while preserving the page history for those who need to see it. –Jonesey95 (talk)23:04, 4 December 2025 (UTC)[reply]

Or we could just program the bots and tell the humans not the edit the disused pages, and it would have the same impact, right? Sorry if there's something I'm missing, but the lint errors, broken templates, deleted categories, they don't suddenly become less broken, deleted, or errorful when you have to look at an old revision, right?GreenLipstickLesbian💌🧸23:08, 4 December 2025 (UTC)[reply]

Retell the humans not the edit the disused pages: Pages with errors that show up on reports, lists, or error categories but should be ignored make those reports/lists/categories less manageable, because other pages with problems become less visible. I have not found ignoring some pages on reports to be a useful strategy in my years of gnoming dozens of error reports and categories. Do you regularly monitor reports/lists/categories that have a subset of pages to be ignored? –Jonesey95 (talk)23:50, 4 December 2025 (UTC)[reply]

In this case I have to back @Jonesey95 up, it is very annoying and complicated when gnoming to keep a blacklist, and gnoming often leads to the discovery of thousands of minor problems, but also a bunch of big problems.Polygnotus (talk)23:53, 4 December 2025 (UTC)[reply]

Could you set up the report to not include IP talk pages? Or ask the person responsible for the report to remove all IP talk pages? Or just... fix the lint error so that the page remains readable?GreenLipstickLesbian💌🧸23:54, 4 December 2025 (UTC)[reply]

No, no, and not easily. The reports (I linked to a subset of one above) are generated by the MediaWiki software. The word "just" is doing a lot of work in the last sentence; there are over 7,000 IP user talk pages with Linter errors, with a wide variety of errors. –Jonesey95 (talk)00:01, 5 December 2025 (UTC)[reply]

Okay, then you can talk to the folks who generate the MediaWiki softwaree? It does look like they'r sortable, to some extend -Wikipedia:Linter/reports/IP user talk pages by Lint Errors for example, only has old IP talk pages. Couldn't you just ignore that page, rather than updating it?

Or, at the very least, if you'd like to blank a user page - could you go through every single one of the IP's contributions, check them for PAG compliance, do an exhaustive search for any unattributed plagiarism, source text integrity, hoax material, BLP violations, NPOV issues? And repeat it for any neighboring IP's (like others on the /64) before you hide evidence that those problems existed?

Because that's what I'm trying to do.GreenLipstickLesbian💌🧸00:06, 5 December 2025 (UTC)[reply]

We can't sweep the errors under the rug, that defeats the whole point of them being reported in the first place.Tenshi! (Talk page)00:11, 5 December 2025 (UTC)[reply]

Sorry if I'm misreading you, but is blanking them not sweeping them under the rug?GreenLipstickLesbian💌🧸00:13, 5 December 2025 (UTC)[reply]

It would mean that the lint errors would not be reported, though it doesn't address the issue for anyone looking back at the history before the page was blanked.Tenshi! (Talk page)00:18, 5 December 2025 (UTC)[reply]

Wikipedians love to debate everything but this proposal is an obviousyes. In the past, stale IP talk pages were routinely blanked to reduce confusion if someone new used the same IP years later. That reason no longer applies. Routine blanking of stale IP pages should not occur now because it would be pointless churn and would hide possibly useful information when searching for old copy-vios or spam. By contrast, stale pages withWP:LINT errors should be cleaned up. Removal of weird wikitext that generates such errors is often best because wasting time polishing stale comments would not be helpful. Simply blanking a stale page with linter errors gives a clue about what happened to anyone investigating the history. Painfully fixing or removing multiple errors on a stale page would obfuscate history and not have any benefit.Johnuniq (talk)03:58, 6 December 2025 (UTC)[reply]

On edits, diffs, and posts, the timestamp is always in UTC. Discord has a feature where, when you copy/view a timestamp, it displays the time according to the viewer’s local timezone. For example, if you report a post that occurred at a specific time in your timezone, another user will see the corresponding time in their own timezone, which helps avoid confusion. I believe adopting a similar feature would support the modernization of Wikipedia.Rc2barrington (talk)02:46, 24 November 2025 (UTC)[reply]

You can have that withUser:Mxn/CommentsInLocalTime orWP:LOCO.

This somewhat used to be a built-in feature (m:Help:Date formatting and linking): every date was linked everywhere to automatically convert the timezone according to the user's preference atSpecial:Preferences#ooui-23. However, various things resulted in the feature being disabled and then removed:Wikipedia:Manual of Style/Dates and numbers#cite_ref-5.Aaron Liu (talk)03:22, 24 November 2025 (UTC)[reply]

That feature converted the format, but not the time zone.Also, if we wanted, there's a#dateformat parser function that could be used to format dates according to the user preference. But we've never wanted.Anomie⚔04:05, 24 November 2025 (UTC)[reply]

I know this is the idea lab and we're not supposed to just support or oppose, but I can't really find a "yes and" here. I'm generally skeptical of attempts to make users see something different from what was written, even with an opt-in. Fonts and dark mode, OK, I guess, but not actually changing the text. I think that was a mistake from the beginning. --Trovatore (talk)03:39, 24 November 2025 (UTC)[reply]

The perks of living in England are that UTC is just the current time for me. (outside of summer)GarethBaloney (talk)11:37, 24 November 2025 (UTC)[reply]

For myself, I have my preferences set so that everything is set to my time zone automatically. The only thing that doesn't get converted is dates and time when I am editing the source.

Converting the time and date when I need to is a bit of a pain, but it is better for me as I can see at a glance on talk pages how long ago the last replies were, which is the most common thing I see related to time on Wikipedia.

In short, I think that what we have works. --Super Goku V (talk)05:50, 24 November 2025 (UTC)[reply]

DiscussionTools puts "Latest comment: 41 minutes ago" at the top of every talk page and each ==Section==, so you should be able to see at a glance on talk pages how long ago the last replies were no matter what your timezone settings are.

I used to set my local time atSpecial:Preferences#mw-prefsection-rendering-timeoffset but eventually it became too much of a hassle to keep straight which timestamp on the talk page corresponded to which edit in the page history. I find it much simpler to have the whole thing in UTC. The UTC clock gadget inSpecial:Preferences#mw-prefsection-gadgets-gadget-section-appearance may be helpful, if you are trying to figure out what time it is in UTC right now. (I turned that off with Vector 2022, though.)WhatamIdoing (talk)07:18, 24 November 2025 (UTC)[reply]

So as seen in this image I just really think it would be better to show the time I AM IN. Not the standardized UTC time.Rc2barrington (talk)01:26, 25 November 2025 (UTC)[reply]

Try the scripts I linked above.Aaron Liu (talk)01:38, 25 November 2025 (UTC)[reply]

Apparently I don't use DiscussionTools on Wikipedia, but I recall seeing something like that on other Wikis. Still I feel more comfortable seeing the exact time people made their replies rather than seeing the UTC time of when they made their comments. Besides, I don't need to convert the date and time enough to where that would be the bigger hassle. (And yes, I have the UTC clock in the upper-right corner just to keep myself aware of it.) --Super Goku V (talk)05:56, 30 November 2025 (UTC)[reply]

It seems to be the case that at least some-language Wikipedias have adopted the time zone of where most speakers of that language reside. For example, French Wikipedia seems to use CET/CEST. English Wikipedia could really have adopt ET (the time zone where c. half of Americans/Canadians live). Or GMT/BST (time zone used in UK). But the UTC gives a compromise not only because English speakers live across the globe, but its also the time zone used for computers, aviation, ISS, etc. If anyone wants to ensure that comments are outputted in the local time zone, theWP:Comments in Local Time should be of help.

For me, I live in a place where its at UTC for winter and UTC+1 for summer, so I just remember to subtract 1 in summer.JuniperChill (talk)20:08, 10 December 2025 (UTC)[reply]

If someone has repeatedly used an LLM without adequate verification of its output, I think we should be able to mass-revert their edits. I envisage a system whereby we only have to glance over each edit and check itis AI-generated, rather than the much higher bar of reverting the cases where the AI has caused a definite problem. My rationale is that if someone has repeatedly failed to use AI responsibly, then their other uses can be assumed to be irresponsible as well. Roughly speaking, I imagine the level of abuse required being roughly the current threshold for a dedicated subpage of the AI cleanup noticeboard. It has been remarked on numerous occasions that checking whether AI output is inclusion-worthy is about as hard as writing the material from scratch, so I think requiring other users to perform this level of checking before reverting AI edits is not reasonable. What do people think?lp0 on fire ()22:03, 26 November 2025 (UTC)[reply]

Are we talking about a blocked user? Was there a discussion about their behavior? I could imagine forming a consensus toWikipedia:Rollback all of an individual's edits, but I'm not sure that I'd recommend that an individual editor unilaterally declare that everything you did in the mainspace is definitely AI and should all be reverted.

Also, outside the mainspace, it's a bit more complicated. If an AI-generated comment on a talk page received a reply, it probably shouldn't be reverted.WhatamIdoing (talk)23:42, 26 November 2025 (UTC)[reply]

IDK if a tool like this is a good idea, but if it did exist I'd envision it being used for blocked editors (look up the user whirlingmerc for an example that wasted hours of my time). For editors who have not been blocked, it's appropriate to ask them to clean up their own mess by self-reverting all the problematic contributions. --LWG^talk01:09, 27 November 2025 (UTC)[reply]

I think it certainly applies to talk pages, perwall of text issues. All AI edits should be deleted, per my comment below.Yesterday, all my dreams... (talk)14:54, 27 November 2025 (UTC)[reply]

I agree that if an editor has been blocked for using AI, reverting any of their edits that look like AI output should be allowed. This sounds likepresumptive deletion in copyright cleanup. I don't think we need a special tool for this though.Toadspike[Talk]07:23, 27 November 2025 (UTC)[reply]

That presumptive deletion is exactly the idea I was going for. I wasn't suggesting a special tool, but I think mirroring the wording there pretty much exactly could save a lot of time (i.e. not requiring that the user be blocked). If someone does a long spree of AI additions but leaves the project before anyone notices, there's no need to block them, but being allowed to mass-revert their mainspace edits would still be helpful.lp0 on fire ()07:45, 27 November 2025 (UTC)[reply]

I agree, and think to succeed you need to invent a name for it, say "vagabond AI editor" reverts. I think this is important because the trend is the increase in AI edits. And I think it should also apply to talk pages givenwall of text issues. AI edits are the termite that can ruin Wikipedia.Yesterday, all my dreams... (talk)14:50, 27 November 2025 (UTC)[reply]

I don't see why we can't just call it presumptive deletion. For talk pages, we have{{aitop}}/{{aibottom}} already and I think that's enough.lp0 on fire ()15:12, 27 November 2025 (UTC)[reply]

Or we could make something similar toTemplate:Single-purpose account, except instead of saying:

—Example (talk •contribs) has madefew or no other edits outside this topic.

for AI use, it could say something like:

—WhatamIdoing believes that this comment waswritten by generative AI instead of byExample (talk •contribs).

WhatamIdoing (talk)20:49, 27 November 2025 (UTC)[reply]

Yesterday, I'm not convinced with your view. In fact, you're rapidly making me less supportive of this whole idea. It begins to feel like this:

• We should revert everything.
  • Maybe not talk page comments, if someone's already replied.
• No, really, everything, because it's aWikipedia:Wall of text.
  • Even if it's just a short reply?
• Really, everything, because everything is aWikipedia:Wall of text.

You obviously loathe AI use, which is fine. But what if the comment isnot a wall of text? Would you seriously recommend reverting a one-word reply because a single word is "a wall of text"? How would you even know whether such a short comment used AI?

Would reverting a talk-page comment actually help anyone?WP:REDACT says usually no, particularly if someone's already replied. Would it be better than alternatives such as striking (like we do with socks), hatting (e.g., aitop/aibottom), labeling (like we do forWP:SPAs), or archiving? I doubt it.

I wonder whether your ham-fisted recommendation signals that you're getting burned out. If editing feels like a sisphyean struggle against the forces of spam and stupidity, then you might try to find a way to contribute that feels fun and/or effective.WhatamIdoing (talk)20:45, 27 November 2025 (UTC)[reply]

Well, you know that our agreement rate is pretty low. But that is the nature of free speech. As for "forces of spam and stupidity" being in full swing on many pages, we actually agree on that. And I assume you are also thinking of my talk comment onfuzzy concept. On that page OR and stupidity are in full swing indeed. We can not have a "respectable" encyclopedia with that type of content.Yesterday, all my dreams... (talk)00:44, 28 November 2025 (UTC)[reply]

I have spent no time looking at your comments on talk pages, so no, I had no idea that you posted a comment there (that says nothing about AI use).WhatamIdoing (talk)04:04, 28 November 2025 (UTC)[reply]

I've been thinking about this sort of thing as well. Regardless of the approach we end up taking, we do need to be more proactive in removing unverified AI content and quickly putting a stop to people who add it.Thebiguglyalien (talk)🛸04:57, 28 November 2025 (UTC)[reply]

Agreed. A quick look at the AI cleanup noticeboard will make it abundantly clear how serious a problem this is. As I see it, there are three levels of assuming good faith we could exercise when doing the cleanup (clarifying what I mean here because I think there was some confusion above; sorry in advance for the wall of text).

1. If someone has repeatedly misused LLMs, we go through their contributions and delete anything that violates policy (weasel/peacock words, OR, hallucinations, &c.) but we can't revert anything until we've identified the problem. This might involve verifying sources and/or translations, might require specialised knowledge, and is about as difficult as writing the content from scratch. This is the current standard, and it makes cleaning up after LLM use unreasonably difficult, leading to a growing backlog of additions to Wikipedia that might be nonsense.
2. Like copyright violations, any mainspace edits by an AI abuser can be reverted indiscriminately. This would make cleaning up after AI misuse very easy (although, given how easy it is to write content with AI, this might still not be enough).
3. What I was originally suggesting was a middle ground: if someone has repeatedly misused LLMs, then any edit of theirs that looks AI-generated can be reverted without proof that the AI has hallucinated or otherwise violated policy, because they are presumedincompetent. This would still make cleanup much easier than in currently is, with reduced risk of undoing good contributions.

lp0 on fire ()07:41, 28 November 2025 (UTC)[reply]

Sockpuppet cleanup allows other users to restore sock edits if they are positive (every now and then some are, or partially are), without putting that burden on the cleanup.CMD (talk)09:13, 28 November 2025 (UTC)[reply]

I don’t think it’s a matter of LLM or not LLM; it’s a matter of good editors and bad ones. There were plenty of bad editors who tried to push bad articles before LLM. The fairest way to approach low-quality articles is the same way it has always been done: with tags that can only be removed if an editor has done the necessary work to justify their removal.

We can’t allow LLM to become a reason for people to ban whoever they want, for whatever reason. Take a contentious subject, for example: an editor could be falsely accused of using an LLM in order to censor their vote on articles.Orlando Davis (talk)15:53, 28 November 2025 (UTC)[reply]

Instead of deleting the articles, we can have a 3 strike policy where you get banned for 24 hours if you have 3 strikes, and are banned permanently after enough strikes without an attempt to change your behavior.Orlando Davis (talk)16:29, 28 November 2025 (UTC)[reply]

The difference is that LLMs allow people to churn out huge amounts of bad content extremely quickly without first having to learn how Wikipedia works, which makes it significantly more disruptive than just "bad editors".

I don't think your worries about false accusations make sense. If anyone tried to censor someone by accusing them of using AI, then much like accusing someone of being a sock, that would be highly problematic and likely lead to the accuser being blocked (especially in a contentious topic); however, it's much easier to spot a bad-faith accusation of AI than a bad-faith accusation of sockpuppetry.

Your suggestion of "get banned if you have enough strikes" (I assume you mean blocked not banned) doesn't sound substantially different from the standard system of "you get blocked if you keep doing stuff wrong after being warned" and indeed the template{{uw-ai1}} through{{uw-ai4}} exist for this very purpose.

I think you may have misunderstood the purpose of this proposal: it's not for dealing withpeople who disrupt the project using AI but rather for cleaning up theiredits, which otherwise demands an unreasonable amount of time from the users doing the cleanup.lp0 on fire ()16:43, 28 November 2025 (UTC)[reply]

Couldn’t a way to reduce backlog be to put a cap on how many articles and edits a user can perform per day, to give reviewers enough time to keep up? For example, a 1–2 article per day limit and a 100–200 edits per day limit. What do other editors think?Orlando Davis (talk)17:09, 28 November 2025 (UTC)[reply]

That soundsway out of scope for this issue. Bear in mind most a lot of AI cleanup involves cleaning up after editors who stopped before (or when) they were noticed, so such a filter would have to apply to all users. I also note that 100 edits a day isn't very much for normal editing, but it's a huge amount of work to clean up after 100 edits of AI drivel. For example, seeWikipedia:WikiProject AI Cleanup/Noticeboard/2025-09-17 Thefallguy2025 which is from early September and still less than half done.lp0 on fire ()17:25, 28 November 2025 (UTC)[reply]

What about the cap on edits being applied more strictly to flagged users?Orlando Davis (talk)17:41, 28 November 2025 (UTC)[reply]

Or to newbies. Very few brand-new accounts make even five edits on the first day.WhatamIdoing (talk)01:24, 29 November 2025 (UTC)[reply]

To the extent that new accounts do, they're usually people who have made accounts before (sockpuppets,WP:CLEANSTART)Katzrockso (talk)01:28, 29 November 2025 (UTC)[reply]

Or someone who couldn't figure out how to use the [Preview] button, so it took them five tries to fix the same sentence.WhatamIdoing (talk)21:49, 14 December 2025 (UTC)[reply]

So, #3 is what we've been doing atWP:AINB since around August and it has been working just fine, albeit without any PAG to justify... we typically leave an edit summary like "LLM cleanup, as discussed at AINB and/or ANI". I personally have cleaned ~500 articles in this way and only on one of those articles did someone else complain, and I just reverted my deletion and asked that user to verify/fix the article, which they did. Also agreed with Toadspike that it would be a rare case where a tool would be helpful. In almost all cases this has to be done manually.NicheSports (talk)19:45, 28 November 2025 (UTC)[reply]

Oh, that's encouraging I suppose. It would still be nice to formalize it in a guideline (or at minimum a WikiProject advice page), for the combination of legitimacy and clarity that we get from explicitly writing stuff down.lp0 on fire ()23:05, 28 November 2025 (UTC)[reply]

I feel like we can just use the general provisions ofWP:CHALLENGE etc if it's the usual AI stuff and the sources don't verify.Alpha3031 (t •c)23:50, 28 November 2025 (UTC)[reply]

Also,WP:5P3 exists. I don't really know why this is even a discussion to be honest. Text can be added, changed, or removed at any time, that's the fundamental point of a wiki.Gnomingstuff (talk)01:15, 30 November 2025 (UTC)[reply]

Good idea, any chance you want to give it a whirl? Maybe makes sense to start as an advice page at WP:AIC. Also pointing you to this, which is an idea I had with some support at AIC:WT:WikiProject AI Cleanup/Archive 4 § Guidance on handling article with mostly minor edits subsequent to LLM-rewrite. Maybe this could be incorporated?NicheSports (talk)21:16, 29 November 2025 (UTC)[reply]

I fully agree with implementing a presumptive deletion-esque policy here. Take a look atthis tracker on AINB for example - 74 pages by a chronic LLM user need to be reviewed. I've been doing some myself, and I've found that a lot of it is innocuous AI copyediting, but then on one or two edits, you'll see places where the AI accidentally combines two sentence clauses and changes the meaning, or does a thesaurus rewrite of a direct quotation from a real person; it requires an intense and time-consuming level of scrutiny to pick those out, but I can't simply in good faith revert everything without checking, because a lot of it is useful copyediting changing articles to a more formal tone.

It would be much easier to just go in and revert everything this person has substantively changed.Athanelar (talk)18:16, 10 December 2025 (UTC)[reply]

I wouldn't necessarily want to require editors to revert everything, or to send a bot around, but for individual editors who have been specifically identified as causing problems, I think that it's reasonable to assume a problem unless you can prove otherwise at a glance. For example, @Athanelar, I looked atthat editor's contributions toGeorg Klein (composer). They might be fine. But I can't tell at a glance. And the editor is known to have problematic contributions. So I think that reverting that with a suitable edit summary would be justified.WhatamIdoing (talk)18:47, 10 December 2025 (UTC)[reply]

I'm inclined to agree that the community is currently fairly vigorously contesting LLM-slop. There are even false positives, at least one case of something from 2010 getting tagged. Remember that LLMs are trained on Wikipedia. Nobody tagged me for this but I recently saw text I had written where I used "fostered" and "surpassed," two tagged vocab words, but on double-checking both of which were used by the sources, so I was being faithful by also using them.Shlomo Lambroza (Wikidata) andDiana Dumitru probably didn't use an LLM, they used that vocab because they with precise diction decided that "surpassed" and "fostered" were the best way to express themselves at that moment. Not saying that the slop isn't a big problem but right now I think there is adequate control of it - thanks to a lot of volunteer work, time, energy. See, I did 3 things. But I remember someone telling me about the rule of 3 at least 5 years ago and it had nothing to do with LLMs.Andre🚐02:08, 29 November 2025 (UTC)[reply]

To be clear, I'm not proposing that anyone can delete anything they personally think might have been written by an LLM, but in cases where a user has a long history of LLM misuse, it feels unlikely that they also just happen to write like an LLM. I don't necessarily agree with you that enough is being done to clean up after LLMs to avoid needing a measure like this, but rven if that's true, such cleanup still wastes a huge amount of community time. The current wording ofWP:ONUS means that if a source has been provided, it's the responsibility of the person removing information to check that verification fails. The thing about AI is it's very easy to make something that looks convincing, meaning one often can't tell at a glance whether the sources are okay. This creates aWP:TNT situation where it's easier to blow it up and start over than to fix the problems by manually checking each source, which can take a very long time.lp0 on fire ()13:01, 29 November 2025 (UTC)[reply]

That makes sense. But isn't it pretty easy to make something look convincing without AI? Shouldn't we use a system of cleaning up that isn't so confrontational? Couldn't erasing pages start edit wars? There have been very good alternative suggestions here.Orlando Davis (talk)20:31, 29 November 2025 (UTC)[reply]

It's not true thatWP:ONUS means that if a source has been provided, it's the responsibility of the person removing information to check that verification fails.WP:BURDEN means the other editor has to provide one source (but only one; you can't make themWP:FETCH and endless supply of sources).WP:ONUS says only that it's the other guy who has to organize a consensus to include the information.

One of the footnotes in BURDEN gives a partial list of reasons why one might be justified in removing cited content: removing editors "must articulate specific problems that would justify its exclusion from Wikipedia (e.g., why the source is unreliable; the source does not support the claim;undue emphasis;unencyclopedic content; etc.)". In practice, I suspect that an edit summary along the lines of "Presumptive removal of text from an editor since blocked for abusing AI tools" would be considered an entirely sufficient articulation of a specific problem.WhatamIdoing (talk)21:46, 29 November 2025 (UTC)[reply]

That was my failure to read the footnote; thanks for clarifying. I still think it'd be helpful to formalize allowing such presumptive deletions.lp0 on fire ()22:09, 29 November 2025 (UTC)[reply]

It might be useful to have a short page on when and why aWikipedia:Presumptive removal would be warranted. If it gets used and doesn't create a lot of problems, it would probably be easy to get an "Oh BTW there's this WP:PRESRM thing..." added to a guideline or policy somewhere.WhatamIdoing (talk)23:25, 29 November 2025 (UTC)[reply]

To be clear, are you suggesting a single page that collates all the common kinds of presumptive removal (AI, socks, copyvios, banrevert, arbecp, maybe something else I haven't thought of)?lp0 on fire ()09:11, 30 November 2025 (UTC)[reply]

Yes.

I'm thinking of something that's more of a 'process description' page than a 'rulebook'. It could be a little bit similar toWikipedia:Why was the page I created deleted? orWikipedia:What is significant coverage? After someone reads it, they should know what presumptive removal is (mass removal of edits from known-problematic individuals), why we use it (efficiently protecting Wikipedia), and what to do (careful evaluation).WhatamIdoing (talk)23:45, 30 November 2025 (UTC)[reply]

lp0 on fire, let's not forget about creating this documentation.WhatamIdoing (talk)03:16, 5 December 2025 (UTC)[reply]

I might have time to do this today. Thanks for the reminder.lp0 on fire ()08:49, 5 December 2025 (UTC)[reply]

@WhatamIdoing: Very early stage beginnings of a draft created atUser:lp0 on fire/Drafts/Presumptive removal; I'll work on this when I have some time but feel free to contribute.lp0 on fire ()16:36, 5 December 2025 (UTC)[reply]

It may be relevant to this discussion that Orlando Davis has been temp-blocked following anANI report concerning disruptive editing and LLM use.fifteen thousand two hundred twenty four (talk)02:34, 1 December 2025 (UTC)[reply]

A system of cleaning up that isn't so confrontational is easy to achieve, simply by getting the confrontation over with. Four warnings followed by a site ban. AI use is significantly more damaging than ordinary vandalism.TooManyFingers (talk)01:49, 14 December 2025 (UTC)[reply]

I'm doubtful about your assertion that "AI use is significantly more damaging than ordinary vandalism". Maybe we have different ideas of what "ordinary vandalism" looks like?WhatamIdoing (talk)02:16, 14 December 2025 (UTC)[reply]

In the Wikipedia app, the English Wikipedia doesn't show whether an article is Good or Featured. For example, in the German Wikipedia—likethis good article—this information appears at the bottom of the article in the app, and it even shows the date when the article was selected as Featured. I strongly suggest adding this feature—and the date of selection—to the English Wikipedia app as well.Vastmajority20025 (talk)19:37, 28 November 2025 (UTC)[reply]

Last I heard, readers don't notice or care about those little icons, so why should we bother?WhatamIdoing (talk)21:47, 29 November 2025 (UTC)[reply]

Yeah @WhatamIdoing, but it would be better for the English wikipedia to be more accessible on phone for a better experience, like German wikipedia example, and it doesn't need to be icon, likethis article in German wikipedia, at bottom of it is a section for the date of the article turning Good or Featured, and says it is Good or Featured.Vastmajority20025 (talk)16:59, 4 December 2025 (UTC)[reply]

I don't think this requires much consensus. With enough time, try opening upthe publicly-available code for the apps and implement it!Aaron Liu (talk)20:54, 9 December 2025 (UTC)[reply]

@Aaron Liu, I am not familiar with programming and even with help ofchatgpt, I don't know where to put it. Now If anybody familiar with programming and know what to do, do this, would be great.Vastmajority20025 (talk)13:47, 14 December 2025 (UTC)[reply]

I'm opening this as a more general idea lab discussion since I don't have a specific proposal, but we've reached the point now where we really need to be looking into how we frame Wikipedia's relationship with AI, especially in public-facing areas. There's currently nothing public-facing, not even on the main page, emphasizing that Wikipedia is a human-written encyclopedia (or whatever term you want to use). As LLM content only becomes more common, the fact that Wikipedia is written by humans is going to become one of its defining characteristics and a major reason why it's a better alternative to other sites. Has anyone given thought to how we might incorporate this?Thebiguglyalien (talk)🛸02:57, 29 November 2025 (UTC)[reply]

I do think Wikipedia has always had a human and humanistic aspect, and I support the proposal in the abstract. Maybe we could have a contest for someone to design a banner or an interactive display to promote Wikipedia: The Free as in Libre, Human Encyclopedia. Like we used to do in the old days.Andre🚐03:02, 29 November 2025 (UTC)[reply]

Awful suggestion. 1. Being human-written is not an important pillar of Wikipedia, it is rather the bare minimum for any respectable encyclopedia, book or news article. Hence it's a bad idea to emphasive this fact so prominently. 2. Wikipedia is not "human". That particular phrasing is confusing.

I don't object to including the fact that Wikipedia is human-written in some guidelines, essays or promotions. But it's not the central selling-point of Wikipedia – lots of other outlets are human-written too but inferior to Wikipedia in many ways (e.g. less reliable).Joe vom Titan (talk)13:17, 29 November 2025 (UTC)[reply]

I have some bad news for you about the internet of the 2020s.Thebiguglyalien (talk)🛸15:41, 29 November 2025 (UTC)[reply]

What are those bad news? Has AI slop appeared on nytimes.com or home.cern yet? AI is neither the biggest problem in the world nor the biggest problem on the internet. For one, misinformation spread by oil companies, oligarchs and petrostates to serve their own interests is much more insidious.Joe vom Titan (talk)13:44, 30 November 2025 (UTC)[reply]

nytimes.com almost certainly, can't speak for cern though. The point is, if you google something a good 75% of the time most of the first page results will be SEO infested ai-generated spam that vaguely summarizes a topic instead of providing useful information. Wikipedia is fundamentally not that, and as more and more of what used to be considered "reliable" websites for most people become infested with slop I feel like it's worth highlighting the fact that we aren't doing thatmgjertson (talk) (contribs)20:36, 8 December 2025 (UTC)[reply]

What's the AI slop on nytimes.com?Aaron Liu (talk)20:55, 9 December 2025 (UTC)[reply]

Pretty sure it was Perplexity AI that copied NYT, not NYT posting AI slop.[8]Polygnotus (talk)00:06, 11 December 2025 (UTC)[reply]

Even more bad news—The list (misinformation spread by oil companies, oligarchs and petrostates) includes states, x-archs... that have lots of cash they crave to grow—what better way to get richer than AI (restricted by very high subscription fees). $20USD/mon is my limit. What's Bezos'? Oh, right, Amazon is one of the three largests investors in AI—looked at or listened to the A. website lately? —Neonorange (talk to Phil) (he, they)03:32, 1 December 2025 (UTC) —[reply]

• I am quite keen on the idea of making a statement of principle like this. As for the implementation, I think there are a few possibilities. I can see something being incorporated into theWikipedia:Five pillars. Another possibility is to add something intoWikipedia:What Wikipedia is not, e.g. 'Wikipedia is not written by machines'. The last possibility I can think of is to write a new one-line policy or guideline to the effect that 'Wikipedia is a human-written encyclopaedia', in a similar format toWP:IAR. Whatever is proposed will need wide community support to be adopted.Yours, &c.RGloucester —☎03:59, 29 November 2025 (UTC)[reply]

  Just today I was musing on writing a "Wikipedia is not Grokipedia" essay which stresses that the entire point of Wikipedia is to eliminate error by having different perspectives, opinions, editing approaches etc coming together to make consensus, and how using AI essentially centralises everything into coming from one authorial voice which fundamentally undermines the spirit and purpose of the project.Athanelar (talk)18:24, 10 December 2025 (UTC)[reply]

• It is actually difficult to know post-2022 if Wikipedia was written by human, machine, or machine-human combo. --GreenC04:25, 29 November 2025 (UTC)[reply]

  It's almost as if we should have some sort of clarification in public-facing areas stating that the purpose of Wikipedia is to be written by humans.Thebiguglyalien (talk)🛸15:37, 29 November 2025 (UTC)[reply]

  @Thebiguglyalien So we should just indef you and only allow humans to edit Wikipedia from now on?Polygnotus (talk)00:08, 11 December 2025 (UTC)[reply]

• I wonder how this would interact withm:Abstract Wikipedia. If I write code that produces a sentence – "$company-name is a business in $country-name that produces $product" – is that still "human-written"?WhatamIdoing (talk)05:08, 29 November 2025 (UTC)[reply]

  The interaction is a distinguishing point between English Wikipedia and Abstract Wikipedia (is that the final name?). Auto-generated text is not human-written.CMD (talk)05:39, 29 November 2025 (UTC)[reply]

  'Language-independent articles'? How has the world become so dystopic? Each language has its own mode of communication, its own mode of thinking. There is no one-to-one relationship between a concept in one language and a concept in any other. Even if we could modify language to allow for such things, this would destroy the organic diversity that is the body of human language. God knows I don't want to read an article that is written in a manner inconsistent with the thought process that is associated with the language in which it is written. I can only imagine the horrible damage this will do to languages other than English. Haven't we done enough harm with the likes of the Scots Wikipedia?Yours, &c.RGloucester —☎06:05, 29 November 2025 (UTC)[reply]

  On the other hand, there are quite a few articles that exist in fr, de, etc and nobody has created in en. Google Translate does ok, but affects ease of discovering information and browseability. So if we had a way to conceptualize a layer between factoids and prose, it could be useful to aid in translation or spreading knowledge further and sooner. At any rate, this is only theoretical. If and when it is accomplished, it may or may not even achieve critical mass.Andre🚐06:12, 29 November 2025 (UTC)[reply]

  Our goal is not to have more articles for the sake of more articles, but to have articles that meet our quality standards. Usually, there is a reason why an article may exist on a non-English Wikipedia, but not on the English Wikipedia. The English Wikipedia has much higher standards in terms of referencing. Very often, articles found on other Wikipedias lack sources at all, or rely heavily on niche sources that would be insufficient to establish notability here. Additionally, they are frequently written from a perspective that is insufficiently global for the English Wikipedia. I have many times endeavoured to translate an article from one Wikipedia to another, in the languages that I know, only to be stymied by the poor quality of the content. It is often easier to start a new English Wikipedia article from scratch, using some of the sources from the other Wikipedia as a foundation.Yours, &c.RGloucester —☎06:19, 29 November 2025 (UTC)[reply]

  Not necessarily always the case. There are many good quality articles on fr or de that if I could snap my fingers to port over with an idiom-proof translation would be worthwhile in edifying readers, and have appropriate references.Andre🚐06:28, 29 November 2025 (UTC)[reply]

  Ask a translator for assistance, there are plenty of volunteers willing to help. No translation can be 'idiom-proof', unless the fundamentals of language itself are to be destroyed.Yours, &c.RGloucester —☎07:21, 29 November 2025 (UTC)[reply]

  (I wouldn't use the German-language Wikipedia as an example of appropriately cited articles, as their standards are very different from ours.)WhatamIdoing (talk)21:49, 29 November 2025 (UTC)[reply]

  I am aware that a human translation can't be idiom-proof, but that is the promise of an abstract Wikipedia, a syntactically complete database-frontend of facts that takes Wikidata beyond simply data and makes actual articles. I mean another way to do that would just be to feed Wikidata to an LLM that doesn't have other knowledge or the ability to call out to random tools and make things up, but simply weaves Wikidata into article form. That wouldn't work though without a lot more UX work and volunteer time on data entry. At any rate, I don't necessarily think the articles I'm personally interested in are the ones that translators need to work on, so it kind of feels like an imposition to dump my requests into that list. I'm sure there's a backlog. Instead, I'm dumping them into Wikiprojects that will potentially have a contributor write an English article while just consulting the other articles. But I do know that there are many many topics that are adequately covered in international Wikipedias. It seems silly to ignore the possible technological developments that will make reading content in other languages more accessible. Here's an example:Mikhail Kulisher (he;ru;uk). The articles seem fairly complete and are referenced. There is a whole pile of similar articles.Andre🚐05:59, 1 December 2025 (UTC)[reply]

  Your claim thatThere is no one-to-one relationship between a concept in one language and a concept in any other sounds a bit overstated. Simple facts (Angela Merkel was Chancellor of Germany; calculus is a type of mathematics; carrots are edible) seem to translate quite well between most languages. There are individual instances of non-translation (家は青い – the house is, um,blue or green or thereabouts; ), but it's not true that there are no concepts that map to the same concept in any other language.WhatamIdoing (talk)22:10, 29 November 2025 (UTC)[reply]

  I said that there is no 'one-to-one' relationship, not that there was no relationship. The process of translation is a delicate one. What you call a 'simple fact' could potentially be translated tens of different ways. The meaning of 'edible' can be rendered many ways in English, and it is likewise true in most other languages. I could say 'can be eaten', 'able to be consumed', 'safe to eat', 'comestible', depending on context, register, &c. By creating an artificial one-to-one relationship between words, whereby 'edible' can only be rendered as one specific term in another language, you destroy the organic diversity of that language, and the naturalness of the text produced. It is very likely that whatever term is chosen may end up being inappropriate in the relevant context, because the person creating this artificial one-to-one relationship will not have a full grasp of the relevant language, and will rely on horrible dictionaries or computer code. The end result will be Scots or Greenlandic Wikipedia, redux.Yours, &c.RGloucester —☎07:51, 30 November 2025 (UTC)[reply]

  And yet, somehow, I think that if it offered me a sentence like "carrots are edible^[source]", and I didn't think it was appropriate in the relevant context, had the wrong register, etc., then I could probably either reject it or re-write it without destroying either the organic diversity of the English language or the naturalness of the text in the Wikipedia article.WhatamIdoing (talk)23:49, 30 November 2025 (UTC)[reply]

  Sure, if you're a speaker of English and a speaker of the source language, you will be able to evaluate whether the machine's output is suitable or not, though I don't see how this will save any time as compared with traditional translation. However, I expect that this 'abstract Wikipedia' will mainly be used for minor languages, with few available editors qualified to make such judgements. It is a recipe for disaster.Yours, &c.RGloucester —☎11:05, 1 December 2025 (UTC)[reply]

  I think it will get used in a variety of ways, many of which involve numbers that change in a more or less predictable fashion. For example: "According to $source, the current population of the world is estimated to be $world-population.^[source]"

  Speaking of which, I frequently wish that the second sentence ofWorld population had an up-to-date number in it.WhatamIdoing (talk)03:20, 5 December 2025 (UTC)[reply]

  I'm a native Anglophone, and I wrote poetry in Hebrew that I had trouble translating.user:RGloucester is absolutely right that there are things that don't translate well. "Traduttore, traditore" --Shmuel (Seymour J.) Metz Username:Chatul (talk)14:30, 1 December 2025 (UTC)[reply]

  Of course there are things that don't translate well. I object to the overbroad statement that there isno one-to-one relationship between any part of one language and any part of any other language, for any statement.WhatamIdoing (talk)03:22, 5 December 2025 (UTC)[reply]

  Indeed in closely related languages it is likely that there are very many concepts and phrasings that correspond 1:1 and while I haven't attempted to verify this I would be astonished if a phrase like "Thryduulf is a living person." could not be directly and accurately translated into the majority of the world's languages without any change of meaning or nuance.Note I explicitly don't say "all" as I'm sure there will be some exception somewhere, perhaps there is a language that mandates specifying whether this is direct, second hand or inferred knowledge or requires an explicit indication of gender.Thryduulf (talk)03:53, 5 December 2025 (UTC)[reply]

  Angela Merkel was Chancellor of Germany "Okay, well, what's a "chancellor?" We don't have a word for that in Examplese, so we could keep it untranslated, but that might be confusing, so I'd rather try to pick an equivalent word in our language."

  Well, in the context of Germany, the chancellor is the executive leader of a federal republic; i.e., an electoral-democratic state divided into smaller polities with some degree of independence, which is governed by elected representatives in charge of each administrative subdivision, where the chancellor acts as theprima inter pares of the representatives, representing the whole federal state rather than an individual subdivision. Suddenly the Examplese-speaking editor has quite a lot more translating to do.Athanelar (talk)18:30, 10 December 2025 (UTC)[reply]

  @Chipmunkdavis, please seem:Abstract Wikipedia/Abstract Wikipedia naming contest. I gather that the team would very much like to have a different name (though I don't have any insight into why).WhatamIdoing (talk)21:52, 29 November 2025 (UTC)[reply]

  I was pretty sure that I had proposed Wikigenerator, but I guess great minds think alike.Andre🚐21:58, 29 November 2025 (UTC)[reply]

• It is fair to exclude LLM written content from Wikipedia on the grounds that they're currentlynot very competent at the task of writing an encyclopedia article, but I am opposed to any display of human or "humanistic" chauvinism, specially anywhere as prominent as the front page. It is also not practical to uphold this claim/promise, as it basically impossible to be certain whether any text is "really human" or has had a partial/full LLM contribution behind it.TryKid ^{[dubious –discuss]}14:48, 29 November 2025 (UTC)[reply]

  Seconded. The LLM text is more prevalent than some people realize, and certainly more than laypeople realize. Making such a claim after 2 years of having no AI policy or guidelines would be telling our readers a lie.Gnomingstuff (talk)05:16, 30 November 2025 (UTC)[reply]

  I agree on all counts. LLM is both unsuitable for writing new articles, but it's also not outright banned by policy (at least not yet). Even if it were banned, there are still articles out there that have been written partially using LLM.

  Wecould theoretically ban any LLM use, but that still wouldn't make the statement "Wikipedia is entirely human-written" true. –Epicgenius (talk)23:53, 30 November 2025 (UTC)[reply]

  Gnomingstuff &Epicgenius, I don't know if you're referring to this or if you haven't seen it yet, but as of last week there is in fact a content guideline in regard to creating articles with LLMs, and there are ongoing discussions to decide how its scope will be expanded beyond simple article creation:Wikipedia:Writing articles with large language models.Thebiguglyalien (talk)🛸06:04, 1 December 2025 (UTC)[reply]

  @Thebiguglyalien, thanks for the ping. I did see this, but it doesn't apply retroactively, nor does it cover LLM-assisted expansions of existing articles. We'd need to ban LLM for at least the latter before we can claim that WP is human-written (and even then, people will try to sneak in LLM text constantly, so vigilance will be required).Epicgenius (talk)06:44, 1 December 2025 (UTC)[reply]

  To clarify, when I said "2 years" I meant the prior 2+ years' worth of accumulated AI edits. (The guideline was approved just days before the 3-year anniversary of ChatGPT.)Gnomingstuff (talk)13:47, 3 December 2025 (UTC)[reply]

• First, let us ask if a bicycle is "human powered"? It is, but provides more power than walking. Wikipedia can be human powered but with bicycle type tools. The human decides where the bicycle goes. Secondly please let me introduce the concept ofclosed loop system to the discussion. The LLM nightmare is when other sources pick half baked content from AI generated sources, and said sources pick it up again themselves. The term to User then is jambalaya knowledge.Yesterday, all my dreams... (talk)16:00, 29 November 2025 (UTC)[reply]

  This has nothing to do with whether bots arewriting the content. No one said "human powered" until you did.Thebiguglyalien (talk)🛸16:06, 29 November 2025 (UTC)[reply]

  Yes, I used human powered, because I think the term should be considered. If you do not like it, do not use it. Others may consider it and it will linger in their minds.Yesterday, all my dreams... (talk)16:22, 29 November 2025 (UTC)[reply]

• Even ignoring all the AI-related issues, there are many articles (partially) written by bots - see for example the article about nearly any small town in the United States - so the statement isn't true.Thryduulf (talk)00:00, 1 December 2025 (UTC)[reply]

  agreed. i think wikipedia's greater value is its verification and its robust community that debates articles. unfortunately, thats not as pithy as "human-written"User:Bluethricecreamman(Talk·Contribs)05:31, 1 December 2025 (UTC)[reply]

  Maybe it should be Wikipedia: The encyclopedia of human ideas and discussion? Surely we agree the ideas and discussion are human even if we can't, as Gnomingstuff and Thryduulf point out, actually claim the articles are all human-driven, aside from LLMs, due toRambot and similar automation that has been around almost as long as the project.Andre🚐05:47, 1 December 2025 (UTC)[reply]

The Wikimedia Foundation itself seems to be much less wary of generative AI, using it in some of their TikTok videos (one on Wicked (film), if I do recall) and advertising in their 25th anniversary video how Wikipedia trains AI. If there is a community consensus that Wikipedia and generative AI are not allies, should we address this with Foundation leaders so they can alter their messaging?✨ΩmegaMantis✨^blather20:19, 1 December 2025 (UTC)[reply]

The foundation has long since even pretended to represent the consensus of wikipedians (as evident by the temporary account roll out)mgjertson (talk) (contribs)20:38, 8 December 2025 (UTC)[reply]

It certainly doesn't have to stay that way. The Foundation has relented to our demands involving AI in the past, like through halting the Simple Summaries feature. The temp account rollout seems to be prompted by a legal issue faced by the WMF.✨ΩmegaMantis✨^blather00:11, 9 December 2025 (UTC)[reply]

Yes, I think the temp account thing was a band-aid solution to legal pressure on the WMF. I think if we can come up with something better that addresses the relevant legalities, we could probably get it implemented (I think I'm of the 'requiring registration' camp)Athanelar (talk)18:31, 10 December 2025 (UTC)[reply]

There was some kind of "leaked" memo over at Meta regarding challenging the community to accept more AI features. If you have sysop rights there you'll find it by inspecting my deleted contribs.ChildrenWillListen (🐄 talk,🫘 contribs)00:23, 9 December 2025 (UTC)[reply]

I'm not even extended confirmed yet, so I definitely can't view it, but that is incredibly concerning. Since Wikimedia is inherently a movement driven by its contributors and community, it seems to be another dangerous step of the WMF to negate this mission by concentrating their own power. Perhaps it should be proposed on Meta's Wikimedia Forum to make some larger change involving greater community election of board members so the WMF is more Wikimedian and isn't trying to thwart its own community.✨ΩmegaMantis✨^blather00:28, 9 December 2025 (UTC)[reply]

not surprising, given that they pushed the simple summaries feature through with the rationale of "editors will hate this but it'snot for them"Gnomingstuff (talk)03:37, 9 December 2025 (UTC)[reply]

I think "by humans, for humans" would be a great (if a bit cliché) tagline for Wikipedia to have somewhere.Athanelar (talk)18:17, 10 December 2025 (UTC)[reply]

The slogan that you're looking for is thatWikipedia is the free encyclopedia that only humans can edit. The trouble is that we often don't know who's doing the editing and so can't verify such a claim.Andrew🐉(talk)20:12, 12 December 2025 (UTC)[reply]

Okay, if for a moment we were to ignore the ideas where we welcome and accept AI content as part of Wikipedia's identity, what could we hypothetically do as a project to make it clear what separates reading Wikipedia from things like asking ChatGPT, searching Grokipedia, or using the Google AI Overview?Thebiguglyalien (talk)🛸05:59, 1 December 2025 (UTC)[reply]

Perhaps we can mention that it's "human-vetted" or "human-curated"? Even the AI-generated content is (usually) detected, and tagged or removed, rather quickly. However, Thryduulf also has a good point that many articles have at least some non-human input. –Epicgenius (talk)15:50, 1 December 2025 (UTC)[reply]

Even the AI-generated content is (usually) detected, and tagged or removed all we can say is that theproblematic AI-generated content is usually tagged and/or removed. Any AI-generated content that is stylistically similar to a Wikipedia article and which contains no errors (e.g. incorrect statements, non-existent references, etc) will almost always not be flagged because doing so wouldn't benefit the encyclopaedia. Accordingly it is impossible to know whether there have been 1 or 1 million edits of this nature.Thryduulf (talk)18:23, 1 December 2025 (UTC)[reply]

If an article contains no errors (e.g. incorrect statements, non-existent references, etc) do we actually care what wrote the text? Just asking on behalf of my human... —GhostInTheMachine^{talk to me}18:32, 12 December 2025 (UTC)[reply]

That's one of the major points of debate right now. Myself and some others would say yes, we absolutelydo care.

My thesis is thus; if we allow AI usage provided the output is Wiki-suitable, we will inevitably trend towards a higher and higher percentage of the encyclopedia being authored by AI, and I don't think that's desirable in the same way it would be undesirable if any large percentage of the wiki were authored by a single human person. It is agood thing that we have such a wide variety of authorisl voices.Athanelar (talk)18:37, 12 December 2025 (UTC)[reply]

if we allow AI usage provided the output is Wiki-suitable, we will inevitably trend towards a higher and higher percentage of the encyclopedia being authored by AI firstly, why? Secondly, if the output of AI reviewed by humans to the point that it is of the same standard as directly human authored, why is that differently good (or bad) than content directly written by humans?Thryduulf (talk)19:31, 12 December 2025 (UTC)[reply]

The 'why?' is the same reason it's happening to the rest of the internet. Just like it takes far less effort to pump out SEO-maximising AI listicles that are now dominating google search, AI-generated wikicontent would be much faster to produce than the human alternative. One needs only look at how widespread the actions of chronic AI-abusing individuals can get to see what I mean; the tracker for User:A Touch of Humanity still has some 60 articles that need review.User:Gnomingstuff has already hazarded a guess at the potential volume of AI text that might already be on Wikipedia, and it's concerning.

As for the second point, it's a philosophical more than ptactical thing. I don't think it'd be a good thing if, say, 20-30% of Wikipedia's text was authored by some individual human John Wikipedia, no mstter how good the content actually was. We're a communal project, that is inherently undermined if much of the new content is coming from a single authorial source.Athanelar (talk)19:55, 12 December 2025 (UTC)[reply]

If AI were producing Wiki-suitable output, why would a high rate be a problem?Levivich (talk)19:57, 12 December 2025 (UTC)[reply]

I think @Athanelar's point is that it's because it contradicts the idea of Wikipedia. The whole reason why there was a Wikipedia in the first place (instead of simply accepting any normal peer-reviewed, low interaction encyclopedia) is because it's a communal project since we believe in the strength of the commons, and that any monopolization of edits from a single source goes against that mission and founding idea, making it inherently anti-Wikipedian.then again, what do I know, I am not John Wikipedia✨ΩmegaMantis✨❦blather |☞spy on me20:37, 12 December 2025 (UTC)[reply]

I'm John Wikipedia, and I endorse this message.Athanelar (talk)20:49, 12 December 2025 (UTC)[reply]

Why is anti-Wikipedia a bad thing? Which is another way of asking: if a machine could do what Wikipedia can do, but faster, then why is that a bad thing? If we had a fast machine-written Wikipedia, and a slow human-written Wikipedia, and they both produce articles of the same quality, then what use is the latter?Levivich (talk)20:50, 12 December 2025 (UTC)[reply]

What you're essentially asking is "why would it be a problem if Wikipedia were fundamentally transformed into a completely different thing?" Which, like, I see your thought experiment, but I'm beginning with the assumption that we generally want to maintain the overall ethos of the project, and we don't want "the free (as in libre) encyclopedia that anybody can edit" to turn into "the free (as in gratis) encyclopedia which is mostly edited by a content engine developed by OpenAIet al which anybody can double-check the output of."Athanelar (talk)20:55, 12 December 2025 (UTC)[reply]

But you're not really answering the question, you're just restating the position: "Wikipedia is better than..." I'm askingwhy is it better? Why don't we want "free (as in gratis) encyclopedia which is mostly edited by a content engine developed by OpenAI et al which anybody can double-check the output of"? I think that's better, and I could tell you my reasons. Why do you think it's worse?Levivich (talk)21:07, 12 December 2025 (UTC)[reply]

Wikipedia's popularity resulted in the major decline of traditional written encyclopedias. If we assume that this decline was a rational decision by the people because of the merits of the Wikipedia project in contrast to traditional encyclopedias, and that at least one of those merits was its diversity of voices (likely so, as Wikipedia branded itself as the encyclopedia any one can edit) than that means that people like encyclopedias that are based off a diversity of voices. If we take @Athanelar's argument that AI destroys this diversity, then this means people won't like an AI-generated encyclopedia, and therefore it is not suitable for Wikipedia as it will cause us to lose readers. It's sort of anargumentum ad populum, but we certainly do depend on thepopulum for donations.✨ΩmegaMantis✨❦blather |☞spy on me21:16, 12 December 2025 (UTC)[reply]

Well, and it's also an argument from the idea that I don't want any more of the world's information to be monopolised by tech corporations than already is.Athanelar (talk)21:27, 12 December 2025 (UTC)[reply]

If we assume that this decline was a rational decision by the people because of the merits of the Wikipedia project... I think it's more likely that people just found it easier and quicker to look things up online than to go to a library to look in a traditional encyclopedia. (Sets of encyclopedias were not cheap and most homes didn't own a set.) Wikipedia is basically one-stop-shopping, the Amazon-equivalent of information.Schazjmd (talk)22:27, 12 December 2025 (UTC)[reply]

Yeah, that's probably right. The "rational"Homo econominus model is a quite inaccurate one. Still, I think it's hard to wave away how much the collaborative wiki model mattered and influenced perceptions of Wikipedia -- and probably of knowledge as a whole. It certainly made it novel and unique, and even now it still mostly is. That has to count for something.✨ΩmegaMantis✨❦blather |☞spy on me22:32, 12 December 2025 (UTC)[reply]

@OmegaMantis said:The whole reason why there was a Wikipedia in the first place...is because it's a communal project since we believe in the strength of the commons.

And here I thought Wikipedia was created becauseJimmy Wales' dot-com venture wanted to attract more men by hosting factual information about sports, cars, pornography, etc., andthe non-wiki version was moving too slowly to grab their intended audience.WhatamIdoing (talk)03:01, 14 December 2025 (UTC)[reply]

Fair, fair. Still, the community seems to have generally been motivated to keep up the project, and start the Wikimedia movement, based off some sort of collaborative ideals that stem from free software and free culture. These ideals pair much better with our encyclopedia than top-down, knowledge monopolizing overuse of large language models -- even if the latter may be sexier to the commercial eye, at least totoday'sdot-com ventures.✨ΩmegaMantis✨❦blather |☞spy on me03:22, 14 December 2025 (UTC)[reply]

Slightly relevant to this is today's donation appeal banner:December 4: Knowledge is human. We're sorry we've asked you a few times recently, but it's Thursday, December 4, and this fundraiser matters. We're nearing today's goal, but time's running out. If just 2% of our most loyal readers gave $2.75 today, we'd reach our goal quickly. Most people donate because Wikipedia is the internet we were promised: useful, free to use, and filled with reliable, human-created knowledge. If you agree, consider giving $25 or even just $2.75. So apparently the WMF is leaning into this kind of messaging. --LWG^talk05:52, 5 December 2025 (UTC)[reply]

WMF leans into whatever they think will get them donations, if they think AI hype will get more they'd mention that insteadmgjertson (talk) (contribs)20:40, 8 December 2025 (UTC)[reply]

I think this is a good idea, and it seems already integrated into the banner campaigns. I don't think Wikipedia has much advertising, though, so it'd be difficult to adapt our message when we don't really have one.Aaron Liu (talk)20:57, 9 December 2025 (UTC)[reply]

Based off @Athanelar's belief that Wikipedia is powerful as a content source not monopolized by one voice (as it would be if dominated by AI) I would advocate for the pitch of "Wikipedia remains one of the few sources of knowledge thatall humans, including you, can contribute to." This wouldn't make any claims about what content on Wikipedia is AI-generated or how it differs from bots, but basically rephrases the "anyone can edit" by emphasizing the control of humans (contrasting with AI) while implying that now, a lot of what we learn with AI we have little human control over at all.✨ΩmegaMantis✨❦blather |☞spy on me22:38, 12 December 2025 (UTC)[reply]

... make it clear what separates reading Wikipedia from things like asking ChatGPT ...

• Wikipedia: the free encyclopedia that's probably still more accurate than ChatGPT
• Wikipedia: the free encyclopedia that contributes less to climate change than LLMs
• Wikipedia: the free encyclopedia whose job hasn't yet been completely outsourced to AI
• Wikipedia: the free encyclopedia that doesn't write better than you
• Wikipedia: the free encyclopedia that doesn't talk back

I wrote those without the assistance of an LLM.

Wikipedia is like a big school book written by lots of teachers. Every fact has to be checked. All the teachers agree on what goes in the book. It explains things the same way for everyone.

ChatGPT is like asking a super-smart robot friend. It explains things in whatever way helps you understand. You can ask follow-up questions. It can give stories, examples, or simpler explanations. But sometimes the robot might guess wrong, so you still have to be careful.

Wikipedia is like a museum: Everything on display is curated, sourced, and labeled. You see stable information. You walk through and learn at your own pace. It does not answer you directly; you explore it.

ChatGPT is like a personal tour guide: You can ask anything: "Can you explain that again, but simpler?" The guide adapts to your interests. It connects ideas across rooms ("Here’s how this painting relates to that sculpture.") But occasionally, the guide might misremember or over-explain something, so you verify if it matters.

The table and everything after it was written by ChatGPT.Levivich (talk)20:59, 11 December 2025 (UTC)[reply]

These seem to be catered to the audience that is already LLM-skeptical. But most average people do not necessarily share an LLM-skeptical view or care about the ethical aspects of data centers. That is why ChatGPT and Gemini are growing and slowly eating the rest of the internet's lunch. The confident mistakes is important, but Wikipedia can also confidently report a hoax for years.Andre🚐21:23, 11 December 2025 (UTC)[reply]

The proposal would be misleading. If you say "human-written" the layman would understand "written by a single and definite human", ideally someone the layman already knows and trusts (like a newspaper article or opinion piece). Wikipedia is not AI but, for the layman who ignores our internal procedures, it's something in-between: it's not AI (at least not what the layman understands for AI, a chatbot or a similar service that generates content in answer to a query), but it's not something written by an identifiable someone either. It's stuff written by several nobodies with usernames.

Also, "human-written" is only a virtue in the eyes of people with a strong anti-AI sentiment. There is a sizeable group of people like that, but they are not everybody. There are people wholike AI (some even to insane degrees), and others who just don't care, and just think of asking "who is this guy?" to ChatGPT as a more evolved version ofgoogling it. If everybody was as anti-AI as the anti-AI guys pretend, AI would not be the success it is.Cambalachero (talk)13:47, 12 December 2025 (UTC)[reply]

If you say "human-written" the layman would understand "written by a single and definite human" I don’t see how. The lack of any article (a, the…) makes it indefinite and plural, as it is understood in both academic and common speech.Aaron Liu (talk)17:22, 12 December 2025 (UTC)[reply]

Being written by "an identifiable someone" isn't necessary for something to be written by humans. One doesn't look atBeowulf and think "Oh, dear, the author is unidentifiable, so it might have been written by non-humans".WhatamIdoing (talk)19:31, 12 December 2025 (UTC)[reply]

There was no AI back in the 700 ADCambalachero (talk)22:51, 12 December 2025 (UTC)[reply]

I agree. But I think that it is a strong counter-example to the idea that knowing the identity of the author ("written by a single and definite human"...written by an identifiable someone) is what makes people think that something is human-written.

So, what is human-written? I think "human-written" encompasses works (like Beowulf) for which the author is unidentifiable and also those for which there is more than one author. "Human-written" indicates that humans (not AI, notmonkeys, notMartians) do a particular thing (write, which I distinguish here from other activities, such as "prompt a chatbot to write" or "copy and paste" or "prettyprint the wikitext" or even "add citations to", though some of those are sometimes desirable activities). Wikipedia is (still, or at leastmostly still) human-written because it is written by humans (Monkey selfie excluded). It might be difficult to draw an exact line between how much automation is possible before it stops being human-written (see alsoShip of Theseus), but I think "human-written" is a fair description of both what we have now and what we want for, say, the rest of this decade.WhatamIdoing (talk)23:37, 12 December 2025 (UTC)[reply]

The human chauvinism is really uncalled for. No species other than the Homo Sapiens has human-like intelligence (except perhaps other long extinct hominids), there is no evidence that alien civilizations even exist, and artificial general intelligence is still firmly in the realm of science fiction.All knowledge is human knowledge,all writing is human writing, and that includes writing generated by AI.Cambalachero (talk)20:52, 13 December 2025 (UTC)[reply]

I don't think that "writing generated by AI" is written by humans.WhatamIdoing (talk)03:05, 14 December 2025 (UTC)[reply]

As pointed out previously, "human-written" is incorrect because of bots that have been part of Wikipedia for over 20 years and long-predate AI.Thryduulf (talk)19:33, 12 December 2025 (UTC)[reply]

I think I would accept some of Wikipedia's "bot-created" content as being human-written in the end: If the bot is filling in the blanks to create a sentence like "$Name is a city in $State with a population of $population as of the 2000 US Census", then a human is still significantly responsible for it. That's closer tomail merge for a form letter than to bot-created.WhatamIdoing (talk)23:29, 12 December 2025 (UTC)[reply]

If your criteria is "a human is significantly responsible for it" then an AI prompted by a human, trained on information written by humans, and with output fully reviewed by humans must also count as human written.Thryduulf (talk)23:42, 12 December 2025 (UTC)[reply]

No, I don't think so. In the case ofUser:Rambot geography articles, we can name a specific human (User:Ram-man) who created the blanks for the bot to fill in and prepared the list of the exact things that the bot was to put in those blanks.

In the case of AI, I'm not sure that "written" is even the correct verb (maybe "assembled"?), but it's not really a human writing it. I'm not sure where you would draw the line between the total human control of Rambot-like scripts and AI, but at some point, it stops being human written.WhatamIdoing (talk)03:08, 14 December 2025 (UTC)[reply]

I would not say that bots have beenwriting content. They've been reverting edits and making technical and style fixes.Aaron Liu (talk)03:33, 13 December 2025 (UTC)[reply]

They also generated many wholesale articles about small localities.Andre🚐03:34, 13 December 2025 (UTC)[reply]

Okay, I would concede on that point. It irks me, though, that such a low-visibility area stops us from claiming "human-written"...Aaron Liu (talk)03:36, 13 December 2025 (UTC)[reply]

It's not just small localities, although that's the greatest number, there have been articles about species and multiple other topics too. Just because it's low visibility to you doesn't mean it's not important, or not high visibility to someone else (I'd be surprised if most readers don't look up articles related to their local area at some point). Thirdly, fixing spelling and grammar errors, updating templates, fixing links, and many other small tasks that bots do (solo or in conjunction with a human) are also very much part of writing an encyclopaedia.Thryduulf (talk)04:46, 13 December 2025 (UTC)[reply]

On third: I agree that they are vital work, but I wouldn't call them "writing".Aaron Liu (talk)19:45, 13 December 2025 (UTC)[reply]

There's little scope for creative writing on Wikipedia because the articles are supposed be entirely derivative, being based on the writing of others, and presented in a bland, dispassionate and formulaïc style. This is perhaps why forums like the Village Pump are so popular – they enable editors to express themselves more freely.Andrew🐉(talk)20:10, 13 December 2025 (UTC)[reply]

A major bot editor of Wikipedia isCluebot NG which has made over 6 million edits and counting. This explicitly uses anartificial neural network which has been weighted with training data. This seems to be much the same as the new AIs and its scope includes all articles not just the cookie-cutter stubs created by bots likeRambot.Andrew🐉(talk)08:12, 13 December 2025 (UTC)[reply]

That's somewhat true, in that GPT is also a type of a neural network, a more specialized type for generating text. The problem is when people use it as a reasoning tool instead of a fancy autocompleter or a search tool. It is searching to come up with something, and not really thinking. Cluebot is trained with edits as data and is returning basically a binary decision, what is the score of likelihood that this is a bad edit and I should revert it. Therefore it works well only because some common types of vandalism look similar. But it doesn't actually read or write articles.Andre🚐08:22, 13 December 2025 (UTC)[reply]

Cluebot is performingrecent changes patrol and this requires both reading and writing. Few human editors read or write entire articles; much of the activity is piecemeal.

Looking atan example of a recent edit by Cluebot, note that the article source looks more like code than English prose. That's because of the heavy use of markup, templates and tables. Many ordinary humans would find this incomprehensible. Wikipedia is not just some English text; the whole thing is a complex bundle of software.

Note also that the vandalism was made by a temporary account and so we don't really know who or what did that. But the trusted version to which Cluebot reverted, was created byCitation bot. In this case, the bots seem to have more presence and standing than the putative human.

Andrew🐉(talk)08:48, 13 December 2025 (UTC)[reply]

I'm quibbling a bit but Cluebot doesn't really write the article. It reads the diff, determines it looks like vandalism, and reverts. Citation bot only modifies the URL inside the citation. In neither case are they actually composing article text.Andre🚐20:39, 13 December 2025 (UTC)[reply]

That's editing rather than writing, but humans often do this too. And such bots can be versatile – an earlier version of Cluebot created thousands of articles such as1803 Zwicky.Andrew🐉(talk)20:59, 13 December 2025 (UTC)[reply]

That's true. More in the vein of the Rambot type stub articles.Andre🚐22:07, 13 December 2025 (UTC)[reply]

I don't think that reverting is "writing" content.WhatamIdoing (talk)03:10, 14 December 2025 (UTC)[reply]

To this I would add that the assumptionbut we've reached the point now where we really need to be looking into how we frame Wikipedia's relationship with AI, especially in public-facing areas is unfounded. I don't know why there is such huge worries and people describing the situation as complete chaos with Wikipedians scrambling to manage the huge flood of AI problem and stay relevant.
People basically know Wikipedia is written not by AI but by humans.
The WMF and news orgs currently reporting on Wikipedia communicate over and over that Wikipedia stays human.
Lots of other websites are also written by humans, just because LLMs are there now doesn't mean it's soon all just AI texts. Basically nobody uses Grokipedia. So reone of its defining characteristics that's also kind of false because people don't expect it to change otherwise and lots of other text websites and text are also still written by humans.

Andrew, Cluebot NG doesn't write content but just reverts edits so it doesn't really add content.Prototyperspective (talk)21:08, 13 December 2025 (UTC)[reply]

Cluebot II created about 20,000 articles; that's plenty of content.Andrew🐉(talk)22:01, 13 December 2025 (UTC)[reply]

Not to go off-topic, but I find it weird how this (and many other discussions about AI use on WP) seem to laser-focus on edge cases ("how can we say we should limit LLM use when we've always used bots?") while totally ignoring the big picture. The claim that"human-written" is only a virtue in the eyes of people with a strong anti-AI sentiment may or may not be true, but is ignoring the fact that people who come to WP choose not to use AI, which is now unavoidably built into every Google search and thus more easily accessible. I don't much care to defend the idea that WP isbetter than some hypothetical future LLM output, only that it'sdifferent, and we should be able to offer such an option to those who want it for whatever reason. Should black licorice be pulled from store shelves and replaced with chocolate because "the taste of black licorice is a virtue to only those few people who hate the taste of chocolate"?

To those who say that AIscan write brilliant articles at scale...well, sure, daily Coca-Cola consumptioncan be a part of a healthy diet and lifestyle. But look at Facebook, look at LinkedIn, look at the dozens of listicles that CNN is trying to pass as news on their homepage. They'reutter total predictable garbage at a machine-gun pace. The ability of AI to make good things at scale isvastly dwarfed by its ability to make garbage at scale. Why are so many people here focusing on thecan and themight and theif it's done right while ignoring the internet that is right in front of their faces, from which WP is an increasingly rare island of refuge? Isn't that worth preserving?WeirdNAnnoyed (talk)22:00, 13 December 2025 (UTC)[reply]

Faulty example. Rather than pull black licorice from store shelves, the correct analogy would be to advertise it as "not chocolate". Would would care about such detail, other than those who do not want chocolate?Cambalachero (talk)22:15, 13 December 2025 (UTC)[reply]

That's also a faulty example, but only because the whole premise of both examples is faulty too.

Why are we afraid to publicly say what's actually going on? Who are we trying to defend or appease?

I propose adopting the following attitude, even if we don't literally go for my new tagline:

Wikipedia: Our editors are 100% real, because AI is bullshit.TooManyFingers (talk)02:10, 14 December 2025 (UTC)[reply]

With how popular explanatory footnotes are, a feature in the section of the visual editor citation button for creating footnotes could be pretty useful. A section to the visual editor link button for reusing previous links could be useful considering how many times I find myself linking to the same article. A more secondary visual feature is that instead of citations next to each other being distinct like [1][2], they could be merged like [1,2].Misterpotatoman (talk)07:07, 29 November 2025 (UTC)[reply]

Like the idea on footnotes.

For reusing previous links, you just need to type '<ref' where you want to put your source in the Visual Editor, and then a pop-up would automatically appear where you would get 3 options 'Automatic', 'Manual' and 'Re-use'.

Merged citations [1,2] would be too close for comfort, and could result in mis-taps on smaller handheld devices, alsoWP:AINT.Cdr. Erwin Smith (talk)12:46, 29 November 2025 (UTC)[reply]

If you'd like to see merged footnote markers, then seew:fr:Moineau mélanure#Description. The proposal is similar to the style used at the French Wikipedia.WhatamIdoing (talk)22:13, 29 November 2025 (UTC)[reply]

That requires manually insertingfr:Template:, between each<ref>.jlwoodwa (talk)04:15, 30 November 2025 (UTC)[reply]

no, i mean reusing links as the wikipedia feature that let's you link to links, in not talking about citations, also i think if it was merged, it should pull up a screen where all the citation links appear, i think it will actually make it easier on smaller devices.Misterpotatoman (talk)22:32, 29 November 2025 (UTC)[reply]

It sounds like you're thinking about the scenario in which I go from one article to the next to add a link to (for example)Rare disease (real example, BTW), and instead of clicking the link button and typingrare dise in the search box until it pops up the link to the correct article, it would have a list of the most recent ones I've added links to, and I could just click on one of those instead of typing.

As someone who never edits from a smartphone, this would not be efficient for me. But since you're on mobile, where IMO typing is practically impossible, let me ping @PPelberg (WMF) and ask him to please make a Phab task suggesting this new feature idea for themw:mobile visual editor.WhatamIdoing (talk)23:53, 30 November 2025 (UTC)[reply]

exactly, that's what i meant, and thats even better than my first idea for link reuse.Misterpotatoman (talk)23:58, 30 November 2025 (UTC)[reply]

“recently created citations” would be a goodMeta:Community Wishlist entry ~ 🦝Shushugah (he/him • talk)23:26, 9 December 2025 (UTC)[reply]

FWIW you can just click on the "Cite" buttonAaron Liu (talk)20:29, 11 December 2025 (UTC)[reply]

Imagine that you're adding the same two or three sources to multiple articles. The "Cite" button will let you re-create the source each time (hand-correcting each time whatever it gets wrong). You could alternatively copy/paste the citations between articles. But I believe the request is for something like:

1. Click the "Cite" button
2. See everything that's there nowplus a short list of the last few citations you generated (including the last few that you used in other articles).
3. Pick one from the pre-loaded list of recently used citations.

WhatamIdoing (talk)00:16, 12 December 2025 (UTC)[reply]

Yeah, that seems about right. I was pointing out to @Cdr. Erwin Smith that you can do the same without having to type out the ref syntax.Aaron Liu (talk)17:24, 12 December 2025 (UTC)[reply]

The 'Citation' button beside the 'Link' button looks much more like aQuotation [" ] button than a reference one. Me being an erstwhile Quora user assumed it as such :DCdr. Erwin Smith (talk)18:59, 12 December 2025 (UTC)[reply]

Yeah me too.c: User:Jack who built the house/Convenient Discussions even uses it as a quote button!Aaron Liu (talk)19:02, 12 December 2025 (UTC)[reply]

It would be nice to have some essay or guideline or such forWP:GPTZERO/WP:AIDETECT software. As is, there is a good understanding that such software is highly error-prone and subject to an unacceptably high false positive rate, and yet they are also regularly used as additional evidence, often with other signs. I myself have used it as evidence sometimes, though interpretation of such output remains highly subjective.

There seems to be an exponential rise in AI conduct reports at ANI[9], so having more guidance seems useful.
I saw we still lack a useful metric for definitively determining AI usage, but this seems like an easier question to solve, and I think one the community may already have a good idea on.In what circumstances are AI detectors useful, and when should they not be allowed?User:Bluethricecreamman(Talk·Contribs)16:55, 30 November 2025 (UTC)[reply]

I've been thinking about it a bit since i wrote this comment from a report I filed a bit ago here:but i think determining AI is like diagnosing a rare disease, the probability of AI given any one sign is low, but the conjunction of multiple signs, previous use of AI, hallucinated URLs, and human judgement is important to determine AI usage. even gptzero is useful here, though its high FPR should be understoodUser:Bluethricecreamman(Talk·Contribs)16:59, 30 November 2025 (UTC)[reply]

GPTZero is actually a very good indicator when percentage is high. You can't really "disallow" it. I think people who dismiss it out of hand are missing the boat. Of course it can make false positives, but I've never seen it not detect AI that was in fact AI: a high percentage is a significant data point. Also they keep improving the algorithm it only keeps getting better. --GreenC17:24, 30 November 2025 (UTC)[reply]

Automatic AI detectors can have false negatives, but generally you have to dosome fine-tuning of the output to get it there. (very depressing study btw)Gnomingstuff (talk)20:58, 30 November 2025 (UTC)[reply]

Controversial opinion, but I think if there's considerable human editing in an AI output, it should be allowed. We should alsoAGF in this whole process.

As such, if the final result of these Anti-AI tools is Unclear/Mixed/Medium/<50% probability of being written by LLMs, we should favour the editor in our verdict. Ofcourse the final verdict should be on the hands of an actual experienced human.Cdr. Erwin Smith (talk)08:09, 1 December 2025 (UTC)[reply]

My opinion has changed. In general if AI is used responsibly so as to not violate wikipedia policies its not technically prohibited. It just so happens, though, that many editors are using it blatantly, without disclosure, and in ways that do violate our policies.

We really shouldnt be trying to identify every bit of AI, an obviously impossible goal. We should consider AI usage an aggravating factor (WP:CIR) when considering other policy violationsUser:Bluethricecreamman(Talk·Contribs)14:18, 1 December 2025 (UTC)[reply]

I disagree -- a lot of what AI detectors are checking are things unrelated to Wikipedia policy. If an editor generates text with AI, does not verify the claims, but does change the sentence structure enough to get it to 50%, then the core problem has not been addressed.Gnomingstuff (talk)14:36, 1 December 2025 (UTC)[reply]

This is exactly why I said that the final verdict should rest in the hands of an experienced human.

Also, LLMs are getting better rapidly with every passing generation, and can not lie like Humans. If you instruct them to write a Wiki article, they will write so abiding by most, if not all the existing policies.

Ofcourse, they can still make some errors, and that's why we Humans are here to weed such articles out!

So basically,

• >50% AI Probability > Insta-discard
• ≤50% AI Probability > We decide with leniency

Cdr. Erwin Smith (talk)19:06, 1 December 2025 (UTC)[reply]

Given the error rate of most of the checkers they should pretty much never be used on their own and when they are used taken with a huge grain of salt, while keepingWP:AGF in mind. So with that I wouldn't be opposed to that being banned as a main resource on the topic.PackMecEng (talk)17:48, 30 November 2025 (UTC)[reply]

Giventhe error rate of most humans who claim to have spotted AI use, we shouldn't rely on human detection, either. The accuracy rate is barely better than a coin toss for editors like me (who don't use AI tools regularly). This study says thatpower users of generative AI tools may "only" be mistaken 10% of the time. I wonder how that compares to the common AI-detecting tools?WhatamIdoing (talk)00:58, 1 December 2025 (UTC)[reply]

Oh for sure, people are not great at it either. All options kind of suck at the moment.PackMecEng (talk)01:44, 1 December 2025 (UTC)[reply]

Sometimes a duck is a duck, but I agree with the general sentiment that we are stuck between a rock and a hard place hereKatzrockso (talk)08:14, 1 December 2025 (UTC)[reply]

That 10% error rate is mostly false negatives (that is, sometimes AI can slip past even expert eyes) and included LLMs that were specifically tuned to defeat detection. The consensus opinion of the experienced humans in that study correctly identified 99.3% of AI-writted articles as AI, and never once falsely identified human-written text as AI. Quoting from the articleDespite our best efforts to generate articles that our experts would find undetectable, most of their detection rates remain largely unchanged from prior experiments, and the expert majority vote is again perfect on all 60 articles. --LWG^talk16:46, 1 December 2025 (UTC)[reply]

The other thing that worries me about this report is that the "expert" is someone who uses ChatGPT or other LLMs themselves, a lot. It's not the person who thinks they're an expert, or who (to use our context) spends their day accusing other people of using LLMs and therefore develops thepossibly mistaken self-impression of their own expertise.WhatamIdoing (talk)03:27, 5 December 2025 (UTC)[reply]

I get the impression that many, possibly most, editors who are accusing (correctly or otherwise) others of using LLMs (almost) never (knowingly) use LLMs themselves.Thryduulf (talk)03:58, 5 December 2025 (UTC)[reply]

They should certainly never be used on their own, as they have a high rate of both false positives and false negatives. In conjunction with other signs they can be interesting but never particularly useful as whenever there are enough other signs that you can trust the output those other signs are enough to be determinative on their own. I'd support a page explaining this.Thryduulf (talk)18:43, 30 November 2025 (UTC)[reply]

One sign that I think may be helpful is "paste check" (detecting when a new editor copy/pastes a large block of text into an article). However, ittriggered only a couple of times recently, and it's not specific to AI. It could be copyvios, someone pasting in a valid quotation, or other things.WhatamIdoing (talk)01:01, 1 December 2025 (UTC)[reply]

Some folks seem to also prefer using google docs for creating articles, the mad lads.User:Bluethricecreamman(Talk·Contribs)01:26, 1 December 2025 (UTC)[reply]

From what I understand, paid AI detectors are much better than the free ones in terms of accuracy, at least for the kind of generic, non-fine-tuned output that we are probably getting here, and are very good at this point. Pangram seems to be the best-performing.

That said I don't personally use AI detectors, if only because of optics -- I don't have the patience to deal with endless "well automatic AI detectors get it wrong so please take your ugly tag offMY beautiful writing."Gnomingstuff (talk)20:57, 30 November 2025 (UTC)[reply]

Note: I've gone ahead and boldly made a section onWP:AISIGNS forWP:GPTZERO. Feel free to add collective wisdom.User:Bluethricecreamman(Talk·Contribs)05:20, 1 December 2025 (UTC)[reply]

Two thoughts:

1. Even though no tool is perfect, in my experience tools like GPTZero are very accurate at detecting raw LLM output that was copy-pasted into Wikipedia. In thestudy WhatamIdoing linked GPTZero had a 0% false positive rate on every test set except the one where the LLM was specifically fine-tuned to defeat AI detection. I have never yet seen GPTZero return 100% AI on human-written text or 100% human on copy-pasted LLM output. So for our use case, where our primary concern is novice users who copy-paste large quantities of slop, we can expect the tool to be helpful, and it would be counterproductive to tell people to ignore it.
2. What are the consequences of the tool being wrong? If the tool gives a false-negative, the result is that we fail to detect AI content, which is the same outcome as if we don't use the tool at all. If the tool gives a false positive, the result is that we incorrectly believe content to be AI-generated, possibly leading to the content being reverted or the editor being asked to explain and justify their contribution. But if the content is not in fact AI generated, then all the editor needs to do is accept theirWP:ONUS and acquire consensus for inclusion of their content, which is the same as the normal wiki process.

So basically, I don't understand what harm we are trying to prevent by discouraging editors from using AI detection tools to assess content. --LWG^talk16:28, 1 December 2025 (UTC)[reply]

The harm is false positives leading to harassment and sanctions. Also the study notes that humans were far more successful than AI detectors in this. So let's not give a not terrible tool amd say its accurate. The study does not fully support that and I do not understand the help it gives on its own given its well know deficiencies and the harm it can easily cause to our most vulnerable user base.PackMecEng (talk)18:29, 1 December 2025 (UTC)[reply]

How are we defining "harassment and sanctions" here?Gnomingstuff (talk)00:44, 2 December 2025 (UTC)[reply]

The usual way.PackMecEng (talk)13:56, 2 December 2025 (UTC)[reply]

So if I understand correctly, you are concerned that a new editor who does not use AI might still write edits that GPTZero identifies as AI, and that would cause that editor to be inappropriately blocked, or to become a target ofWikipedia:Harassment? That seems unlikely, since editors aren't normally blocked based on one bad edit with no chance to explain themselves. If someone just hates new editors and decides to falsely accuse one of them of using AI without giving them a chance to explain themselves, the accuser will getWP:BOOMERANGed. --LWG^talk14:51, 2 December 2025 (UTC)[reply]

it is naturally stressful to deal with claims that your work may be AI. it sounds like an insult, it makes one wonder why an editor is picking a fight, trying to involve additional folks through any noticeboard may escalate the situation.

its not even about accusing, its naturally stressful to be randomly flagged. we should ofc ask and investigate as appropriate, but we should not be doing a giant fishnet unless theres broadly more policy viosUser:Bluethricecreamman(Talk·Contribs)15:11, 2 December 2025 (UTC)[reply]

If by fishnet you mean running some sort of broad scan of all new contributions and reporting every hit to ANI, then I agree with you. But I think choosing to engage in this community means exposing your writing to scrutiny, and if the stress of having to explain your contributions is too much for you, then a collaborative encyclopedia that seeks to have high sourcing standards is probably not the place for you. If a new user contributes large quantities of text without explaining themselves, I think it's reasonable to run their edits through GPTZero, especially if subjective tells of AI writing are present. If GPTZero returns a high AI percentage, I think it's entirely reasonable to reach out to the editor asking for an explanation, and to remove the content if no satisfactory explanation is given. We aren't under obligation to give content the benefit of the doubt here, theWP:ONUS is on the contributor. False content is much more harmful to the Wiki than missing content, since missing content can always be added eventually, but false content calls the entire rest of the wiki into question. It's also much easier to identify and add missing content than to identify and remove false content. --LWG^talk15:32, 2 December 2025 (UTC)[reply]

Actually I think you made a very good point without realizing it. AI generated content is not in and of itself a reason to revert something. That falls to if it is poorly or falsely sourced. If you are reverting because it was a large block of content and you ran it through a dubious AI detector with it coming back positive, you need more than that to revert it otherwise you are the problem there. That seems to be the general rub, blanket this is bad and going after people as you just described is the problem we are talking about. Heck there was even a recent ANI thread where someone was trying to mis-apply G5 and G15 and finally then when that didn't fit try to IAR just because it was AI.[10]PackMecEng (talk)17:19, 2 December 2025 (UTC)[reply]

Yes, some people take the position that all AI-generated content is inherently bad and has no place here, whether for ethical or copyright or content quality or general project vision reasons. That's not my position, I'm with you that the problem with LLM content is that itfrequently fails other policies, however it's also my position that we don'tneed LLM content, and we're currently facing a flood of bad LLM content that is overwhelming our normal mechanisms for dealing with bad content, so if this community can't find any way to navigate between the two slippery slopes here then I'd rather slide down the one that leads to no AI. --LWG^talk17:46, 2 December 2025 (UTC)[reply]

This is a good way of putting it. LLMs completely invert the effort difference between writing and reviewing. The issues that led to theWP:MASSCREATION policy, but possible with any text anywhere.CMD (talk)02:10, 3 December 2025 (UTC)[reply]

It's also incredibly bad optics, especially when WMF is currently in the middle of an advertising drive about how Wikipedia is thegreat human alternative in the age of AI -- an advertising drive no doubt directed at prospective donations from people who support Wikipedia forexactly that reason -- when in reality a substantial amount of articles have quietly been AI-generated in part or full for several years. I'm actually kind of shocked that the media hasn't picked up on the fact that Wikipedia has only just now gotten around to creating real AI guidelines, given the response to the Simple Summaries debacle earlier this year.

So yes, weabsolutely should be doing a "giant fishnet" to determine the extent of the problem. If we had started doing that in November 2022 like we should have, then it wouldn't be a "giant" undertaking, but we didn't, and so now it is.Gnomingstuff (talk)03:04, 3 December 2025 (UTC)[reply]

Not really though, because if it has not been a problem for years and most dont know it was AI generated why remove it? Seems counter to bring here to build an encyclopedia.PackMecEng (talk)13:58, 3 December 2025 (UTC)[reply]

Just because no one noticed a problem until recently doesn't magically make it not-a-problem. To take an extreme example,blatant vandalism has sometimes undetected for 10+ years.Gnomingstuff (talk)17:10, 3 December 2025 (UTC)[reply]

The point is that if you read a given bit of text and there are no problems with it, either stylistically or factually, then it does not become a problem when you find out it was (or might have been) written by (or with the assistance of) an AI.Thryduulf (talk)17:21, 3 December 2025 (UTC)[reply]

And my point is that just because no one changed a piece of text doesn't mean there are no problems with it.Gnomingstuff (talk)23:51, 3 December 2025 (UTC)[reply]

That's a strawman argument. I specifically said if thereare no problems with a given bit of text, not that there were problems which hadn't been noticed. Text being AI-generated is not a problem in and of itself. Itmight contain problems, for example stylistic errors, factual errors, non-existent references, etc, but it is possible for every single one of those problems to also be present in human-written text. The different types of problem occur at different frequencies in differently-originating text (AI-generated text is very significantly more likely to include meta comment, human-generated text is very significantly more likely to include spelling errors) but the only type I can think of that only ever appears in one but not the other is copy-paste errors (e.g. copying one too few characters) and that's a mistake only humans make (although LLMs can obviously propagate such errors I'm not aware they can originate them). In at some circumstances LLMs are more likely (but not guaranteed) to produce text with issues than an equivalent text produced by humans (a 1000 word submission is more likely to contain issues than a 10 word submission, regardless of origin), but such problems are identifiable specific things not the mere fact of being written by AI. That is to say that the problem with a text containing a non-existent reference is that the reference does not exist, not that it might have been written using AI.

Text that objectively contains no issues when assumed to be human-written still contains no issues when alleged (or even proven) to be LLM-generated.Thryduulf (talk)04:10, 4 December 2025 (UTC)[reply]

There's several problems with this line of argumentation, which you've repeatedly expressed in every AI discussion I've seen you in. I recognize that your position is coherent, but the lack of recognition from you that other positions are also coherent is getting tiresome.

1. It's clear from these discussions that for many people (both editors and readers) text being AI-generatedis a problem in and of itself, whether for ethical grounds due to the provenance of the technology or the economic dynamics of its implementation, or for legal concerns about the still-developing copyright landscape in the field, or for philosophical reasons about the overall vision of our project, or for whatever other reason.

2. Even setting that aside, AI text is still qualitatively different than human-written text in that the authorship is different, and authorship can change the acceptability of text totally independently of content, seeWP:NOSHARE andWP:COIEDIT. So it's not automatically a given that all edits can be judged purely by the bytes they contain.

3. Even settingthat aside, in the real world we never actually get your hypothetical "text with no problems in it", because our ability to assess text is not perfect. All we get is text with noknown problems, which is acceptable if the text has has adequate scrutiny. Unfortunately, our resources for scrutinizing text are dramatically inadequate to the scale of the task, so we constantly prioritize our attention with various heuristics. Because the types of errors that tend to come up in AI text are different, the type of scrutiny they require also tends to be different, so knowing whether text is AI generated may change whether we feel it has received the scrutiny it needs.

All three of those are very valid reasons why text that we would accept when written by a human might be rejected or subjected to additional scrutiny if we later discover it was written by an AI, and even from your position as I understand it point 3 should motivate different treatment of AI content. --LWG^talk05:12, 4 December 2025 (UTC)[reply]

The only problem with that is that itisn't coherent to say that an identical string of characters is unproblematic when human-written but problematic when (partially) LLM-written.Thryduulf (talk)13:46, 4 December 2025 (UTC)[reply]

@LWG, my experience with AI detection toolswas very different from yours. Maybe they've improved since then?WhatamIdoing (talk)03:35, 5 December 2025 (UTC)[reply]

• I think many editors are quietly using AI as a co-worker. They are careful about following policy, verifiability. They use GPTZero and other tools to check their work and copyedit. There is considerable human involvement. It's not "AI generated", it's something else, it's a mixture. We don't have a good name for this, most discussions revolve around the worst case scenario of a chatbot-cut-paste-save. We would be fooling ourselves to ban AI entirely, and when used appropriately, what difference does it make, it's part of a complex process of humans and machines working together. Statistical fuzzy matching algorithms are the basis of spell checkers and search. They are often incorrect and cause problems. We still use them because humans are in the loop error-checking. --GreenC17:53, 3 December 2025 (UTC)[reply]

  If they edit the model output to unrecognizability or don't use it directly then it won't be detected at all and accusing AI use at that point would be frivolous and aspersive (?) without more evidence. ~212.70~~2025-31733-18 (talk)18:08, 3 December 2025 (UTC)[reply]

  At least some LLM-detectors (both human and non-human) flag non-LLM text as being AI (a false positive). I've seen (both on Wikipedia and elsewhere) humans who suspect someone of using an AI repeatedly hound that person if they do not admit to using AI - regardless of whether they have actually used AI or not. This is exactly as unacceptable as hounding an editor for any other reason.Thryduulf (talk)19:27, 3 December 2025 (UTC)[reply]

I realize this is a completely different approach, but maybe we need a tool that detects not LLMs but hallucinations, fake sources and the like. Such a tool would cover the problems that LLMs cause just as well.Jo-Jo Eumerus (talk)08:01, 4 December 2025 (UTC)[reply]

Yes, fake sources are a problem. Whether they are added by and AI or human is irrelevant.Thryduulf (talk)13:47, 4 December 2025 (UTC)[reply]

Which is a good reason right there to look for a tool that detects fake sources, rather than LLM style.Jo-Jo Eumerus (talk)14:13, 5 December 2025 (UTC)[reply]

issue is also that sometimes folks are using real citations, but using facts that are hallucinated and not validated from it. repeating my comment from aWP:ANI report here[11]his diff triggers an AI hit. the source exists, but its about teaching reading to children and phenome recognition, not about speed reading. the 400 words per minute figure doesn't appear as far as I can tell.User:Bluethricecreamman(Talk·Contribs)16:11, 5 December 2025 (UTC)[reply]

That's aWP:V problem, whether AI was used is irrelevant.Thryduulf (talk)17:00, 5 December 2025 (UTC)[reply]

agreed. just mean to point out that even if citations are real, facts may still be hallucinated. hallucinated sources could be trivial to check, but other more subtle forms may require a human.User:Bluethricecreamman(Talk·Contribs)17:15, 5 December 2025 (UTC)[reply]

In exactly the same way as a human edit needs to be checked. References that don't exist may be due to hallucination, minor errors (e.g. transposition of digits in an ISBN), major but still fixable errors (e.g. reusing the wrong source, citing the wrong chapter), not having checked the source actually says what the summary/snippet says it does, cluelessness, intentional misleading by a human, and possibly other things. Only a human can reliably determine which it is (a machine will be able to resolve some errors in some cases but not more than that).Thryduulf (talk)18:57, 5 December 2025 (UTC)[reply]

We have the pending pages review mechanism for semi-protected pages. Perhaps, instead of blocking an editor, their edits could be automatically flagged as a pending change. It would stop them as effectively as a block and it would create an opportunity to educate them. It would allow them to continue the dialog.

I see this option being selected mainly for new editors.Constant314 (talk)18:00, 1 December 2025 (UTC)[reply]

With the current pending changes implementation, an article with pending changes protection continues to have a single, linear history of changes. Implementing a way to flag an individual edit as requiring review, while still allowing others to make changes that are visible to non-logged in readers, would require implementing a branching history, and would require someone to merge the pending change if approved. It would be more complex for the editor in question to make successive unreviewed pending changes to the article, as they would have to understand the branching model. It would be a significant amount of development effort, changing fundamental aspects of how articles are stored and edited.isaacl (talk)19:48, 1 December 2025 (UTC)[reply]

Good-faith editors who ignore repeated talk-page warnings are often partially blocked from article space, which forces them to engage in discussion with no restrictions on where the discussion occurs. How would this proposal be more effective than a partial block?

If an editor fails to change in response to repeated talk-page warnings for edits that had to be reverted, how would this restriction convince them to stop editing disruptively, given that they can still edit almost as freely as before?

Pending changes is mainly intended to filter blatantly inappropriate drive-by edits such as vandalism and spam. If a user is "pending-changes restricted" for a subtle or complex issue, would reviewers be expected to check for that issue?Helpful Raccoon (talk)04:45, 2 December 2025 (UTC)[reply]

I'm new here, so if this idea has already been discussed or should be posted elsewhere, please tell me. I think that Wikipedia's cite tool, the one that comes up when you press the Cite button on the visual editor, should be improved to allow uploading a .RIS file and automatically populating all of the necessary fields based on that. I got the idea by using tools like Scrible, which have this functionality. I was thinking you could just implement this as an upload button in theAutomatic tab of the Cite button, but I would be happy if it went anywhere in the tool. Unfortunately, I have no idea how to code this, so I would need a lot of help making it to theWP:VPR. Hopefully we can make this work.Mxwllhe (talk)17:23, 2 December 2025 (UTC)[reply]

More information onRIS.Mxwllhe (talk)17:34, 2 December 2025 (UTC)[reply]

Pinging @Quiddity (WMF) to figure out who's dealing withmw:citoid these days.WhatamIdoing (talk)03:37, 5 December 2025 (UTC)[reply]

There's an existing older task (T128405) that I've now linked to this thread to show that there's user-interest. Thanks for sharing the idea.Quiddity (WMF) (talk)19:25, 5 December 2025 (UTC)[reply]

It's taking up like 80% of this page. I have no formal proposal, but it might be a good idea to have a separate talk page/notice board for this.-1ctinus📝🗨20:47, 2 December 2025 (UTC)[reply]

Wikipedia talk:Writing articles with large language models additional participation is welcomed, especially as a lot of the discussion here is redundant to discussion already happening over there. --LWG^talk20:57, 2 December 2025 (UTC)[reply]

Maybe a banner would be helpful to redirect prospective posters?-1ctinus📝🗨22:27, 2 December 2025 (UTC)[reply]

That's a good idea -- it's a little concerning how many people weren't aware this RfC even happened (not saying it's their fault, the topic just seems strangely under-publicized somehow despite taking up volumes of space)Gnomingstuff (talk)03:24, 3 December 2025 (UTC)[reply]

I agree with you. It's getting annoying seeing all these LLM/AI discussions clog up the Village Pump. Maybe another Village Pump tab (e.g.Wikipedia:Village pump (AI/LLM)) is needed, I dunno.Some1 (talk)18:22, 7 December 2025 (UTC)[reply]

I feel the same, but I wonder whether general community oversight of those discussions would be better than letting them 'hide' on a page only frequented by people with a particular interest/POV.WhatamIdoing (talk)22:21, 9 December 2025 (UTC)[reply]

Agree with the above. An idea: putting all LLM- or all AI-related threads into one thread here.

Then there's fewer page headers about these and maybe one could collapse it to scroll over the page. Maybe a feature to collapse threads on desktop is missing.
Another feature that would be great is getting notifications for new threads as another approach to watchlisting this page. See this idea in the wishlist:W370: Mute some discussions on a busy page. Narrowly scoped proposal pages are problematic for several reasons.Prototyperspective (talk)12:37, 14 December 2025 (UTC)[reply]

When I look at a country's infobox and see that the type of government listed in the infobox is vague or missing critical details like with"Republic" or"Fascist dictatorship", I really dislike that as you're gatekeeping the information from our readers on what type of government it is, seems empty and missing to look at, and is unwikipedian of us.

And so, my proposal is that when you're writing a country (whenever it's either in the present or past tense of history), that you fill inthree basic criteria:

1. Is it Unitary? Federal? Confederal?
2. Is it Presidential? Parliamentary? or something else entirely?
3. Is it either a Republic or a monarchy?

The sentence structure should be something likeFederal presidential republic,Unitary absolute monarchy, orConfederal directorial principality.

It can also cause something that I would call a useless conversation, something like this:

"Hello! what type of government isthe Gambia?"

"Presidential Republic."

"But is it Unitary or Federal?"

"Presidential Republic."

"Could you give me more details on the type of government?"

"Presidential Republic."

"Is this working?"

"Presidential Republic."

It doesn't make sense to leave out correct and good information and overall just looks bland.

If you have any questions, just ask and I'll try to answer.

GuesanLoyalist (talk)11:12, 3 December 2025 (UTC)[reply]

We can’t make it mandatory to do anything (seeWP:PRINCIPLE), but we could make this part of the guidance atWP:COUNTRYLEAD orTemplate:Infobox country if others think it’s a good idea (notifiedWP:COUNTRIES)Kowal2701 (talk)12:01, 3 December 2025 (UTC)[reply]

The guideline seems to work fine as I can see that as a good compromise and what I can see as improving.

GuesanLoyalist (talk)20:35, 3 December 2025 (UTC)[reply]

A very bad idea, based on a complete misunderstanding of how articles are created and how they evolve over time. We don't police articles to ensure they comply with arbitrary criteria invented to correct 'blandness' or some strange urge to emulate WikiData. More so when things like 'government type' are frequently contested and per policy shouldn't be reduced to bald assertions in infoboxes anyway.AndyTheGrump (talk)12:13, 3 December 2025 (UTC)[reply]

Infoboxes should only include the most important information. They are generally too bloated, not too short. (For example, the infobox atUnited States still helpfully converts GDP from US dollars to US dollars and lists both values.) If it is not important to scholars of the Gambia whether it has a unitary or federal system of government, then that shouldn't be in the infobox. I also agree with Kowal that we can't make anything "mandatory".Toadspike[Talk]12:53, 3 December 2025 (UTC)[reply]

GuesanLoyalist: please keep in mind that{{Infobox country}} is used in over 7,000 articles, including articles that are not about countries. The|government_type= parameter is currently used in only about 3,800 of those. Using some technical means within the infobox to require a value in|government_type= would not make sense for at least some of those articles (e.g.Benelux,Council of Europe,Central America,Central Powers). –Jonesey95 (talk)01:29, 4 December 2025 (UTC)[reply]

Sorry, I have not known about that

Maybe make it a basic guidance like what @Kowal2701 suggested for the actual countries?

GuesanLoyalist (talk)01:38, 4 December 2025 (UTC)[reply]

You could look at theList of countries, check each infobox, and fill them in yourself, if they're missing.WhatamIdoing (talk)03:40, 5 December 2025 (UTC)[reply]

One of the main issues with the most important pages is they require expert editors on the topic to improve them to GA status. These people are busy IRL, and are unlikely to take Wikipedia seriously. Peer-reviewed journals get these people toreview for free, and this can count as service for tenure packets. One issue with using Wikipedia for this is that accounts are generally anonymous, and anyone can claim to be anything or anyone here. Recently we introduced temp accounts, could a non-anonymous account that requires a .edu email to sign up for, combined with some collection of access to sources and letters of thanks that tracks service that could be put in a tenure packet, be possible/useful? Is there anything else that could be used as bait for expert editors?GeogSage ^{(⚔Chat?⚔)}18:49, 3 December 2025 (UTC)[reply]

Possessing a .edu email address (or equivalent) is not restricted to subject experts or even just academics. For example by virtue of being a life member of the computer society atSwansea University, which I got being serving as the society secretary for a year about 25 years ago, I have an @swan.ac.uk email address despite not even being a graduate. I have a friend with a dot .ac.uk email address because they work as an administrator at asixth-form college.

Secondly, not everybody who is a subject matter expert is an academic and/or works in academia. I have acquaintances who are experts in different aspects of railway history but they are retired railway professionals not academics. I spoke with one of them a few years ago about editing Wikipedia, but they were simply not interested - their primary interest was in conducting the original research. There is also the issue that much of what they would want to write about if they were interested in doing so would be regarded as too niche for a general purpose encyclopaedia.Thryduulf (talk)19:48, 3 December 2025 (UTC)[reply]

I'm aware that you don't need to be an academic for a .edu email, it is one possible limit though, especially if the email is made public and the account is not anonymous. Trying to recruit experts outside academia is another challenge, I'm trying to focus on one approach to getting one possible group of people who have a potential institutional motivation to do service. If you have suggestions on ways to recruit and motivate other groups of experts like those you mention, please suggest it.GeogSage ^{(⚔Chat?⚔)}19:59, 3 December 2025 (UTC)[reply]

Students get them en masse. At minimum, it would have to be restricted to non-students, and I don't think that's feasible, so, long story short, this is not workable._{Piotr Konieczny aka Prokonsul Piotrus| reply here}12:53, 10 December 2025 (UTC)[reply]

The WMF has programs like Wikimedian in Residence that encourage universities to support Wikipedia by encouraging them to create Wikipedia-oriented positions for academics. But that involves a lot of resources to get a single position at a university. I wonder if we could encourage more editors by asking the WMF to also try encouraging universities to promote Wikipedia as a option for fulfilling faculty service requirements.

On the front of experts outside of academia, expanding Wikipedia Library offerings and publicizing them more might attract some contributors._signed,Rosguill^talk01:41, 4 December 2025 (UTC)[reply]

If we could get universities to accept Wikipedia work as service, through whatever means, I suspect we would have a large volume of academics editing. I use Wikipedia as a means to help me actually read the stack of PDFs I download for work on other projects and broaden my understanding of my discipline, the instantaneous gratification of including a source or bit of information is a great motivator, but most professors I know consider it a waste of time they could spend on things they get credit for. Even if the University doesn't consider it as part of a tenure packet, "verified" profiles could help overcome this by allowing a professional to demonstrate some outside work in a qualitative way (even outside academia).GeogSage ^{(⚔Chat?⚔)}02:19, 4 December 2025 (UTC)[reply]

User:GeogSage, I already count my Wikipedia work as "service", but let's be clear: very few people in academia need more "service" to put on their annual report (for the outsiders, we typically get evaluated on teaching, research, service). What we need is for Wikipedia to count as "research", and that's not going to happen until Wikipedia's status in academia goes up. My dean tells me every year "yeah we can't count that as research" and he bases that, pretty much, on what he sees as a rough consensus, nationwide, in the profession: that writing up articles, whether GA or FA, even within one's own field, does not constitute what we call "research". Writing up stuff for online databases, that counts, but the various stigmas associated with Wikipedia continue to prevent us academics from getting credit for work done here. Look at my contributions: I've given up on getting them recognized professionally, and that is one factor in my no longer being so active in actual writing and improving articles.Drmies (talk)16:41, 10 December 2025 (UTC)[reply]

The problem with counting Wikipedia work as research isWikipedia:No original research. Fundamentally, research as I understand it requires the creation of original thought, and should be original. Wikipedia is an aggregator of that thought, and by its nature is not original. One of the pages I'm the most proud of isTechnical geography, and I could improve it tremendously if I could use my own thought on the topic, there are things I know about it through synthesis that are just not in the readily available literature. However this requires I first publish that synthesis in a reliable outlet, which would itself count as research on my annual report. Based on Wikipedia's own policy, I don't see getting it counted as research duties, which is why I started with service.GeogSage ^{(⚔Chat?⚔)}19:58, 10 December 2025 (UTC)[reply]

User:GeogSage, "research" in my business is not necessarily original thought: research comes in all kinds. The problem with it not being weighted as research is the status of Wikipedia, not the nature of the writing. I got two publications in the pipeline--one is of the kind that you're thinking of, with me doing thinking and interpreting, but the other, for the most part, is a biography of the kind that we write here. And I got a couple articles inthis series--there's a mix of "original research" there, along with regular biographical/historical writing. But if Eric Corbett and I had written upGreen children of Woolpit outside of Wikipedia, I am sure I could have found an academic journal that would take it.Drmies (talk)20:42, 10 December 2025 (UTC)[reply]

That is true, however in my experience different types of research are weighted differently. Peer-reviewed publications are the gold standard, other stuff is nice but given as much weight. This is a problem in itself, I have some publications that are not in journals, but they aren't valued as highly.GeogSage ^{(⚔Chat?⚔)}20:51, 10 December 2025 (UTC)[reply]

There's also a corollary issue, which is that irrespective of what one's current institution thinks of Wikipedia activity, most academics also need to think about building a resume of publications for future jobs._signed,Rosguill^talk20:55, 10 December 2025 (UTC)[reply]

The argument I made, and I made this in my promotion file as well, is that FAs and to a lesser extent GAs are in fact peer-reviewed, as are DYKs. It didn't fly, but it should have. On an average FA one gets more peer-review than for most journal submission. For my book, I got two reviewers. For a recent book chapter, two; for a biographical article, one. But for one article inStudies in Medieval and Renaissance Teaching, I hadseven reviewers. My point is that "peer review" (and you know this also of course) isn't always the same thing, and to fetishize it for journal articles and deny it happens on Wikipedia, or doesn't count for anything, is just wrong. But this is a problem in the profession--it's not a problem Wikipedia caused or can do much about. It's up to the T&P committees (our colleagues) and the deans (our supreme rulers).Drmies (talk)21:06, 10 December 2025 (UTC)[reply]

One of the main issues with the most important pages is they require expert editors on the topic to improve them to GA status. Except that isn't true. Anyone who's reasonably careful and willing to do some background reading if necessary should be able to raise most articles to GA status. (Our most technical math articles may be an exception, but more or less everything else is fair game.) I'm also a little confused which of our articles are now the "most important".Cremastra (talk ·contribs)16:52, 4 December 2025 (UTC)[reply]

I'm generally referring to articles rated highly by the vital articles project. While anyone can technically put the work in to get an article to GA status, an expert editor will already have that background. Finding sources is not always straight forward, and the knowledge of how to navigate the literature landscape is not something that happens over night. There are concepts that are not common knowledge that people won't even know should be included in an article without some background. Furthermore, in my narrow area of knowledge, I see that there are errors on Wikipedia that are major but that, no matter how many sources I provide, most editors don't even understand the issue. There are some things that are really hard to self teach, but really easy to think you've mastered without an outside opinion.GeogSage ^{(⚔Chat?⚔)}21:13, 5 December 2025 (UTC)[reply]

Wikimedia sponsored some research on the problem of low academic engagement[12].Alaexis_¿question?21:02, 5 December 2025 (UTC)[reply]

Some of our editors are (or at least claim to be and I have no reason to doubt them) academics, and seem to spend a lot of time editing Wikipedia.Piotrus andDrmies, can you say anything here?Phil Bridger (talk)23:07, 9 December 2025 (UTC)[reply]

I have written articles, in newspapers and in peer reviewed journals, arguing that "If we could get universities to accept Wikipedia work as service, through whatever means, I suspect we would have a large volume of academics editing" and that this is ethically a good idea. More influential folks than me have done the same, but clearly, we are a voice crying in the wilderness. I hae no idea what could be done better. I could say that WMF could use some of its funds that it is wasting on some stuff to do PR for this idea, but honestly, I doubt it would help much, the organizational intertia is just too big to deal with. Universities are not accepting Wikipedia as service, because it is not a component ofuniversity rankings, and this is the main thing that matters for bureaucracy (as rankings=student draw=$$$). It's as simple as that._{Piotr Konieczny aka Prokonsul Piotrus| reply here}12:56, 10 December 2025 (UTC)[reply]

User:Piotrus, as I said above, I count my work here as service and I don't think many in academia would have a problem with that, but service is typically only up to 15% or 20% of the evaluation. We need it counted as research.Drmies (talk)16:42, 10 December 2025 (UTC)[reply]

That too. And I read your point above about OR. It's valid, and other encyclopedias usually allow OR. That said, OR is often in the eye of beholder, particularly in cases ofWP:SYNTH, and when we create articles on topics that don't have proper treatment. Again, folks disagree. Recently I talked with a collegue of mine (academic who also occasionally dabbles here with small edits). I believe that articles such as a book writeup, summarizing reviews and academic analysses and creating the first poper overview of said book is valuableresearch, even if it is just compiling existing knowledge. He doesn't think so. Anyway, to keep it short, while OR is not allowed on Wikipedia, R (research) is, and what we often do is research, as defined and explained in tha article. So, sure, it should be counted. And we know it is not going to happen soon, due to organizational intertia, lack of understanding and incentives for change. I mean, academia has more serious problems it cannot deal with (peer reviews, closed access parasitism, degree inflation, etc.). Shrug._{Piotr Konieczny aka Prokonsul Piotrus| reply here}10:20, 11 December 2025 (UTC)[reply]

Is thewikiversity:WikiJournal User Group still active?WhatamIdoing (talk)03:18, 14 December 2025 (UTC)[reply]

@OhanaUnited may know_{Piotr Konieczny aka Prokonsul Piotrus| reply here}03:52, 14 December 2025 (UTC)[reply]

See the pages inCategory:Wikipedia expert help.

Some ideas and Pros and Cons can be found atthis spot in the structured argument map "Should scientists contribute to Wikipedia?".

.

I agree with what Piotrus somewhat but also see potential issues of the intrinsically-motivated genuine-volunteering and NPOV principles being undermined by such to some degree. I think a quite effective approach would be anonymous Wikipedia contributions certificates where academics could show that they contributed substantially constructively without having to reveal what they did (1. safeguards privacy 2. and neutrality and 3. addresses potential conflict of interest issues). This concept isn't far developed so more R&D on it would be great. Also relevant to recognition of open source development contributions as 'volunteering' (see petition). This maybe could also be used the other way around to verify one's academic experience without harming privacy albeit I don't think that would have much of an impact (could make it easier to find relevant users for a topic or by suggested tasks).

.

Secondarily, I think when it comes to effectiveness it's maybe less about "bait" and incentives and more about making the potential expert editors find places where they're needed and about them learning Wikipedia editing / getting them signed up and to explore a bit. The latter could e.g. be addressed by universities showing a demo of how Wikipedia works or somehow incentivizing such potential editors to sign up etc. The former could be partly addressed via what I proposed atW316: Suggested tasks based on contributions history (user interests) also for experienced editors. Tasks (& articles) would basically find their relevant experts who may spend only very short times on the site and aren't looking much / exploring around to find such.Prototyperspective (talk)12:31, 11 December 2025 (UTC)[reply]

Recently, @LuniZunie and myself createdWikipedia:WikiProject AI Tools, aimed at working on tools leveraging AI models to help Wikipedia editors with tasks other than content writing. However, the line appears to have quickly been blurred. Some of the proposed tools have been focused on tasks such as generating edit summaries, which we've historically been using as a warning sign to stop generative AI abuse. More worryingly, others (Flow Checker,AI Proofreader) will review an article's writing, which might risk editorializing or pushing a POV (even something as innocuous as afalse balance) without the AI writing words itself.

Beyond the question of the WikiProject's scope, there is a fundamental question of what the community is okay with in terms of AI-assisted tools, and it is crucial that we workshop a policy or guideline regarding what is or isn't accepted by the community.ChaoticEnby (talk ·contribs)20:32, 5 December 2025 (UTC)[reply]

As long as the final output is verified by a human then it doesn't matter what an AI did or didn't do before the human reviewed it. If the final output is not verified by a human then that's not acceptable regardless of what it is wasn't reviewed.Thryduulf (talk)21:40, 5 December 2025 (UTC)[reply]

That's easy for you to say, but we'd have to actually establish what "meaningful human review" is and how we can confirm it has happened.Cremastra (talk ·contribs)21:42, 5 December 2025 (UTC)[reply]

Random thought: what do you guys think about GenAI contributions being posted to talk pages in the form of edit requests, to be implemented by another human after review? --LWG^talk21:56, 5 December 2025 (UTC)[reply]

what problem does this solve/benefit does it add relative to human-written edit requests?NicheSports (talk)21:58, 5 December 2025 (UTC)[reply]

None at all, human-written is always preferred. But in the case that we end up landing as a community on "some generated text is acceptable to be inserted after review" the advantage of keeping that text in edit requests is that it prevents harm to the wiki without consuming experienced editor attention, since if the influx of requests exceeds the capacity to review, they can simply be ignored until more capacity is available, as opposed to the current case, where the text is inserted directly to the article and remains there in unreviewed state until someone devotes the effort to review and possibly remove it. --LWG^talk22:28, 5 December 2025 (UTC)[reply]

I think we'd end up swamped in requested edits, some of which were good and many of which were posted by new users who can't see which changes were good and just got an AI to scan the article and then tried to be helpful.

If we're going to allow any AI use (i.e. for identifying typos, etc., not in agenerative sense) it should be restricted to a set of trusted editors who are experienced and smart enough to know what changes to implement.Cremastra (talk ·contribs)22:02, 5 December 2025 (UTC)[reply]

Good suggestion indeed. Maybe make this a new user right? We've had issues with EC users misusing AI many times before, so it isn't just a matter of edit count.ChaoticEnby (talk ·contribs)22:10, 5 December 2025 (UTC)[reply]

Policies and guidelines already require a meaningful review. Use the exact same standard.Thryduulf (talk)22:02, 5 December 2025 (UTC)[reply]

@Thryduulf What policies and guidelines have established detailed procedures to review articles? And why in earth would those procedures be useful forreviewing the accuracy and usefulness of AI-generated content?Cremastra (talk ·contribs)22:13, 5 December 2025 (UTC)[reply]

I objected to the proposals that introduced these policies and guidelines because I believed they were vague and did not take into account details like this one. However the community consensus rejected this viewpoint, therefore sufficient procedures must exist to make it workable. I can't tell you what these are, you need to take it up with those who introduced the relevant policies.Thryduulf (talk)00:33, 6 December 2025 (UTC)[reply]

The AI Tools project was potentiallya little premature, given that the community is actively wrestling with what the limits on AI use should be. My recommendation would be that until our policies on LLM use stabilize, the AI Tools project should avoid advancing any use cases that 1) generate content (including edit summaries) or 2) review or adjust the meaning of article content. Catching typos does not adjust the meaning, so something along those lines would be fine.NicheSports (talk)21:45, 5 December 2025 (UTC)[reply]

That's what I would support too, and I hoped that the project would develop along these lines. Also interested by Cremastra's idea of additionally restricting this to a set of trusted editors, which might provide regulation from another angle.ChaoticEnby (talk ·contribs)22:09, 5 December 2025 (UTC)[reply]

I have suggested restricting LLM-assisted content generation to editors containing anllm-user right several times, so would certainly support this :) I'd prefer similar requirements to autopatrolled for that right, but could discuss.NicheSports (talk)22:12, 5 December 2025 (UTC)[reply]

The AI Proofreader is just that. The prompt is:

• Spelling and Typos: Look for misspelled words, especially proper nouns, technical terms, and common words.
• Grammar and Style: Identify grammatical errors, awkward phrasing, run-on sentences, and violations of Wikipedia's manual of style.
• Factual Inconsistencies: Point out contradictory information within the article.

So it won't editorialize or push a POV. It just helps identify internal inconsistencies, bad writing and mistakes.Polygnotus (talk)22:18, 5 December 2025 (UTC)[reply]

If it only identifies, and does not suggest new content, then I think that will be compliant with any future PAGs we develop for LLMs. Sounds fine to me?NicheSports (talk)22:22, 5 December 2025 (UTC)[reply]

Yup. And it can't do anything, it just tells the user "This is possibly a typo, there is a missing word in this sentence". Stuff like that. Feel free to give it a try.Polygnotus (talk)22:23, 5 December 2025 (UTC)[reply]

My worry is that, even though this is the intent, issues can easily creep in. For instance, fixing "grammar and style" might sound straightforward to us, but has often been used in AI-generated content as a justification for changes in tone or in due weight. Same for "factual inconsistencies", where it might make inferences from its knowledge base on matters that might not be clear-cut inconsistencies. The intent is noble, but I am worried about the doors it might open.ChaoticEnby (talk ·contribs)22:34, 5 December 2025 (UTC)[reply]

Yeah on second thought I think "factual inconsistencies" is too close to "article meaning" for me to be comfortable with. CE, would you be fine with use cases specific to identifying potential typos and MoS violations?NicheSports (talk)22:37, 5 December 2025 (UTC)[reply]

A factual inconsistency would be "According to the infobox this dude was born in 1765 but in the body of the article it says 1865". Allowing an AI to actually make such edits would be bad of course, see Grokipedia, but telling a human editor that is fine.Polygnotus (talk)22:38, 5 December 2025 (UTC)[reply]

Potential typos is fine with me. MoS violations could be okay, although, given the amount of conflict onone ortwo aspects of it, some care should be needed. Restricting it to specific violations (formatting-focused rather than prose/content-focused) could be ideal.ChaoticEnby (talk ·contribs)22:41, 5 December 2025 (UTC)[reply]

Seconding Chaotic Enby. In my experience, AI edits that "rewrite for neutral tone" actually introduce promotional tone, AI edits that "make things more concise" often do so by synthesizing material, and AI edits that summarize sources are currently awful as they simply tack on a source citation to their own homegrown AI slop. (In particular, AI tics like "highlighted" and "emphasized," in my experience, are almost always mischaracterizations of the source -- "emphasizing" means actuallyemphasizing, but AI will use it for even a passing throwaway mention, which can create a false impression ofWP:SIGCOV.)

You can see these issues crop up in the many ANI cases of people mass-rewriting leads with AI, creating problems (one example). I've also been gathering a dataset of AI snippets of article text with human snippets (Q2 2022 and earlier) of the same article to analyze later, and where possible I've used the same paragraphs. In my experience, even stuff that claims to be "just copyediting" tends to introduces tone issues, editorializing, or new information marked. (It also usually makes the text sound like AI, which is bad optics for readers.) Going through them now, will post some illustrative examples once I do.Gnomingstuff (talk)23:15, 5 December 2025 (UTC)[reply]

I just did an experiment on concision. Today I wroteNorthern emerald § Taxonomic history, an interesting couple of short paragraphs on the history of the taxonomy of a species of dragonfly. I asked Gemini to "make it more concise and neutral". Here's what it gave me: (I removed the stupid bolding it scattered all over the place).

"Here is a concise and neutral revision of the provided text, focusing on the key taxonomic and historical facts:
The northern emerald dragonfly was first described by Johan Wilhelm Zetterstedt in 1840 asAeschna arctica, based on specimens found in sub-alpine areas of Finland and Norway. In the same year, Edmond de Sélys Longchamps described a similar species from Belgium,Cordulia subalpina, a name he later listed as a junior synonym to Zetterstedt's species. In 1850, de Sélys reclassified the species asCordulia arctica, documenting its distribution across Belgium, Norway, Sweden, Finland, and Scotland. De Sélys further reclassified the species in 1871 into the subgenusEpitheca, and finally established its current classification,Somatochlora arctica, in 1887. He noted its range included cold and temperate Europe, Siberia, and Transcaucasia. A separate Siberian species,Somatochlora gratiosa, described by Aleksandr Bartenev in 1909, was synonymized withS. arctica by Syoziro Asahina in 1958. As of 1985, three forms ofS. arctica are recognized:fuscoptera,infuscata, andornata."

It removed all the citations and the writing is possibly the dullest thing I've ever read (which is saying something because my original writing was pretty damn dull unless you're excited by taxonomic revisions), but I don't see anyWP:SYNTH.Cremastra (talk ·contribs)23:31, 5 December 2025 (UTC)[reply]

I tried exactly the same task, and the sources were not removed. I always tell AI "You are an English Wikipedia editor", provide the article name for context, feed in wikitext, and request the same in return. --Викидим (talk)02:49, 14 December 2025 (UTC)[reply]

Викидим "Always"? You have used Ai for copyedits (substantially) before? Did you note this in the edit summary, and have these edits been safely reverted?? This is avery,very serious issue.Cremastra (talk ·contribs)03:06, 14 December 2025 (UTC)[reply]

Yes, I use AI. Yes, I always disclose its use, including prompts used. IMHO, a lot of issues reported on this page is due either to using bad (or simply old) tools or bad prompts. I fully expect the same users to produce bad texts without the use of AI. Again IMHO: the problem is real and serious, yet it is not in the AI itself, but in the sheer quantity of text that unskilled editors can generate in a minute. This problem will not be solved by aProhibition, it will just drive the same unskilled hands underground.Викидим (talk)03:21, 14 December 2025 (UTC)[reply]

See, e.g.,Georges Le Turcq andTalk:Georges Le Turcq#AI promts. Викидим has objected to "AI slop" but seems very comfortable with creating articles using AI. Presumably he does not deem all AI uses to be "slop".WhatamIdoing (talk)03:25, 14 December 2025 (UTC)[reply]

No, I do not think that this article is slop. Yes, I think that the current state of AI allows it to write better texts than produced by many human editors. Yes, I am comfortable with creating articles with the assistance of AI. I always check and review personally the texts I create.Викидим (talk)03:37, 14 December 2025 (UTC)[reply]

Whelp. One way or the other, this has drifted far beyond the initial scope of the conversation (which was explicitly about non-generative AI) and we're circling back into the old "AI-written articles, good or bad?" debate.ChaoticEnby (talk ·contribs)07:21, 14 December 2025 (UTC)[reply]

Cremastra has draftified that page since I posted the links last night. Based onDraft talk:Georges Le Turcq#Draftified, it appears that Cremastra wants a reader-facing disclaimer in the mainspace/article itself saying that AI was used to write the article.WhatamIdoing (talk)21:41, 14 December 2025 (UTC)[reply]

OK, here's a selection, all of which are copyedits from the Newcomer Tasks copyediting task, made throughout 2024 (circa GPT-4/GPT-4o). These are all from one user, but they don't read particularly differently from the (too many) other AI copyedits I've seen, and I have no reason to think this user has an unusual prompt. I'm also comparing them to the previous diff rather than the pre-2022 paragraph text; this does mean it's possible that it's AI copyediting AI, but I wanted to remove any intervening changes.

I've marked these edits up accordingly:

• Blue = introduced new information, removed information for unclear reasons, or changed meaning
• Green = introduced puffery
• Orange = introduced clunky, wordy, or otherwise bad phrasing

Hopefully this illustrates the issue.Gnomingstuff (talk)23:52, 5 December 2025 (UTC)[reply]

This is an excellent piece of work from you, thank you! Here is what I got for Flower Parade out of Gemini Pro 3.0 with just a default prompt as listed inUser:Викидим/AI prompts and instruction "rewrite the paragraph in Wikipedia style. If the articles for nontrivial terms terms do not exist in English wikipedia, but are present in other languages, use the ill templates:". The result looks actually good to me (I fed in your text as a plain one, without wikilinks, in real life an editor would check the links and remove the bad ones for Charoen Muang Road andBuak Hat Park), so I would say that the problem here most likely wasbetween the chair and keyboard and unrelated to AI:

Викидим (talk)03:21, 14 December 2025 (UTC)[reply]

Here's some more, from a different user's batch of rapidfire AI copyedits, all of which claim to rewrite for neutral tone but actually introduce puffery (and other issues). Same markup:

(I've omitted a lot of edits that didn't introduce promotional tone but didn't remove it either even when the problems are screamingly obvious, such asthis.)Gnomingstuff (talk)00:29, 6 December 2025 (UTC)[reply]

Thanks a lot. That's making a good case for these tools to be restricted, at most, to just fixing typos. Although even then, some people (and AI models) have a very generous definition of what counts as "fixing typos", beyond unambiguous spelling mistakes.

I've ran the latter three (with a typo deliberately added in each one) through Gemini with the prompt "Please fix any typos you may find in the following paragraph", here are the results:

In all three cases, Gemini managed to find the typo I added and correct it, without adding any extraneous material, which makes me confident that it can be trusted with this taskChaoticEnby (talk ·contribs)01:24, 6 December 2025 (UTC)[reply]

That's making a good case for these tools to be restricted, at most, to just fixing typos. No, that is making a good case for (re)writing articles with GenAI to be restricted. I think we already have consensus for that.Polygnotus (talk)07:00, 6 December 2025 (UTC)[reply]

The thing is that I suspect most people would view the above edits as closer to "fixing typos" than to "writing articles."Gnomingstuff (talk)18:23, 6 December 2025 (UTC)[reply]

That reminds me ofUser_talk:WorldPeace888#COI_/_PAID.Polygnotus (talk)07:12, 6 December 2025 (UTC)[reply]

I wouldn't worry much, these tools seem to require paid for API keys that almost nobody has (I was excited to try some of these tools and bounced back hard). Oh well._{Piotr Konieczny aka Prokonsul Piotrus| reply here}13:39, 10 December 2025 (UTC)[reply]

@Piotrus Gemini is free. And I don't mind giving you a Claude API key. And I already got pregenerated suggestions inUser:Polygnotus/barfoo &User:Polygnotus/barfoo2 (please remove em from the list when you've implemented the suggestions or decided thhat they are incorrect).Polygnotus (talk)13:52, 10 December 2025 (UTC)[reply]

@Piotrus, would be interested in your feedback if you can make it work.

Also, maybe it would be possible to give established Wikipedia editors API credits (in the spirit of the Wikipedia library) that they can use in apps like this one. From the user perspective it happens behind the scenes and they shouldn't be aware of it, unless they hit usage limits. The costs would be minimal but someone would have to bear them.Alaexis_¿question?16:32, 10 December 2025 (UTC)[reply]

@Alaexis Good idea. WMF has more than sufficient funds, but it's overly bureaucracized these days, so I don't even know what procedure if any would be relevant here. Anyway, after I get the tool to work, and I'll post my thoughts about it on its talk page._{Piotr Konieczny aka Prokonsul Piotrus| reply here}09:15, 11 December 2025 (UTC)[reply]

@Piotrus, I've found a free open-source model that we can use as long as it stays small-scale. I've told u:Polygnotus about it, hopefully it'll be implemented. In the meantime I've added it tomy standalone citation checker, feel free to test it. It's a standalone app rather than a script so it's functional but less convenient.Alaexis_¿question?17:00, 12 December 2025 (UTC)[reply]

@Alaexis Oh, citation checker? That sounds like an awesome tool I very much need for my grading of wiki student submisions :) Will try it out ASAP!_{Piotr Konieczny aka Prokonsul Piotrus| reply here}01:13, 13 December 2025 (UTC)[reply]

This has been touched on above, so I'm creating a new section to avoid the above discussion getting too derailed.

If we do create a user right for users trusted to use LLMs for non-generative purposes in articles (e.g. no changes that alter the meaning of the article and do not expect the LLM to check references) what should the minimum requirements for that right be?@Chaotic Enby andNicheSports:Cremastra (talk ·contribs)22:22, 5 December 2025 (UTC)[reply]

Willingness and ability to check every single (proposed) edit and take responsibility for it. We should demand the same for all edits. CIR is not a joke. AI slop and human brain slop ain't that different.Polygnotus (talk)22:24, 5 December 2025 (UTC)[reply]

Support this as a baseline, but I would also add as a soft requirement a demonstrated track record of transparency and responsibility (e.g. no issues of playing fast and loose with source verification). Stating willingness is good, but admins granting the right might want to also rely on evidence from the user's contributions.ChaoticEnby (talk ·contribs)22:38, 5 December 2025 (UTC)[reply]

I'm not sure I understand your response in the context of the question – should everyone have to apply for a user right before being allowed to make any edits?

Now I have said before that in the areas I edit, there is so much poor editing from humans that edits with program-generated content would just a drop in the bucket. The existing content dispute resolution processes are very inefficient, costing a lot of effort to deal with editors making poor edits. So I agree we need better processes to handle all those who submit poor writing, no matter how it was created.isaacl (talk)23:01, 5 December 2025 (UTC)[reply]

Hmmm. I am certainly in the more restrictive camp when it comes to desired LLM policies, but do you guys really think the community will support this direction? My idea for anllm-user right has been to restrict LLM-assisted contentgeneration to highly experienced and trusted users, with the same requirements as autopatrolled, although having to apply separately. Frankly I think the ship has sailed when it comes to using LLM tools for unambiguously non-generation tasks like finding typos.NicheSports (talk)22:47, 5 December 2025 (UTC)[reply]

As someone who uses AI a lot (too much) I believe AI content generation should simply be banned.

Using AI tosupport a human editor is fine tho, as long as the human makes the decision and takes the responsibility. If Claude gives me bad advice I'll just ignore it.Polygnotus (talk)22:49, 5 December 2025 (UTC)[reply]

@Polygnotus I understand this and it makes sense, but I think in practice it would just create an AI free-for-all anyway. Unless someone like you is willing and able to police every single edit, no one is going to "take the responsibility" as you described.

I would support your first sentence if a couple of words were removed:

"... AIcontent generation should simply be banned."

AI isn't worth the costs it imposes.TooManyFingers (talk)02:26, 14 December 2025 (UTC)[reply]

There are 3 distinct ways of working:

• AI makes the edit -- bad
• AI suggests the edit and the human can accept or skip with a single button press -- bad
• AI suggests improvements and the human can make the edit if they agree -- good

Polygnotus (talk)22:54, 5 December 2025 (UTC)[reply]

Oh, LLM content generation should be banned, outright. I'm in the "insanely restrictive" camp on LLM use.Cremastra (talk ·contribs)22:58, 5 December 2025 (UTC)[reply]

@Cremastra Well, it sure looks like we (mostly?) agreeUser_talk:Cremastra#LLMs.Polygnotus (talk)22:59, 5 December 2025 (UTC)[reply]

Yeah I mean my preference would be to completely ban LLM-assisted content generation because evidence (at AFC, NPP, AINB,1346 (hist ·log), etc.) has shown that the vast majority of users will not or cannot sufficiently review LLM output to make itWP:V compliant. My suggestion above is a compromise that I think would solve 99% of the problem so I am fine with it as well. Either works for me.NicheSports (talk)23:04, 5 December 2025 (UTC)[reply]

As I mentione before, this is closest to my preferred policy too. As far as specifics, I think there should be an application process. To apply, someone should at minimum:

• Describe,in detail, their entire process, including the tools they use and versions thereof, the exact prompts, the exact review process, etc. Obviously they should write it themselves, not use AI.
• Walk through an example of that. Provide the raw LLM output or iterations thereof, go through it line by line, check every statement against every source, change problematic material accordingly, and explain why they made every change.
• Then a reviewer -- preferably one familiar with AI writing who knows what issues are likely to crop up -- would need to also double-check the verification behind them, as well as review the prose. If a reviewer flags any issues, the person applying should take that into account.
• Upon getting the right, the user shoulddisclose AI use in all edit summaries and ideally on the talk page -- in part to indicate to anyone coming along later that the AI use was in fact reviewed (which they wouldn't otherwise know). In my ideal world there would also be a note on the article page that AI was used to write the article, because I think readers deserve to know this and because there's precedent in stuff like articles based onCatholic Encyclopedia text. I don't expect anyone to agree with me on that.
• The right can be revoked if there is a pattern of bad AI-assisted edits.

I don't think this process is too extreme -- it's what copy editors and fact-checkers do every day as their job -- but I don't think it is likely to gain much traction, because it's a lot of work for both parties.Gnomingstuff (talk)00:44, 6 December 2025 (UTC)[reply]

I would support this, but it could be made much simpler as we're talking about using AI-powered on-wiki tools for non-generative purposes, rather than asking outside LLMs for raw content generation. In that case, a lot of steps (like disclosing AI in edit summaries, or having to describe their process in detail) would be simplified.ChaoticEnby (talk ·contribs)01:08, 6 December 2025 (UTC)[reply]

Oh I should clarify -- this is regarding NicheSports'llm-user right for content generation. Personally I would prefer people didn't use AI to write articles but I would be ok with this kind of compromise.

I do still think AI use should be required in edit summaries though, no matter what. I guess that's where I'm a hardliner -- my view is that all AI use should be disclosed in a prominentreader-facing location, not just in places like edit summaries where nobody but editors ever looks. News organizations, research papers, etc. are expected to have these disclaimers, and we should too.Gnomingstuff (talk)01:30, 6 December 2025 (UTC)[reply]

Thanks for the clarification! Since the comment by Cremastra above mentioned non-content generation uses (whichWP:AIT is about), I felt it could be useful to mention it, but we're indeed talking about two different things here (and I would also prefer much stricter regulations, or a total moratorium, on AI content).ChaoticEnby (talk ·contribs)01:55, 6 December 2025 (UTC)[reply]

I'd definitely support this stricter approach as well, but I think it will take some serious admin advocacy to be approvedNicheSports (talk)03:36, 6 December 2025 (UTC)[reply]

Thankfully, admins don't have more authority over policy decisions than other editors. If a few concrete ideas come out of this discussion, which is quite likely, you or me can start a formalrequest for comment for the community to decide on them.ChaoticEnby (talk ·contribs)03:51, 6 December 2025 (UTC)[reply]

just be sure to wait 20 years otherwise it's too soonGnomingstuff (talk)03:57, 6 December 2025 (UTC)[reply]

I believe the introduction of a userright requires the WMF, so that 20 years will fly by.Polygnotus (talk)07:03, 6 December 2025 (UTC)[reply]

On the other hand, what would a user right actually do? MediaWiki can't really restrict which external tools someone uses to populate an edit form or build an API query, and there aren't any AI extensions (yet) that could be restricted likemw:Extension:ContentTranslation is restricted.

We wouldn't really even need an empty MediaWiki group to represent the "right", as assignment could as effectively be done likeWikipedia:Requests for permissions/AutoWikiBrowser andWikipedia:AutoWikiBrowser/CheckPageJSON instead.Anomie⚔15:10, 6 December 2025 (UTC)[reply]

We can't, but we can require (per policy) for these user scripts to be limited to that user right. As a matter of precedent, we already do the equivalent when requiring Huggle and AntiVandal to limit one-click reverts to users in the rollbacker group.ChaoticEnby (talk ·contribs)16:26, 6 December 2025 (UTC)[reply]

No, not really. I don't know how Flow Checker works (you'd have to ask Nullnominal) but you can't really have an opensource JavaScript check if someone has a specific userright.

Or, you know, if you did it would be laughably easy to evade that restriction (like is the case with AWB and its JavaScript equivalent, JWB). I have AWB rights but if I didn't I would still be able to use it if I wanted to; there is no protection mechanism to protect against that.

The problems we have seen with AI on Wikipedia are , like Chipmunkdavis (CMD) points out below, all about generating text and then sticking that text in Wikipedia articles. Give the AI proofreader a try; you'll see that it is not the problem. If you don't have any API keys ping me and I will send you one.Polygnotus (talk)16:48, 6 December 2025 (UTC)[reply]

Based on my experience with AI edits that claim to "improve flow" I do not feel very confident about the whole concept of that tool. The quality of the tool is only as good as the actual editors and how they gauge what good "flow improvements" are. Will try to track some of those downGnomingstuff (talk)19:32, 6 December 2025 (UTC)[reply]

This is like the rules for the first automobiles. Better to just ban GenAI (re)writing of articles and be done with it.

You can't factcheck your way out of AI slop. Maybe someone should write an AI version ofWP:BACKWARDS for that.Polygnotus (talk)07:17, 6 December 2025 (UTC)[reply]

I agree with @Gnomingstuff and @Chaotic Enby that the requirements for AI content generation and AI tool usage should be different. The AI-assisted content generation is a hard problem and if we let everyone do it we'd be buried in slop - I think that a temporary moratorium or a small-scale pilot project would be the best way to proceed.

On the other hand, AI-based tools can be quite helpful (full disclosure: I createdone for citation checking). Editors' time is precious and we should strive to use it more effectively. To take citation verification as an example, we know that some percentage of our citations doesn't support the claims they purport to support (there are 18k failed verification tags). However, assuming that the ratio of bad/good citations is 1/100 we can't reasonably expect editors to sift through 100 citations to find one incorrect one. Now these errors are fixed only if a subject matter expert happens to notice it and cares enough to fix/tag it.

If we use AI tools for this use case or for something else, the AI assistance should be limited to raising the flag and the human editor still has to make an edit and bears the responsibility for it. We should also require the AI tools to be open-source and to make AI prompts easily available to avoid bias.Alaexis_¿question?09:30, 6 December 2025 (UTC)[reply]

The only issue I've seen so far regarding using llms for source checking is over whether the source really doesn't support the fact or not, which can happen as pat of normal editorial processes. Such uses are only getting caught up in these discussions because when yet another content generation issue comes up, AI tool use is raised as a 'well what about this use would you ban this?' or caution about such a question. We should focus discussion only on the generation of text, because that is where 99% of the current issues lie.CMD (talk)12:35, 6 December 2025 (UTC)[reply]

They should show the ability to write decent articles on their own, without the usage of AI assistance, as one criteria. This is important because it signifies that they already have a good understanding of the best practices for writing an article, and are capable of checking an AI's output and modifying it to be of good quality. I think it's important we not allow people to automate tasks we couldn't trust them to do without automation in the first place; that's a recipe for massive destruction. It would be like giving AutoWikiBrowser permissions to a person who has no clue how to edit Wikipedia on their own.aaronneallucas (talk)22:53, 6 December 2025 (UTC)[reply]

I can barely string two words together, but I am a pretty good editor if I may say so myself.

capable of checking an AI's output and modifying it to be of good quality

That is writingWP:BACKWARDS. You can't factcheck your way out of AI slop.Polygnotus (talk)23:45, 6 December 2025 (UTC)[reply]

@Polygnotus That would be why I said a user should show that they can write on their own without AI tools. If you disagree with this criteria, that's okay, but please don't sarcastically misrepresent what I said; it isn't productive to this discussion. Besides that, my reply should not be construed as an endorsement of AI usage in writing Wikipedia articles, merely what should be required before a user is granted a theoretical user right to use AI tools.aaronneallucas (talk)02:08, 7 December 2025 (UTC)[reply]

@Aplucas0703 But I didn'tsarcastically misrepresent what you said. Perhaps I interpreted it differently than intended, in which case you can just explain that my interpretation is incorrect. I don't know you and since we are communicating via written text misunderstandings are basically inevitable.Polygnotus (talk)08:59, 7 December 2025 (UTC)[reply]

It's also not true that the only thing relevant to checking AI output is factchecking. What is actually required depends in large part on what the AI was asked to do and what changes it made/proposed. For example if you've just used an AI to improve your grammar and it made no changes to the sources used, then making sure that it didn't introduce any misleading statements when it changed the grammar (and correcting any found)is the difference between bad and good content.Thryduulf (talk)02:21, 7 December 2025 (UTC)[reply]

It's also not true that the only thing relevant to checking AI output is factchecking. Indeed, but no one made that claim as far as I know.Polygnotus (talk)09:02, 7 December 2025 (UTC)[reply]

That was the implication I took from yourYou can't factcheck your way out of AI slop. comment.Thryduulf (talk)13:48, 7 December 2025 (UTC)[reply]

Ah. No, that was not what I was trying to say. AI output contains a myriad of problems; way too many to list here.Polygnotus (talk)14:31, 7 December 2025 (UTC)[reply]

But that's so overgeneralised as to be untrue. AI outputcan contain a myriad of problems, but not every instance of AI output does. Some AI output contains sufficiently few (in some cases no) problems that it is possible for a human to entirely resolve them with an amount of effort that they deem worthwhile (how much effort that is varies person-to-person, and possibly task-to-task).Thryduulf (talk)16:02, 7 December 2025 (UTC)[reply]

@@Thryduulf Apologies for stating the obvious, butAI output contains a myriad of problems does not have the same meaning asevery single instance of AI output contains a myriad of problems.Polygnotus (talk)17:45, 7 December 2025 (UTC)[reply]

Literally no, but in the context of your comments on this page there is no meaningful distinction between the two.Thryduulf (talk)17:47, 7 December 2025 (UTC)[reply]

Yes, there is. Don'tstrawman.Cremastra (talk ·contribs)18:02, 7 December 2025 (UTC)[reply]

Could you identify a meaningful distinction, then? It seems to me like you and Polygnotus are making some very black-and-white claims and then backing up when someone points out the situation is more nuanced than you originally portrayed.Loki (talk)18:59, 7 December 2025 (UTC)[reply]

@LokiTheLiar Would you be so kind to explain why you think there is no meaningful distinction betweenAI output contains a myriad of problems andevery single instance of AI output contains a myriad of problems, either in the context of my comments of this page, or in general?

And can you post 2 or more links to Cremastra and myselfbacking up when someone points out the situation is more nuanced than weoriginally portrayed? Thanks,Polygnotus (talk)19:26, 7 December 2025 (UTC)[reply]

You started withYou can't factcheck your way out of AI slop, then backed up toAI output contains a myriad of problems; way too many to list here when Thryduulf pointed out that factchecking isn't necessarily relevant, and then further backed up to the claim that some AI output contains a myriad of problems when Thryduulf pointed out that some AI output contains few problems.Loki (talk)22:27, 7 December 2025 (UTC)[reply]

@LokiTheLiar Thanks. Probably best if I wait to respond until you've answered the other 2 questions right?Polygnotus (talk)22:32, 7 December 2025 (UTC)[reply]

Yes, there is a difference.

The statementAI output contains a myriad of problems means that AI-generated output as a whole produces many problems, such that, say, in a set of 10 pieces of AI outputs there will be 25 "problems" large and small. It makes no comment on how those problems are distributed among the texts. Some texts may be fine, but "AI output" as a whole contains problems.

The statementevery single instance of AI output contains a myriad of problems means that every single on of those texts above contains a large number of problems.

Obviously, these are different: the first is true; the second, an exaggeration no-one is arguing.

Identifying blatant logical fallacies does not constitute casting aspersions.Cremastra (talk ·contribs)21:00, 7 December 2025 (UTC)[reply]

The interesting thing is that one is overly broad ("every single instance") and the other is overly narrow ("the only thing").Polygnotus (talk)22:42, 7 December 2025 (UTC)[reply]

My comment was not a strawman and I would ask that you refrain from making such aspersions in the future. If you believe there is a meaningful difference between the position you are arguing for and my statement then you should have no trouble explaining what that difference is and why it is meaningful. Even after re-reading this discussion multiple times, I'm still unable to identify any.Thryduulf (talk)20:40, 7 December 2025 (UTC)[reply]

That is not anaspersion. The burden is on those who make a claim, especially when its an extraordinary claim like that 2 pieces of texts with different meanings are functionally the same.Polygnotus (talk)22:34, 7 December 2025 (UTC)[reply]

I would argue a candidate for such a pseudo-right should demonstrate

1. A history of understanding content policies and guidelines
2. Competence checking that sources verify
3. An understanding of the uses and limitations of large language models
4. A legitimate use-case

If at any time it were believed they no longer meet these criteria, the pseudo-right could be revoked.

My reasoning:

1. LLMs produce many issues, such as weasel and peacock wording, that a user would need to know how to identify. This could be demonstrated by a history of high-quality content contributions or detailed and skilled copy editing.
2. Not a difficult skill, but essential for anything involving AI-generated content. Not sure how this could be demonstrated beyond not introducing misinformation, but it would likely be the most common grounds for revocation.
3. Obviously necessary to avoid relying on the AI excessively. Could be demonstrated by time doing AI cleanup, or just an experienced admin asking the user questions.
4. Duh.

I think such a pseudo=right is a good idea, if only because it makes it easier to tell other users their AI contributions are prohibited.lp0 on fire ()19:42, 9 December 2025 (UTC)[reply]

@Lp0 on fire These are for a userright allowing the user to paste AI-generated text into Wikipedia, right?

I think I know only 4 or 5 people irl who understand the limitations of current AI models to a reasonable degree. People just think its some magic box that spits out the truth, or that its a magic box that spits out lies. Few people have a more nuanced opinion than that.Polygnotus (talk)19:48, 9 December 2025 (UTC)[reply]

And for that reason, few people should be allowed to use AI to write for Wikipedia. I'm not sure how broadly this should be scoped, but the question was what the requirements should be, so I was giving some suggestions. I suppose the "understanding of limitations" clause should only apply to understanding the limitations of the specific task they want to use AI for. Something more than "magic box give me answers" is definitely necessary but people don't need to have a PhD in AI. Or at least that's my take.lp0 on fire ()20:55, 9 December 2025 (UTC)[reply]

the "understanding of limitations" clause should only apply to understanding the limitations of the specific task they want to use AI for this would be the reasonable approach. If I wanted to use AI for copyediting articles about Indian settlements (I don't, but this is a class of article that, generally speaking, would benefit from copyediting) it is reasonable to expect me to understand the strengths and weaknesses of AI copyeditors (or at least the model I will be using) and possibly how they interact with Indian English. It is not necessary for me to have any particular understanding of the limitations of a different LLM regarding creating articles about contemporary scientists.

Obviously if my editing history shows that I spend more time creating articles about living scientists than I do editing existing articles about Indian settlements then it would be prudent for those evaluating the request for the right to ask about this if not addressed in the request itself. It should not be automatically disqualifying as there might be a legitimate reason for that (e.g. they might make tens of edits writing each new article but make only one edit per existing article when copyediting and state that an LLM wouldn't help them with their personal article writing process but would fit well with how they copyedit).Thryduulf (talk)23:03, 9 December 2025 (UTC)[reply]

That is a good point, although when giving a right that gives access to a broad variety of tools, we can't predict that the user won't start employing it in more problematic use cases later down the line. Someone could be very good at understanding that LLMs can, in fact, find typos, and then slide from "move typos" to "rewrite paragraphs of text" with little supervision.ChaoticEnby (talk ·contribs)23:13, 9 December 2025 (UTC)[reply]

At some point we have to assume good faith. I imagine it working something like how bots are currently authorised - a request is made to run a specific bot for a specific task (or set of tasks). Whether that task is desirable, whether a bot is suitable for carrying out that task, whether there is consensus for a bot to do that task, the suitability of the applicant to be a bot operator (in general and for the specific task) and the suitability of the specific bot for the specific task are all evaluated, and if approved, the bot flag is granted. There is no technical restriction on running a bot without the bot flag and/or for tasks other than those approved, however when we detect someone doing those things we stop them and, if appropriate, revoke the right. It wouldn't be identical (for obvious reasons) and we'd have to make it absolutely explicit that any comments on requests that there are no tasks suitable for LLMs (or similar) should be struck and/or ignored as contrary to community consensus (and repeatedly leaving such comments should explicitly be disruptive editing).Thryduulf (talk)00:48, 10 December 2025 (UTC)[reply]

Bot operators are presumed to make a request for approval for each new task, as they are usually well-defined, easy-to-track matters, and bots are relatively rare all things considered. Here, it would be much more likely that, once approval is given to someone to use LLMs for one purpose, they won't make a separate request for each task – similar to how folks can be granted the page mover right to draftify pages, but won't be expected to ask for it again if they want to help out atWP:RM/TR. It works for page mover as it is a position of pretty high trust (only a few hundred non-admin page movers!), but a "LLM user" right might be more widespread, meaning we would put trust in a lot more users to know their own limits.ChaoticEnby (talk ·contribs)00:56, 10 December 2025 (UTC)[reply]

Please stop assuming bad faith of LLM-users. An LLM right will be as high or low trust right as we choose to define it, it might be something akin to autopatrolled in which case treating it like a bot authorisation probably wouldn't work, it might be something on the level of edit filter manager (which is arguably a higher trust position than botop) in which case something like my thoughts above absolutely would work. Realistically it would almost certainly be somewhere between those levels (and that is where I would argue for placing it although I couldn't tell you exactly where right now). Similarly we would be free to define the scope of authorisations to be like botop, page-mover, adminship or anything else. As long as we are clear about what authorisation is being granted for (which can be as general or specific as we choose) and what the expectations are regarding doing/wanting to do things other than authorised, then the majority of those granted the right will meet those expectations. Those that don't will have the right revoked and, if appropriate, other action taken in exactly the same way that those who fail to comply with the expectations of other rights are treated.Thryduulf (talk)01:21, 10 December 2025 (UTC)[reply]

If it is as high as autopatrolled or higher, I would be comfortable with it. I was afraid that it would be something easily given out like TAIV or rollbacker, which would be a lot more problematic.ChaoticEnby (talk ·contribs)01:30, 10 December 2025 (UTC)[reply]

They didn't assume bad faith of LLM-users.Polygnotus (talk)09:38, 10 December 2025 (UTC)[reply]

As I said, it will be as easy or hard to get as we (community consensus, not you and me) choose. I don't have a good feel for what level others (other than those who oppose (almost) all LLM use) would desire.Thryduulf (talk)03:26, 10 December 2025 (UTC)[reply]

If it's easy to get, it might be easier to remove.WhatamIdoing (talk)04:49, 10 December 2025 (UTC)[reply]

As far as assuming good faith: there's one other entity involved here, and that's whatever AI company made the tool. AI companies are not known to be transparent about... well, anything, and they change shit all the time. So if an editor appears to slide from basic grammar-fixing copyediting to more substantive and problematic "copyediting," the thing doing the sliding or "acting in bad faith" might not be the editor but ChatGPT (or whatever).Gnomingstuff (talk)21:24, 10 December 2025 (UTC)[reply]

 You are invited to join the discussion atWikipedia talk:Arbitration Committee Elections December 2025 § Workshopping changes to the ArbCom election schedule.Toadspike[Talk]17:29, 6 December 2025 (UTC)[reply]

I've been thinking a lot about the current level of misinformation floating around on the Internet, and how it can be difficult for news websites and fact-checking services to keep up with the demands brought on by it. I had been thinking that we could make a kind of sister project called Wikifact (and just now checking to see that the name was available, turns out someone proposed this exact thing under the same name a while back, go figure, though it didn't receive much attention). It would function similar to websites like PolitiFact: just straight-up dedicated to fact-checking and nothing much else. However, of course, this would rely on verifiable sources and not original research, and wouldnot have to be framed as a declaration of "true" or "false," but perhaps framed as "supported by reliable sources," "partially supported by reliable sources," "not supported by reliable sources," and "contradicted by reliable sources".

I'm curious to hear what other think of this as a sister project (or being incorporated elsewhere), what you would like to see from such a project, potential problems you think we could encounter if this project were made live (and solutions), and what safeguards you would want to see put in place for something like this? I think we're at a place right now where more accessible fact-checking might be a good thing.aaronneallucas (talk)02:37, 7 December 2025 (UTC)[reply]

This is the wrong venue for this discussion - seem:Proposals for new projects.Thryduulf (talk)02:42, 7 December 2025 (UTC)[reply]

...but we'll talk about it here anyway, even though nothing can come out of the discussion. Didn't people understand what Thryduulf wrote?Phil Bridger (talk)19:33, 10 December 2025 (UTC)[reply]

@Phil BridgerUser:DVRTed/move-talk-section.js Enjoy!Polygnotus (talk)19:37, 10 December 2025 (UTC)[reply]

On the contrary, by restarting this discussion we could remind @aaronneallucas to create the actual new project proposal. :PLoki (talk)21:43, 10 December 2025 (UTC)[reply]

There is alsohttps://wikispore.wmflabs.org/wiki/WikiFacts_Spore seehttps://meta.wikimedia.org/wiki/WikiFactsPolygnotus (talk)14:35, 7 December 2025 (UTC)[reply]

That doesn't seem to me like that's the same thing. That WikiFacts would be a list of facts; this WikiFacts would be for fact-checkingLoki (talk)16:59, 10 December 2025 (UTC)[reply]

See alsohttps://captainfact.io/ (this one was featured in a documentary partly about Wikipedia).Govdirectory may also be relevant. I think the biggest hurdle is that it would likely be rather useless and unknown because a niche website nobody uses has little impact. What would have some impact if e.g. bots commented underneath posts claiming verified false info, if Web browsers added a note at the top that the page one is reading contains several false claims, etc. I would start with thinking about how misinformation can be effectively addressed and then from there see where the potential for a wiki project is. Moreover, as you more or less implied, "supported by reliable sources" does not make something true and "not supported by reliable sources" does not make something false. Often, things are not clearly true or false and if a source supports something, it depends on on what basis it's supporting the statement (data suggesting so, proof suggesting so, people the journalist interviewed claiming so, the journalist's opinion, sth else?).Prototyperspective (talk)17:08, 10 December 2025 (UTC)[reply]

I think that is somewhat of a good point, too. Though I don't necessarily think we have to say that we are the arbitrators of truth, which is why I thought indicating the sources themselves might be more useful. I think we could add more context along with a statement, like "verified by Labor Department survey data" or "contradicted by study conducted by Smith et al. (2024)" or even adding an additional label like "mixed evidence in reliable sources". We could even add a tag to all pages stating: "If you have a reliable source you can add relevant to this fact check, please do so". I also think, of course, the full page for it would provide an in-depth explanation of all sources.

Perhaps, alternatively, we could add a short summary to display on a mainpage, along with the full summary on the fact check's page, ditching any type of universal labeling system all together.aaronneallucas (talk)18:53, 10 December 2025 (UTC)[reply]