There isconsensus for better guidelines along the lines of and/or in the spirit of this draft, but unfortunately this version isn't quite it; it's close but no cigar. There is a weighted majority in favour of such a guideline but there are also issues that require addressing prior to promotion.
As an overview there remains three distinct factions within LLM discussions that are represented in this RfC; those that want stricter regulation of LLM, or their complete prohibition; those who favour softer guidance such as that of the currentWP:NEWLLM; and those who quite frankly at this point favour some sort of compromise between the hardline anti-LLM viewpoint and the current two-line guide that's insufficient. The initial concerns were with the RfC being rushed, very soon after a draft, even if substantial discussion had occured previously. This snowballed into strong criticism of the wording of the guidelines that strays from the intended meaning, even if arguably boiling down to semantics. Others criticised the watered down nature of the guideline, either from an oppose or supporting standpoint, or otherwise considered it vague and ambiguous, even if supporting the guideline in principle. This did not make consensus easy to determine, likewise with the support!votes lacking rationales, some weak oppossing arguments, bludgeoning, and even the weak support for the guideline under the guise of being an improvement rather than ideal or polished.
The following discussion is closed.Please do not modify it. Subsequent comments should be made on the appropriate discussion page. No further edits should be made to this discussion.
The new draft guideline defines an LLM, strongly advises editors not to use LLMs to add content to Wikipedia, and describes how to handle LLM-generated content that is already present.
Support, per a lot of the arguments made in the previous RfC. I quibble with "unreviewed", but regardless it’s a big improvement on the current policy.'Perfection' is not required, even an incremental improvement is constructive, and future RfCs can always address quibbles.Kowal2701 (talk)12:43, 4 December 2025 (UTC)[reply]
Support Consistent with current practice. I'd like future RfCs on mandating LLM-use disclosures and clarifying acceptable limited uses like source-searching and copyediting.Catalk to me!13:29, 4 December 2025 (UTC)[reply]
That's one reason why a little bit longer RFCBEFORE might have been helpful, as I know I for one strongly supported the previous version but frankly don't have time to keep up with the pace of this conversation. --LWGtalk16:29, 4 December 2025 (UTC)[reply]
Don't beat yourself up about it; there's a decent support base here. I think a fair amount of us are tired of endless workshopping and RFCBEFORE and just want to get things over the line. There's never going to be something perfect which will satisfy everyone and cover all the bases, and we can always modify once it's in place (which is literally what we're doing currently with NEWLLM, so...)Athanelar (talk)14:45, 5 December 2025 (UTC)[reply]
"tired of endless workshopping" and "we can always modify once it's in place" are directly contradictory. Tired of workshoppibg, but we can always workshop more later? Have you considered just stopping? NEWLLM is only two weeks old, it doesn't need to be expanded so quickly. If this RFC doesn't succeed, consider just walking away from the whole issue for six months or a year. If editors who have worked on this are tired of working on it, consider stopping working on it, rather than "just want to get things over the line".Levivich (talk)15:53, 5 December 2025 (UTC)[reply]
No need to apologize. I'm more frustrated with a large part of the community that is allowing their anti-AI dogma to force us to rush decisions that ought to be made deliberately. The gudieline as it currently exists, for example, is so vague as to be useless, but it was nonetheless forced through because editors felt the need to "do something" without thinking about what it is we're trying to accomplish. We are shooting ourselves in the foot by encouraging PAGs that are hostile to LLMs and will result in people being rude and dismissive towards new editors who might innocently, but improperly, use an LLM.voorts (talk/contributions)15:06, 5 December 2025 (UTC)[reply]
Discussion was ongoing for 2 weeks with 100s of comments, it's challenging to make a single proposal for an RfC because people with strong and polar opinions engage more while the majority that are moderate don't. People being rude can be dealt with by our current policies. It's not a coincidence that the most hard-line opposition to LLM-use comes from NPP, AFC, and LLMN/AINB, it's not right that people defend LLM-use but don't engage in any clean-up.Kowal2701 (talk)15:27, 5 December 2025 (UTC)[reply]
To be fair, I haven't done a ton of NPP recently. I still think we can deal with incompetent AI use without banning good AI use. To paraphrase myself from one of the discussions at the new gudieline talk page, it's fantasy to think that banning LLM use will stop new editors and bad actors from misusing AI (which is already a CIR issue); all it will do is bar productive AI use and result in litigation about whether AI was used, rather than focusing on whether its use was disruptive.voorts (talk/contributions)17:09, 5 December 2025 (UTC)[reply]
The community isn't really rushing if anything it took Wiki a long time to get a guidline on AI(and personally I think while the current one can be improved it is not useless but beneficial and long overdue.)GothicGolem29(Talk)22:03, 5 December 2025 (UTC)[reply]
Oppose The "Do not use an LLM to add unreviewed content" section is self-contradictory about what is and is not allowed. The guideline as whole makes no mention of disclosure, per Ca needs to mention accepted uses, not relying on machine detection and (sadly) also needs to be explicit about assuming good faith and not harassing editors.Thryduulf (talk)13:54, 4 December 2025 (UTC)[reply]
@Thryduulf Since you replied to me below: what exactly do you find contradictory? That section disallows a bunch of stuff, but I don't see where it explicitly allows anything, so I'm not seeing any contradiction.Toadspike[Talk]02:05, 5 December 2025 (UTC)[reply]
The section starts off by saying "Don't Editors should not use an LLM to add content to Wikipedia, whether creating a new article or editing an existing one." which is contradicted by nearly everything else in the section which is more nuanced and implicitly allows stuff. It also contradicts other policies and guidelines that (implicitly or explicitly) allow AI in some circumstances.
If multiple people see a section as self-contradictory (which they do) and multiple other people read the exact same words and don't see that (also apparently true), then that's evidence that the text is not clear enough to be a useful guideline.Thryduulf (talk)02:30, 5 December 2025 (UTC)[reply]
I think the different wordings between the first sentence and the rest of the section are deliberate. The first sentence is stronger, yes, but having a single clear sentence would be useful when dealing with people who refuse to listen to or read much of anything, which is in my experience most people using LLMs poorly. I do not see any direct contradictions within this proposed guideline. Could you please name the "other policies and guidelines that (implicitly or explicitly) allow AI in some circumstances" that are contradicted?Toadspike[Talk]08:25, 6 December 2025 (UTC)[reply]
The first sentence states don't add content using LLMs, with no qualification, while the other sentences state that unreviewed LLM content should be added, which is a qualification.
The examples listed below by Anne drew of 'good LLM use' (whatever your position on that) make the contradiction clear: the first sentence makes the good LLM use impermissible, the other sentences imply it is permissible in some cases.Katzrockso (talk)08:34, 6 December 2025 (UTC)[reply]
Oppose - first, that RFCBEFORE was open for like 12 hours. That's not enough time. Substantively, I oppose "do not use LLMs..." as overly broad, totally unenforceable, and just a bad idea overall, like saying "Do not use a word processor...". We want to prohibit people from cutting-and-pasting LLM output wholesale onto Wikipedia (in articles or on talk pages); we do not want to prohibit people from using LLMs for any reason or in any way. If I want to use an LLM to start a draft and then I personally edit that draft to improve it, fact check it, etc., nobody on the internet will ever know I did that and, assuming my output is policy-compliant, there is absolutely no reason to prohibit me from doing it. You all have no idea if I started this comment with an LLM, or had an LLM write it. Or if I asked an LLM to check the grammar, or suggest improvements. Frankly, LLMs write better than some people write on their own; for some, an LLM-written comment is better than one written without the help of an LLM. We don't want to discourage people from using LLMs in any way, because they are useful. They can be used as dictionaries, thesauruses, search engines... wholesale copy-and-pasting articles/comments isn't the only way to use them. Additionally,A large language model (LLM) means any program that can generate natural-language text either in response to prompts or by transforming existing text. is not what alarge language model is; see the Wikipedia article for a definition. Please, please, please, not yet another Wikipedia guideline that redefines words to mean something different on Wikipedia than they do in the real world. If you want to expand the existing guideline (which should be expanded), say "do not copy-and-paste without checking..." (it doesn't need to be said three times, just once is fine) and then have the "handling LLM-generated content" part (the word "existing" is unhelpful, because it would apply to "new" LLM-generated content as well as existing content).Levivich (talk)14:50, 4 December 2025 (UTC)[reply]
V3 was posted 22:37 Dec 3 according toSpecial:Diff/1325583018. Was it posted somewhere else on Nov 24? Best practice, especially for site-wide new rules, is to announce the intent to end the RFCBEFORE and launch the RfC, and ask if anyone objects, and then wait at least a day, preferably more than one, for any objections. "Are we going to categorically prohibit all usage of LLMs on Wikipedia" is a question the community has already answered with "no" in prior RFCs, and should probably be asked on its own before we try to write guidelines that implement such a major change.Levivich (talk)15:03, 4 December 2025 (UTC)[reply]
I'm sure I speak for everyone when I say no apologies necessary, we all make mistakes :-) Another thing on the substantive draft: the reason that the currentWP:NEWLLM guideline is so short, almost comically so, is because that's all that we've got consensus for so far. Butthat guideline, those words, are what has consensus. And it says,Large language models (LLMs) can be useful tools..., so to rewrite the guideline into an absolute prohibition runs contrary to that consensus statement. LLMscan be useful, which is why their useshould be permitted within certain boundaries. The current guideline continues... but they are not good at creating entirely new Wikipedia articles. That has consensus, too; almost everyone will agree that having an LLM write an entirely new article, and copying-and-pasting it onto Wikipedia, is disruptive because of the high risk of error. An expansion of WP:NEWLLM needs tobuild upon that foundation, not try to re-write it into an absolute prohibition. A better approach to expanding the guideline would be a "LLM Do's and Don'ts"-style guide, that teaches people the proper and improper ways of LLM use--when are they useful and when are they disruptive. It should explainwhy the "don'ts" are "don'ts," and suggest better "do's".Levivich (talk)15:14, 4 December 2025 (UTC)[reply]
It's worth reading through the RFCBEFORE if you haven't already, as much of this has already been discussed there. Namely, a few of us pointed out concerns with carving out acceptable use cases in a guideline that is essentially fundamentally designed to prohibit, not to permit. The goal is to tell people what theyshouldn't do. If you include a section explicitly detailing the acceptable uses, then anyone who gets accused of using them unacceptably is just going to claim they were doing one of the acceptable things (I've already seen it happen; "oh, I didn't generate my comment with an LLM, I just used Grammarly on it and that's why it sounds like LLM prose")
This also really isn't an absolute prohibition; it still provides a carveout in the language of "unreviewed" content. It says you "should not" use an LLM to generate content for Wikipedia, not that you "must not." The only thing it absolutely prohibits is generating new articles from scratch with LLMs, which is (from what I see) generally interpreted as the meaning ofWP:NEWLLM anyway, so there's no change there.Athanelar (talk)16:58, 4 December 2025 (UTC)[reply]
I disagree with "a guideline that is essentially fundamentally designed to prohibit, not to permit. The goal is to tell people what theyshouldn't do." - the guideline should be designed toguide, meaning toinstruct, which means both prohibitions and permissions, what they should and shouldn't do.
I disagree with "anyone who gets accused of using them unacceptably is just going to claim they were doing one of the acceptable things". This is an old argument on Wikipedia that I've seen raised many times, and I think it's bad advice, contrary to the fundamental purpose of guidelines, which is to teach. The notion that we shouldn't outline what is acceptable because people who do unacceptable things will claim it's acceptable is nonsensical to me.
I disagree with "It says you 'should not' use an LLM to generate content for Wikipedia, not that you 'must not.'" That is the exact reason why "should" shouldn't be used in policies and guidelines. "Must" and "may" are clear; "should" is ambiguous. It's too bad that many of our policies and guidelines use "should" go mean both a requirement and a suggestion. We must not continue this practice.
Basically, I disagree with you on the fundamentals for what makes a good guideline, and what the purpose of the LLM guideline ought to be. Fundamentally, you want a clear rule you can use to take action against people who do something you think shouldn't be done. That's not how rules on Wikipedia work, or should work. Our rules are not laws, they are guides and descriptions of practice (descriptive, not prescriptive).Levivich (talk)17:33, 4 December 2025 (UTC)[reply]
Completely agree, and frankly I don't understand the justification for keeping important background information out of the guidelines.Guidelines aren't information pages. We have plenty of information already about why using LLMs is usually a bad idea atWP:LLM. is the reason given for why NEWLLM does not justify itself. I disagree. Guidelines areabsolutely information pages. If we don't explainwhy we don't want articles to be generated with LLMs, it gives the impression that their use is banned based only on Wikipedians' opinions.
If WP:LLM is the page that explains why using LLMs is a bad idea,that should be the guideline, not another entry in a series of incremental restrictions on LLM use. Making such incremental restrictions is more confusing to newbies then just explaining everything in one place.
I prefer to read NEWLLMnot asa ban on LLM-generated articles, but asa piece of generally accepted advice saying "LLMs are bad at writing Wikipedia articles, so it's generally not a good idea to use them." Although a significant number of editors read NEWLLM as "LLM-generated articles are prohibited and subject to deletion."
Confession: Other, experienced editors already pointed out the problems with vague guidelines in the RfC. I argued against them. I now understandwhy they opposed.SuperPianoMan9167 (talk)20:20, 4 December 2025 (UTC)[reply]
Support per my previous reasoning, which I'm just going to copy-paste because it remains unchanged: "Having no policy effectively is encouraging LLM use. The longer we remain in that state, the longer we are encouraging LLM use, and the longer it accumulates on Wikipedia [...] Going three years [without] AI policy has done substantial damage to the integrity of Wikipedia, and frankly we're probably near the precipice where that damage becomes irreversible." To this I will add: the longer we remain in this state, the longer the project is fundamentally misleading bordering on lying to our readers, given the huge advertising in the past few months about how Wikipedia is the human alternative to AI.Gnomingstuff (talk)15:05, 4 December 2025 (UTC)[reply]
@Gnomingstuff This is not proposing to adopt a policy but to replace one guideline with a different guideline. I'm also going to have to ask for a citation for pretty much all of your justifications, particularlyGoing three years [without] AI policy has done substantial damage to the integrity of Wikipedia.Thryduulf (talk)15:27, 4 December 2025 (UTC)[reply]
First, stop pinging me.
Second, I know what this is a proposal for. My view is that we need a policy but in the absence of that a guideline is the next best thing, and that this guideline is the improvement.
Most of the justifications should speak for themselves, but regarding the advertising push I am referring to stuff like[1] (which, among other things claimed that Wikipedia had AI guidelines when it didn't at the time)Gnomingstuff (talk)15:47, 4 December 2025 (UTC)[reply]
Not exactly. The article you link to says "In all cases, volunteers create and enforce guidelines for responsible use of AI tools...", that was on Nov. 10, and by thenWP:NEWLLM had already been created and was in the middle of the RFC to promote it to a guideline. I am also curious why you think that "substantial damage to the integrity of Wikipedia" has occurred in the last three years due to the lack of an AI guideline. Can you point to an example of a false statement in a Wikipedia article introduced by LLM usage that remained in mainspace for a significant amount of time, that wouldn't have remained if we had had an AI guideline?Levivich (talk)16:12, 4 December 2025 (UTC)[reply]
Not to speak on @Gnomingstuff's behalf, but from my point of view the lack of guideline has meant that disruptive use of LLMs by editors has meant protracted discussions at ANI and other noticeboard while editors and admins um and ahh over what to do about it.qcne(talk)16:28, 4 December 2025 (UTC)[reply]
As far as damage to the integrity of the encyclopedia: I refer you to your previous stancehere:
We don't know how many [need to be dealt with]. But even if a mere 10% are bad -- 9,000 articles -- PRODing 10 a day would take 3 years. And if it's higher than 10%, if it's 50% bad, then we're talking over a decade to PROD them all. To go through them case-by-case is totally impossible, in my view.
Substitute "PRODing them all" with "fact-checking them all" (fact-checking being much more time-consuming than PROD), and "9,000 articles" with "some unspecified but large amount of edits" (edits being much more numerous than articles).
As a data point, there are currently around4,376 drafts identified and/or declined as likely AI-generated in the past 6 months, not counting anything deleted for being older than that, and around3,900 articles identified as possibly containing AI-generated text, not counting anything already fixed or still undetected. Are we at 93,000 unreviewed AI-generated edits? I think we're easily on track for that in the next couple of years.Gnomingstuff (talk)17:11, 4 December 2025 (UTC)[reply]
unsurprisingly there has been no response to this; guess potentially problematic edits are only a problem when there's a real person you can scapegoat and bully over themGnomingstuff (talk)23:39, 7 December 2025 (UTC)[reply]
Support I think this is a very reasonable (and workable) proposal. I also think if future edits are desired, this structure is a good base. --Enos733 (talk)16:04, 4 December 2025 (UTC)[reply]
Oppose this gutted version, support version 2 from the RFCBEFORE As much as this would be an improvement onWP:NEWLLM, in it's current state it's just going to require an immediate RFC to strengthen it, and I'm tired of spending all my Wikipedia time on AI issues. Specifically we need these sentences back in the guideline:
Define appropriate scope:LLMs, if used at all, should assist with narrow, well-understood tasks such as copyediting. New editors should not use LLMs when editing Wikipedia.
Clearly state what CIR looks like for LLMs:If the editor cannot confidently check and correct the output, they should not use an LLM for that task. LLMs should not be used for tasks in which the editor has little or no independent experience.
Set an expectation of disclosure:Editors should disclose LLM assistance in the edit summary (e.g. "copyedited with the help of ChatGPT 5.1 Thinking"). This helps other editors understand and review the edit.
I also support Thryduulf's suggestion of adding something about AGF and biteing, since that is a common concern raised about stronger AI guidelines. --LWGtalk16:45, 4 December 2025 (UTC)[reply]
WP:LLM would need changes if it became a guideline, which brings back the original issue: there is no agreement on the changes. --not-cheesewhisk3rs ≽^•⩊•^≼ ∫ (pester)19:43, 4 December 2025 (UTC)[reply]
I would prefer to use@Qcne's proposal as a basis for a guideline thanWP:LLM, which I find bloated. If we can resolve the contradictory language and section header, and potentially incorporate some of LWG's suggestions, I think we can get something pretty strong approved!NicheSports (talk)19:58, 4 December 2025 (UTC)[reply]
How is WP:LLM bloated? Aren't guidelines meant to educate? The rules are not intended to be laws. I don't think "you can't use LLMs because Wikipedians say so" is a good approach to a guideline.SuperPianoMan9167 (talk)20:02, 4 December 2025 (UTC)[reply]
Oppose as not strong enough. This isn't even an improvement on the status quo, so I see no reason to support this.The weakest of weak supports this version is a little tougher than previous versions, and it is a faint improvement of the status quo of no guidance at all. Nonetheless! – AI-generated text isn't appropriate for a human encyclopedia, period. Re Ca and Thryduulf: thereare no accepted uses.Cremastra (talk·contribs)16:48, 4 December 2025 (UTC)[reply]
They can disagree all they want. Doesn't make the facts any different. There areno valid uses of LLMs on Wikipedia. If someone cannot summarize or paraphrase a source without resorting to AI slop, they lack competence to edit the article in question.Support stronger prohibitions.oknazevad (talk)07:02, 8 December 2025 (UTC)[reply]
There are many valid uses of LLM models on Wikipedia. An important one, which a guideline on LLM should perhaps encourage, is to ask the LLM to identify factual mistakes in WP articles. However, the editor should then check the outcome "manually" (unassisted) using the cited sources and other reliable sources to verify the proposed change.Gitz (talk) (contribs)09:42, 8 December 2025 (UTC)[reply]
So rather than manually verifying the facts in an article, you ask an LLM to do it for you, and then you... manually verify its whole output? I'm not seeing the improvement.Athanelar (talk)09:47, 8 December 2025 (UTC)[reply]
Yesterday I asked ChatGPT to verify an article. It found that the subject had not won a literary award, but had only been shortlisted. It was right and I corrected the mistake[2]. I wouldn't have been able to spot this mistake alone. It's a minutia that doesn't raise a red flag, so it gets unnoticed.Gitz (talk) (contribs)10:15, 8 December 2025 (UTC)[reply]
@Gitz6666except that the source actually said she won the award, which apparently you didn't check:The Unbreakable Miss Lovely (p. 60 on that eBook; orp. 38 on this one): "The first review arrived the morning of her birthday. It was the only one that would turn out not to be positive. (She would win an “Edgar” from the Mystery Writers of America in nonfiction for the book.)" That is the only relevant mention of "Edgar" in the text.
This is pristine evidence that LLMs can't be trusted to check sources, because they don't actually read them. If sources are contradicting each other, a better approach is needed that just doing what the LLM says is correct.Cremastra (talk·contribs)13:31, 8 December 2025 (UTC)[reply]
We already have a "better approach", the one I followed: to check the official website of the award, which unequivocally reports that she was nominated andLegacy of Death by Barbara Levy won the award.[3] This is only one of the many mistakes that can be detected using an LLM.Gitz (talk) (contribs)13:50, 8 December 2025 (UTC)[reply]
OK so it appears as though this award gives out "scrolls" to the shortlisted/nominated books[4], which some have interpreted as "winning" an Edgar Award ([5]). Misleading at best to say that that is "winning" an Edgar Award.Katzrockso (talk)14:08, 8 December 2025 (UTC)[reply]
Oppose with disappointment and some frustration. Agree with Thryduulf onDo not use an LLM to add unreviewed content" section is self-contradictory which was the exact feedback I gave at the RFCBEFORE, which@Qcne didn't engage with. I would enthusiastically supportthis interpretation but it is not clear from the language in the current proposal that this is what is intended. As written this contradictory language will cause significant confusion "in the field" especially for newer editors. Minor changes could have resolved this. The RFCBEFORE should have gone on longer to address this type of feedback.NicheSports (talk)16:49, 4 December 2025 (UTC)[reply]
I know it'sfaux pas to modify a guideline during RfC, but would your vote change if we changed the section header like so;
−
Do not use an LLM to addunreviewedcontent
+
Do not use an LLM to addcontenttoarticleswithoutthoroughhumanreview,ortocreatenewarticles
I think this brings the section header in line with the body of the guideline and the spirit of its intent, and should I think berelatively uncontroversial as a result.
No. In addition to not addressing most of my points this just demonstrates it's still very unstable and thus not ready to be considered for adoption.Thryduulf (talk)17:16, 4 December 2025 (UTC)[reply]
I think it'd be better as "Do not use an LLM to modify or add content to Wikipedia articles, or to create new articles." --not-cheesewhisk3rs ≽^•⩊•^≼ ∫ (pester)17:19, 4 December 2025 (UTC)[reply]
No, but I would support "Do not use an LLM to add content to articles without thorough human review." It should also say something about how we are responsible for what we publish on this website regardless of what tools we used to draft the text, so the review should be as "thorough" as is necessary to ensure compliance with policies, because if the text is not compliant, the editor will be held accountable for it. I wouldn't support the whole proposed guideline just with this one change, as I still don't support other parts of it, but it would bring me a step closer.Levivich (talk)17:25, 4 December 2025 (UTC)[reply]
This doesn't address the contradictory language at all. I can draft something later that would, but unfortunately the RFC is already open. Bah.NicheSports (talk)17:35, 4 December 2025 (UTC)[reply]
LWG's changes are basically just returning to V2 of Qcne's proposal, which was controversial during the RFCBEFORE.
Ultimately, there's no pleasing everyone here. The LLM hardliners like me won't be pleased by more permissive language, and the permissive ones won't be pleased by more restrictive language. I think this version really is the best middle ground.Athanelar (talk)18:08, 4 December 2025 (UTC)[reply]
I'm not opposing because this is too strict or too permissive, I'm opposing because it is contradictory. We need to stop rushing LLM PAG RfCs. Agree with Kowal2701, let's remove the RFC tag and work out this issue. But I can't do that because I have !voted. Qcne can, or an uninvolved editor.NicheSports (talk)18:24, 4 December 2025 (UTC)[reply]
Oppose - It contradicts itself in a few places as pointed out above. Also explicitly states it should not be used in general and that is just wrong. It can be used, but it needs to be reviewed and verified, which is a huge difference.~2025-38536-45 (talk)19:43, 4 December 2025 (UTC)[reply]
Support - this proposed guideline is still insufficiently strong, but it is a step in the right direction, and we should take it, even if we need to fix a wording here and there later, and even if we have to immediately hold a follow-up to try and strengthen it further.Tazerdadog (talk)21:52, 4 December 2025 (UTC)[reply]
Support – I have some qualms about the specifc wording, but this is a massive improvement. Thank you, Qcne. We should be listening to folks like him who have reviewed tens of thousands of AfC drafts, not nerds like me who spend too much time yapping on noticeboards like this one.Toadspike[Talk]22:02, 4 December 2025 (UTC)[reply]
Support - the right approach, and we will certainly continue to tweak. FWIW, I think AGF is implicit and nothing here undermines it so shouldn't need to be spelled out. —Rutebega (talk)22:22, 4 December 2025 (UTC)[reply]
Itshouldn't need to be spelled out, but experience has tells us that it absolutely does need to be. There are far too many editors who believe that a suspicion (correct or otherwise) of using AI is an automatic declaration of bad faith and thus they are free to respond in kind.Thryduulf (talk)23:11, 4 December 2025 (UTC)[reply]
Support I'd prefer stricter but, after a pretty thorough discussion this version satisfies most of the issues raised in the previous rfc, and no policy is ever going to be perfect. So let's not fall for theperfect solution fallacy and pass incrementally better versions instead of failing the imperfect. Also, Randy in Boise might actually realize he cannot simply paste what ChatGPT says about the Peloponnesian War.~ Argenti Aertheri(Chat?)23:45, 4 December 2025 (UTC)[reply]
Perfect is indeed the enemy of good, but I do not believe that something that is self-contradictory and has other issues (as noted above by multiple people) can be accurately described as "good" nor as "better" (incrementally or otherwise) than what preceded it (as that does not contradict itself).Thryduulf (talk)00:47, 5 December 2025 (UTC)[reply]
Oppose as insufficiently clear. This ambiguity will not permit good decision-making and will just cause issues - more time workshopping this proposed guideline would have been appreciated.Katzrockso (talk)02:45, 5 December 2025 (UTC)[reply]
Support as clearly better than the current guideline. We need to have a guideline that will clearly discourage the use of any unreviewed LLM created material in any article, not just in articles created from scratch as at present, and this draft does that. It may be that the guideline needs rephrasing in places (and in particular the first sentence of the "Do not use an LLM to add unreviewed content" section should be rephrased to make it clearer that only raw or lightly edited LLM content is being forbidden here), but we shouldn't let that stop us implementing this guideline as it is, since we can always improve it later.Dionysodorus (talk)08:00, 5 December 2025 (UTC)[reply]
Oppose. I share Thryduulf's concern that the do not use an LLM to add unreviewed content section is contradictory. The section itself only mentions not using LLM's to generate content it does not reference reviewed content or give examples of reviewed content that would be acceptable.GothicGolem29(Talk)14:20, 5 December 2025 (UTC)[reply]
Oppose. This is poor guidance that prevents legitimate uses of LLMs:
Editors should not use an LLM to add content to Wikipedia, whether creating a new article or editing an existing one. Do not use an LLM as the primary author of a new article or a major expansion of an existing article, even if you plan to edit the output later.
Sure. As one example, someone might notice information is missing from an article, use AI to draft a paragraph with that information, comb through it to correct errors or policy/guideline violations (if any), and include it in the article with appropriate sourcing.
But this is a general-purpose technology - there are innumerable valid uses of AI, from research, to copyediting, tofinding errors in articles. The important part is that any AI-generated content must be rigorously reviewed and corrected by a human editor. I would be glad to support a guideline emphasizing this.Anne drew (talk ·contribs)22:23, 5 December 2025 (UTC)[reply]
Legitimate uses:
Human-reviewed article improvement suggestions
Human-reviewed ideas for new articles
Human-reviewed analysis of diffs and usernames for RCP patrolling
Oppose per Levivich, who seems to sum up the issues well. The word processor analogy is a good one - in the right hands, with due care an attention, an LLM can be a useful tool in helping to produce better-quality prose and basic sanity checking of what you've written. And clearly in the wrong hands, LLM use is very dangerous - it gives editors the ability to churn out lots of material if questionable accuracy. The current guideline has an advantage that it is very short and to the point, it's very easy to see what it prohibits. I would support an expansion of that to (equally succinctly) prohibit newunreviewed LLM content of any flavour rather than just "new articles from scratch", but I certainly wouldn't support a change that restricts legitimate LLM use more generally, or makes it more wordy and therefore harder to point new users at the hard rules about what they can and can't do. With thanks to the OP for wading into this difficult debate and attempting to make progress, I unfortunately do find the proposal too wordy (as well as potentially contradictory, as per Thryduilf) to be an a improvement for us. Cheers — Amakuru (talk)17:41, 5 December 2025 (UTC)[reply]
The word processor analogy is a good one. I have to disagree with this point. It's not a good one, as an LLM is a tool for off-loading human cognition to generate text, whereas a word processor's purpose is to digitize written words into text. These two things are not analogous.Revolving Doormat (talk)01:58, 7 January 2026 (UTC)[reply]
Support. I am unclear of proper protocol for this support message, as this is my first time supporting. I agree with pester and Kowal2701.Parameci (talk)18:12, 5 December 2025 (UTC)[reply]
Support. Good step in the right direction. There's definitely a faction of editors that want to continue allowing some LLM use. Forbidding "raw or lightly edited LLM output", while not explicitly forbidding "all" LLM use, is what this RFC does, and is in my opinion a good compromise. –Novem Linguae(talk)23:31, 5 December 2025 (UTC)[reply]
P.S. To address the issues raised above of "Editors should not use an LLM to add content to Wikipedia" (absolute prohibition) being out of sync with "Editors should not Paste raw or lightly edited LLM output" (prohibition of most LLM use), I'd be OK with deleting the sentences that give absolute prohibitions. Just delete that first paragraph of "Do not use an LLM to add unreviewed content" and leave the bullets. Worth it to get this passed. –Novem Linguae(talk)23:37, 5 December 2025 (UTC)[reply]
Oppose. This part is poorly written and stops all legitimate use of LLM in Wikipedia:
Editors should not use an LLM to add content to Wikipedia, whether creating a new article or editing an existing one.
That statement shuts off all legitimate use of LLM on the Wikipedia. This also contradicts the other statements made on the section, especially as the text after this statement implicitly allowed LLM that are are heavily-edited.
Do not use LLMs to write comments or replies in discussions
There's no legitimate reason for this guideline. LLM can be used to fix grammar, which is a legitimate use of LLM. We can also use LLM to rewrite or refine our thoughts, which is another legitimate use of LLM. And if one agrees with what the LLM is spouting up, who are we to judge them? A bad judgment didn't mean that the opinion had to be discarded. We humans made bad decisions and judgments as well, and it is wrong to have it discarded outright.
Where content is largely or entirely based on unedited or lightly edited LLM output, it may be draftified, stubified, nominated for deletion, collapsed, or removed entirely, especially where the content is unverifiable, fabricated, or otherwise non-compliant with existing Wikipedia policies.
I have to disagree on this. A content has to be actioned only if it is unverifiable, fabricated, or troubled. If an LLM content contains no problem, there is no reason to remove the content from Wikipedia. Wikipedia has enough guidelines to keep our content "clean". We should not care too much "how" the content is written; we should focus on "what" the content is. We should judge the quality of the content, not the methods of writing the content.
Repeatedly making problematic LLM-assisted edits may be treated as a competence issue and can result in the editor being blocked.
Same reason as above. Problematic LLM-assisted edits are the same as problematic human edits. There should be no difference at all.I have to emphasize again that Wikipedia has enough guidelines to stop the excesses of LLM problems. There is no reason to add more rulings. Today we have seen the limits of LLM, but what will happen in 20 years ahead? Will Wikipedia expect the LLM to continue to be problematic? Today, our rulings are based on LLM throwing imaginary references, but what about in the next 5 years?✠SunDawn ✠Contact me!06:43, 6 December 2025 (UTC)[reply]
Support Even if this guideline is not perfect, it's significantly more detailed than what we have now. For a topic as complicated as LLMs, I feel that a short guideline on their usage is insufficient as it doesn't explain LLMs, whatspecifically editors should not do with LLMs, or what editors should do upon seeing LLM generated content, all of which this proposal addressed. I don't think it's necessarily perfect, but what matters is that it is a step in the right direction of having a concrete stance on LLM usage, and is far more tenable to changes than the current wording.Gramix13 (talk)15:57, 6 December 2025 (UTC)[reply]
But the guideline contradicts itself. One part says "do not use LLMs to add content", and the other says "do not use LLMs to add unreviewed content" (implying reviewed content is okay). These two parts represent the two major viewpoints toward LLMs on Wikipedia, which makes sense for a guideline that is supposed to be a compromise, but putting them together makes no sense, because we're simultaneously saying "don't do it" and "you can do it if you thoroughly review it".SuperPianoMan9167 (talk)17:17, 6 December 2025 (UTC)[reply]
I really think this issue is a mountain out of a molehill. The section in question says editorsshould not use AI to add any content (which is advice, not prohibition) and only explicitlyforbids using AI to create new articles from scratch. The section header fails to properly include the latter point, but it doesn't contradict it.
The section tells you;
Youmust not use AI to add unrestricted content
Youshouldn't use AI to add any content at all, preferably
Youmust not use AI to create new articles wholesale
The bullet points add some confusion by using 'should' rather than 'must', but we can easily improve the clarity on that post-RfC. I think the point is the spirit of the guideline, which aims toprohibit generating whole articles with AI (in line with NEWLLM) and otherwisestrongly discourage thw use of AI to add other content.
It'll be much easier to improve the clarity post-RFC than it would be to cancel the RFC, make a new draft, take it to RFCBEFORE, leave the RFCBEFORE up long enough to satisfy the people who are complaining about rushing, then bring it back to RFC and get consensus on it. So long as we can get consensus on thespirit of the guideline, the actual written clarity can be improvedAthanelar (talk)19:18, 6 December 2025 (UTC)[reply]
There is no deadline. I'd prefer we resolve the contradiction before proceeding.
Guidelines are not rules that you can enforce against anyone who breaks them; they areguidelines, which are meant toguide editors.
Also, you read the word "should" in different ways depending on the specific guideline: in regards to NEWLLM, you readLarge language models should not be used to generate new Wikipedia articles from scratch. as anabsolute prohibition on generating articles with LLMs, yet here, you say thatEditors should not use an LLM to add content to Wikipedia isnot an absolute prohibition on all LLM content generation. "Should" is therefore ambiguous. The proposed guideline does not have the word "must" in it at all, so you're reading words into the guideline that aren't there.SuperPianoMan9167 (talk)21:24, 6 December 2025 (UTC)[reply]
I don't think the use of 'should' is ambiguous per se, I think it's contextual. In the context of NEWLLM (especially when taken with Cremastra's stance in these subsequent RFCs, as well as the sentiments of those who voted in support of that RfC) I think it's clear that the intent was and is to act as a prohibition. If it weren't for this RfC going on now, then I would be lobbying to get 'should' changed to 'must' there precisely to clarify that, and it was something discussed at the time.
I think 'should' is read as non-prohibitive here precisely because it stands next to and is contrasted with statements that read "Do not..." (which is, I argue, plainly synonymous with 'Editors must not X' which is why I used that language)
NEWLLM is one simple statement which uses the word 'should,' and that's why I read it as prohibitive despite that.
This is multiple statements, some which say 'should not' and some which say 'do not,' and I think that difference in language marks a difference in implementation.Athanelar (talk)21:34, 6 December 2025 (UTC)[reply]
Similar to what Levivich argued inthis comment, NEWLLM has more nuance to it than you are suggesting, and it isnotone simple statement. It is actually three statements, each one with consensus support:
Large language models can be useful tools. (True.)
But they are not good at creating entirely new Wikipedia articles. (Also true.)
Large language models should not be used to generate new Wikipedia articles from scratch. (The conclusion.)
The second sentence indicates that the reason we are banning unreviewed LLM-generated articles is because they're bad. The words "from scratch" are vague, but my understanding of them is that blindly copy-pasting LLM output into the edit window and pressing "save" is not allowed. This is a perfectly reasonable rule becausecompetence is required. My understanding is that if an editor reviewed and fixed up an LLM-generated article so that it meets all policies and guidelines, the guideline would not apply. Even if the guideline is an absolute prohibition, editors are completely free tonot follow it if they are experienced editors who properly review their LLM-generated articles.SuperPianoMan9167 (talk)22:51, 6 December 2025 (UTC)[reply]
@Athanelar, abouttake it to RFCBEFORE, leave the RFCBEFORE up long enough: Please go readWP:RFCBEFORE. RFCBEFORE is about how to not have an RFC at all. (You may be looking for the gentle suggestion in the last lines ofWP:RFCBRIEF, which is about how to get help with wording an RFC.)WhatamIdoing (talk)18:29, 15 December 2025 (UTC)[reply]
Oppose Agree with Levivich that the days spent critiquing other proposals do not justify having less than a day ofWP:RFCBEFORE discussion on this proposal's final wording, especially when more review should have addressed the inconsistency highlighted by Thryduulf. Substantively, I agree with WhatamIdoing that whereasWP:NEWLLM was guided by AFC/NPP experience to focus on new AI-written articles, this proposal is prohibiting a far wider set of uses without demonstrating that those uses are similarly draining on editor time or corrosive to our content.ViridianPenguin🐧 (💬)18:54, 6 December 2025 (UTC)[reply]
Neutral, but positive: I know this process has been a long slog, so it may be time to say this is a 'good enough' draft to replace what exists, with the understanding that a new draft is needed. Some issues I want addressed: I think "unreviewed content" is trending in the right direction but we can be more specific about the problem: Editors should not add content they have not read and understood, and large revisions incorporating a lot of changes are difficult to review compared to small revisions. Generally, I think that Wikipedia's consensus process has been pretty robust in the face of vandalism and poor quality edits - conceptually, that's basically what bad AI editing is. I also agree with Thryduulf's observation about inconsistency. --Edwin Herdman (talk)21:06, 6 December 2025 (UTC)[reply]
Support in spirit. I agree with editors that the finalWP:RFCBEFORE should have been left open for longer. I agree with the prohibition onunreviewed content, but as others have pointed out, the current wording is self-contradictory. Again, I see no reason why we should be prohibiting the addition of policy-compliant content, although I wouldn't necessarily be against restricting it to, say, extended-confirmed editors in order to ensure that they are actually aware of the relevant policies and guidelines that the generated content is supposed to comply with.I think the closer should consider a partial consensus for this RfC, as I think most editors here support forbidding unreviewed LLM-generated content, but they don't support the much broader prohibition on LLM output.Kovcszaln6 (talk)11:29, 7 December 2025 (UTC)[reply]
I very strongly oppose adoptinganything (regardless of subject matter) as a guideline that differs from the words that people were explicitly commenting in the discussion, because even with absolutely impeccable faith there is avery strong risk of unintentionally introducing significant problems or misrepresentation of views.
In this specific case,Wikipedia talk:Large language model policy#RFC found no consensus to adopt "Large language model output, if used on Wikipedia, must be manually checked for accuracy (including references it generates), and its use (including which model) must be disclosed by the editor; text added in violation of this policy may be summarily removed." as a "policy/guideline" and that should not be overturned in a discussion that does not specifically discuss that. FWIW I'd be inclined to support a policy or guideline specifically and only prohibiting the addition of AI content to Wikipedia without human review (although specific wording, e.g. around what counts as a sufficient human review, would be of the utmost importance), but that is not what is proposed inthis discussion.Thryduulf (talk)14:04, 7 December 2025 (UTC)[reply]
After just now reading through yet another AI-generated unblock request which says that policies say things they absolutely don't say in reality, I am affirmed in my position that I think it's better if our net accidentally catches too much AI usage rather than too little. I of course recognise that this is an ideological position that you absolutely won't agree with; but every incidence like this only strengthens my belief that we need stronger, unequivocal restrictions on problematic AI usage, yesterday.Athanelar (talk)18:21, 7 December 2025 (UTC)[reply]
It isfar more important that we get any restrictions right than we get the quickly. Poorly worded restrictions that are confusing, misleading, contradictory, redundant (e.g. a significant proportion of the problematic edits highlighted in discussions like these already fall foul of existing policies and guidelines), too broad or too narrow help nobody but have a high likelihood of making existing problems worse or creating problems where none currently exist.Thryduulf (talk)20:53, 7 December 2025 (UTC)[reply]
Oppose. I find it hard to interpret as written. LLM output can be unreviewed, reviewed but not significantly edited, or reviewed and edited. The guidelines mixes all three, such that it's not very clear what exactly is forbidden. I can also see why previous comments find it somewhat contradictory. This proposal is directionally correct, but I'd rather support an edited and respun version of this RFC. That seems like a safer bet if the goal is to make it easier to make decisions about LLM (ab)use, and less risk of being stuck with a ambiguous guidelines.Mlkj (talk)22:48, 7 December 2025 (UTC)[reply]
MyOpposeorSupport depends on whether the all-encompassing sentenceEditors should not use an LLM to add content to Wikipedia, whether creating a new article or editing an existing one stays or not; in the latter case, I support the proposal. However, I think it would be useful to have aHelp page with LLM do's and don'ts, instructing editors on a helpful use of LLM, as Levivich suggested.Gitz (talk) (contribs)10:03, 8 December 2025 (UTC)[reply]
Based on the ongoing discussion, I'm striking through my conditional "Support": the sentenceEditors should not use an LLM to add content to Wikipedia is going to remain, apparently, as well as the "nutshell"Do not use large language models (LLMs) to write Wikipedia articles. I don't agree. I think thatthis article I created is a good article, and there's no reason to renounce this kind of contribution (when made with disclosure of AI-assisted editing). I think that the emphasis/the essential criterion should unambiguously fall on "raw or lightly edited" and "unreviewed content": if an editor publishes bad-quality content because their review/editing of AI-generated output is not sufficiently thorough, this is misbehaviour or a WP:CIR issue and needs to be prevented with blocks and bans. But if they want to use AI to produce decent content and are transparent about this, all the better.Gitz (talk) (contribs)11:49, 5 January 2026 (UTC)[reply]
When I try such experiments, I often like to post the output for transparency and to share the results. Flat prohibitions on the posting of AI generated text is not helpful in such cases.
This is something that I hadn't thought of before, but posting both unreviewed and reviewed outputin the same comment (and I guess this might be relevant too in an article related to LLMs/generative AI) for the purposes of comparison should be an obvious exception to any prohibition on contributing unreviewed output.Thryduulf (talk)13:46, 8 December 2025 (UTC)[reply]
I'd say that sort of thing falls under a common sense exception anyway, no? Even if a total restriction on unreviewed LLM content does manage to pass, I think any sensible editor would understand that an excerpt of unreviewed LLM text posted in a discussion context for demonstrative purposes doesn't count; any more than quoting a sourced insult made about someone in a BLP is a BLP violation.Athanelar (talk)13:51, 8 December 2025 (UTC)[reply]
@Athanelar (sorry, it looks like I should have edit conflicted with you when I posted by comment below but I didn't and I've only just spotted your reply) I agree it's the sort of thing thatshould fall under a common sense exception, but given the volume of comments in discussions about AI that explicitly or implicitly state there should be no exceptions whatsoever and (in some cases) a lack of belief that common sense exceptions can even theoretically exist when it comes to AI, I'm not convinced that it doesn't need to be explicitly stated.Thryduulf (talk)16:31, 8 December 2025 (UTC)[reply]
Actually there are a couple of other exceptions too -
Posting the unreviewed output from various LLMs given the same prompt as a comparison between them
When the unreviewed output is itself notable, or part of a work that it is notable because it contains such
When quoting a reliable source (e.g. if it is encyclopaedic and due to quote the reaction of $official to $event(which is less often than such quotes are often used), then that official's quote being unreviewed LLM-output should not be a barrier to inclusion).
Given how stringently and literally some editors (wish to) interpret LLM-related guidelines, it is probably necessary to explicitly list these exceptions somewhere.Thryduulf (talk)13:54, 8 December 2025 (UTC)[reply]
Support perTazerdadog andToadspike, definitely a step in the right direction; the process will continue and we will refine this guideline. I do not quite get most of the oppose arguments. Invoking "there is no deadline" (a valid concept about encyclopedic content) for a guideline is unconvincing, while arguing that there has not been enough discussion guarantees that we will never move forward given how fraught this topic is. We should be thankful to the people who deal with the clean-up and help instead of losing sleep over technicalities in wording that can be refined as needed. The infinitesimal minority of editors who know how to use an LLM constructively (spot-checking, summarizing, rephrasing, etc.) are never going to paste any kind of output into articles since it will onlyinform their edits instead of being part of their substance, which is what we are trying to avoid. Let's keep our eyes on the prize here.Choucas0 🐦⬛⋅📬⋅📜14:34, 8 December 2025 (UTC)[reply]
I personally do not see this as a direct contradiction, and while it could be clearer it is far from a blocking issue in my book. This guideline is not perfect yet (none is or can be) but what matters to me is that its practical implications are helpful in the short term and that it is a stepping stone in the right direction.Choucas0 🐦⬛⋅📬⋅📜14:48, 8 December 2025 (UTC)[reply]
@SuperPianoMan9167 responding to every supporting !vote with this argument isbludgeoning. You've made the argument plenty of times here already, please trust that the people !voting in support have already thoroughly read the thread and haven't been compelled by this point.Athanelar (talk)15:08, 8 December 2025 (UTC)[reply]
Oppose: The proposal seems like a half-baked draft. First, there is no lead text. Second, "This includes tools marketed as "AI chatbots" or "AI writing assistants", such as...". Even if a guideline is in itself not an article, this goes against the guideline onexpressions of doubt. Those softwares are not "marketed as" chatbots or writing assistants, theyare (besides, "AI chatbot" is a tautology, there are no non-AI chatbots). Third, "Editors should not paste raw or lightly edited LLM output as a new article or as a draft intended to become an article". Another tautology: with some very limited exceptions (drafts intended to be something else instead of articles, such as the one being discussed here),all drafts are "intended to become an article", or they would have to go to miscellany for deletion. Meaning, AI is not allowed even in drafts. Which is a questionable proposal (and also forgets thatDrafts are not checked for notability or sanity), but later says "Where content is largely or entirely based on unedited or lightly edited LLM output, it may be draftified". If the proposal is actually that AI contentis allowed on drafts, with the caveat that it has to be reviewed or improved before going live, then it has to be rewritten for clarity. If "a draft intended to become an article" means a draft inWikipedia:Articles for creation that has been submitted for review but isn't approved yet, be specific (and don't forget that "pasting raw or lightly edited LLM output" in a draft would be step 1 before starting to edit and review it, so don't drop it among all the other copypasted prohibitions). Fourth, "...especially where the content is unverifiable, fabricated, or otherwise non-compliant with existing Wikipedia policies. So, that means that AI generated text may be removed even if verifiable, non-fabricated and compliant with existing Wikipedia policies? Fifth, "Tag the page as LLM generated under Template:AI-generated.", links to the template directly instead of using TL for easier copypasting. And then there's the missing content: describe the acceptable uses. Some people propose that users should disclose the use of AI, but if the guideline is all "don't"s and no "do"s, the inevitable outcome would be that all such edits would be deleted or removed by users with a strong anti-AI sentiment, and then nobody would disclose anything (and with no reliable AI-detecting tools, that's probably not the thing you would want). By the way, the idea of promoting a guideline against AI got the better of several users, and supported it without noticing the small detail that the proposal was just a useless stub. This proposal is very little better, and it wouldn't pass AFC if it was an article. Many say "let's promote it anyway, we'll fix it later". No:now is that "later". Policies and guidelines are not articles, and the wiki process does not work the same: big changes require consensus, and more so if the page has to be rewritten from the ground up. We are having this RFC because of the short-sight of approving a useless stub guideline, and if we promote this mess we'll have yet another one in a couple of weeks to fix this one as well.Cambalachero (talk)15:34, 9 December 2025 (UTC)[reply]
Err, this is literally aquickly thrown together [guideline] to "fix" issues that isn't even clear about any of what it's trying to fix, why it's trying to fix it, or how it will fix it.Thryduulf (talk)23:13, 9 December 2025 (UTC)[reply]
We kind of need these "bludgeoning" comments when people make confusing or incorrect statements (though all of us have had the experience of posting a comment in the wrong tab or mixing up which page we're talking about – you were probably thinking ofWP:AISIGNS, which is almost exactly two years old now).
Levivich gave us the dates for the new guideline, but if you're opposed to "some quickly thrown together policy to "fix" issues", then I'll point out that what you've voted to "support" was just 10 days old when this RFC started (started on the same day the new guideline was promoted), and the OP made major changes to it just half a day before starting this RFC. The resulting lack of in-depth pre-proposal review means there are still some significant weaknesses, and this is one of the reasons some editors have voted to oppose its adoption for now. I'm personally skeptical that what we need is a long guideline (I think a big box at the top of the editing window that says the equivalent of "No chatbots! 🤮" would be more useful), but if we're going to have a long guideline, we need to polish it up a bit before making a newWP:PROPOSAL.WhatamIdoing (talk)18:55, 11 December 2025 (UTC)[reply]
Support But go further. All use of LLM's to add or change texts should be disclosed in the edit summary. Not following this rule should result in an immediate indef which can only be lifted after aknowledging LLM use, promising to not use LLM's again and promising to clean up all previous use (no other edits befor this is done). Merely using a LLM as a search engine to find sources is fine and shouldn't be disclosed. There should be a banner warning editors before an edit is published.Rolluik (talk)12:37, 10 December 2025 (UTC)[reply]
I couldn't support that - one incorrect accusation of undisclosed LLM use (whether in good or bad faith) would mean either an unappealable indefinite block for doing nothing wrong or forcing people to lie in order to be unblocked. Neither of which are even on the same planet as things that are good for Wikipedia.Thryduulf (talk)13:59, 10 December 2025 (UTC)[reply]
If I was repeating the same comments in response to the same opinions being expressed, then yes it might be. This comment however was definitely not. It was a response to something that was different (qualitatively and in degree) to anything brought up previously, and my comment was focused entirely on that new argument.Thryduulf (talk)17:11, 10 December 2025 (UTC)[reply]
Bludgeoning isn't just repeating oneself, it's also taking up space in/dominating a discussion and pushing back on every comment one disagrees withKowal2701 (talk)17:54, 10 December 2025 (UTC)[reply]
I think Thryduulf is fine here; as they've said, they're being selective about who they respond to and how, not pushing the same counterpoint again and again.Athanelar (talk)17:28, 10 December 2025 (UTC)[reply]
Support. It's an improvement over what we have now. TheDo not use an LLM to add unreviewed content section seems perfectly clear to me in prohibitingraw or lightly edited LLM output but in any case this expanded version will be easier to refine going further. The fact is that most of our efforts at policies surrounding LLMs have stalled because people have demanded perfection or dragged their heels on anything that doesn't completely encompass everything they want; in particular I think it's extremely unlikely that we would reach a clear consensus on unambiguously allowed usages (there are too many people who want it banned in all cases for a consensus to formexplicitly allowing any use, while not being enough to actually ban it in all cases; this is what has stalled out all previous discussions. And if you feel youcan get a consensus to add specific allowed use-cases, it can always be done in a separate RFC - there's no need to try and do everything at once, that's another reason why previous efforts to make this policy fell apart.) Refining and narrowing the allowed usages over time until we hit points where we can't reach a consensus to narrow them further is the easiest way to work around these problems. And in this context the "contradictions" people complain about are a good thing - any ambiguities will be resolved over time as the guideline enters use and practice evolves, with resolutions to them eventually making their way to the page itself. We had our chance to get an exhaustive from-above LLM policy in the past and it fell apart after months of discussion (partially, to be extremely blunt, due to disagreements between people opposing here.) So this is what we get. Is it perfect? No. But it's workable and is an incremential improvement over what we have, which itself was an improvement over having nothing. This is how building policies for controversial aspects of the wiki work. --Aquillion (talk)14:25, 11 December 2025 (UTC)[reply]
Support -the proposed version is compact, gets straight to the point, and is a solid first-stage expansion compared to the one-sentenceold version. The concern (expressed several times above) of self-contradiction is incorrect:Do not use large language models (LLMs) ... add unreviewed content to existing articles is not strictly complete, since taken alone, it allows for adding "reviewed" content, i.e. it literally allows for adding "lightly reviewed" content. However, the content of the guideline clarifies thatEditors should not: ... Paste raw or lightly edited LLM output .... Any human fluent in English who reads this will understand the implication thatheavily edited LLM output can, in principle, be accepted (and that it's up to the editor to do the heavy, thorough editing of the LLM output). There is no contradiction. An "in a nutshell" summary can acceptably be incomplete (which is the case here); that does not make it contradictory with the body of the article; it is not intended as a complete, nuanced summary. It's aTL;DR. It gets to the essence of the guideline. These are instructionsfor humans, not for semantic engines, and not for LLMs either.Boud (talk)00:52, 12 December 2025 (UTC)[reply]
Any human fluent in English who reads this will understand the implication that heavily edited LLM output can, in principle, be accepted (and that it's up to the editor to do the heavy, thorough editing of the LLM output). multiple humans fluent in English have explained that this is not the case.Thryduulf (talk)05:28, 12 December 2025 (UTC)[reply]
What is the difference between "heavily edited" and "lightly edited"? How much edited is enough edited? How do we know how whether editing was heavy or light? And I don't just mean how do I know whether someone else heavily or lightly edited, I mean how do I know whetherI heavily or lightly edited? So not just how do I monitor others' compliance with this guideline, but how do I monitormy own compliance? I think the answer is "you don't know, it's impossible to measure with such vague standards," which is why this proposed guideline shouldn't be adopted.Levivich (talk)05:45, 12 December 2025 (UTC)[reply]
Well, no, I think the answer is always going to be 'you'll know it when you see it' because depending on the model used, volume of text generated etc the amount of review necessary is always going to vary. That's why I'm of the mind that ultimately our prohibition shouldn't be based on human review at all, since it seems to me the only sensible threshold for "well reviewed" is "indistinguishable from non-LLM text anyway"Athanelar (talk)10:30, 12 December 2025 (UTC)[reply]
Weak support. Needs more time to bake, we need to actually agree on what to put in such a policy to actually have a consensus other than oppose too weak and oppose too strong arguments~2025-31733-18 (talk)19:25, 13 December 2025 (UTC)[reply]
This is to modify an existing guideline, not create a new policy. Because it already exists, this would be exactly the place to talk about if the text is to strong or weak, since it is modifying an existing.~2025-38536-45 (talk)17:26, 15 December 2025 (UTC)[reply]
This is an RFC, which comes after the workshopping stage. That there was insufficient workshopping of this proposal prior to the RFC does not mean this is a workshop.Thryduulf (talk)19:43, 16 December 2025 (UTC)[reply]
If it quacks like a duck, it is a duck. So no, this is also part of the overall workshopping. Just less official, but none the less, the same purpose. Unless you want to argue semantics I guess? That feels like a waste of time though.~2025-38536-45 (talk)21:03, 17 December 2025 (UTC)[reply]
The point of an RFC is to establish consensus on a specific thing, to that end changes to the wording of that thing are generally strongly discouraged because they are almost always unhelpful (because it is unclear what someone is actually supporting or opposing). This is in complete contrast to a workshop where there are frequent, often iterative, changes to whatever is under discussion and no attempt to solicit support or opposition to adopting the proposal (and indeed bolded "support"/"oppose" etc. comments are frequently called out as incorrect and/or unhelpful). The two processes look nothing alike, and this is both explicitly and by all appearances an RFC not a workshop.Thryduulf (talk)21:21, 17 December 2025 (UTC)[reply]
Oppose I think its the right direction but missing a few steps. For example, thereare good uses of LLM when you have actually already crafted text but want the LLM to make it more concise (that is, you're not asking for it to create information but rephrase it), or using LLM to actually do initial searching. As long as the editor using the LLM is taking responsibility for any problems it creates (including hallucinations or non-encyclopedic tone or whatever) and corrects them if present, that should not be considered harmful.Masem (t)00:55, 14 December 2025 (UTC)[reply]
Support, it's a great improvement on the existing guideline, which covers only new articles; however, inserting text into existing articles is at least as much of an issue. Things can be tweaked later if needed, but we can't let every quibble hold us back from having a strong guideline on adding AI-generated text to existing articles.Crossroads-talk-22:24, 15 December 2025 (UTC)[reply]
Why do you regard the multiple, fundamental objections to this guideline from multiple independent good-faith editors as "quibbles"? What makes a guideline that many read as self-contradictory "strong"? I'm not trying to bludgeon here, I'm genuinely trying to understand your viewpoint.Thryduulf (talk)23:51, 15 December 2025 (UTC)[reply]
The fundamental issue is that we need a policy addressing people credulously plopping unedited or barely edited AI output intoexisting articles. Right now LLMs have been around for years and there is still no policy/guideline squarely forbidding that. Any subsidiary aspects can be discussed and adjusted later, but statements likeEditors should not:...Paste raw or lightly edited LLM output into existing articles as new or expanded prose are sorely needed.Crossroads-talk-22:29, 16 December 2025 (UTC)[reply]
Please see thePolitician's syllogism - just because (you believe) something should be done does not mean that this is the correct something. If you think itis the correct something, you should address the issues withthis specific proposal brought up by those in opposition. Many (perhaps most) of those opposing agree with your aim, (almost?) everyone agrees that the aim of this proposal matches that aim, however there is significant disagreement that this will achieve that aim and/or that it will not cause more problems than it solves. Almost none (but not quite all) of the support, including your two comments, does not address those issues.Thryduulf (talk)01:52, 17 December 2025 (UTC)[reply]
Support — While I have some minor quibbles over the wording and I'd still prefer a blanket ban on LLM use, this is a goodsecond step towards codifying best practices at dealing with LLM slop.—pythoncoder (talk |contribs)02:19, 16 December 2025 (UTC)[reply]
Support per fifteen thousand two hundred twenty four – If the editors doing a great service to this project by devoting themselves to the cleanup of AI slop think that this is a good change, we ought support it. Anything to give them more tools to tackle what is an existential threat to the project.Yours, &c.RGloucester —☎20:23, 16 December 2025 (UTC)[reply]
Support; I don't see anything substantial to disagree with here. I agree it is challenging to police and to prove LLM use in many cases, but there is still value in a clear statement of principle saying we don't want it - cf our policies on conflicts of interest, etc. I doubt this will be the final policy we have but it is a decent next step. Yes, it is imperfect, but the issues with it do not seem fatal to me, and something like this has been needed for a long time now.Andrew Gray (talk)04:49, 17 December 2025 (UTC)[reply]
The issues is that many people disagree that hit isa clear statement and that statements of principle and guidelines are interchangeable.Thryduulf (talk)14:48, 17 December 2025 (UTC)[reply]
Thryduulf, not every !vote needs your opinion. Seeing an admin argue with every third to forth vote is incredibly discouraging to newer editors who disagree with you, please chill.~ Argenti Aertheri(Chat?)23:04, 17 December 2025 (UTC)[reply]
I was going to say the same thing, but I was getting toasted by autoblock at the time.
The issues is that many people disagree that hit is a clear statement and that statements of principle and guidelines are interchangeable and clearly Andrew Gray does not share this opinion. This has passedWP:BLUDGEONING for which you have been warned repeatedly on this thread. It's really not fitting for an administrator to have to be reminded of conduct rules by multiple non-admins.Cremastra (talk·contribs)01:08, 18 December 2025 (UTC)[reply]
not every !vote needs your opinion. which is why I'm not giving my opinion on every vote, justsome of the ones that appear not to have taken any note of the prior discussion and/or seem not to have engaged with the fact this is a proposed guideline (i.e. intended to provide guidance) not a proposed general statement of principal or philosophy. The number of accusations of bludgeoning is irrelevant to whether someone actually is bludgeoning, and you'll note that the previous accusations have all been disagreed with by those who understand what it actually is.Thryduulf (talk)03:19, 18 December 2025 (UTC)[reply]
By my count at the time of the above comments you'd participated in the replies of 11 out of 28 supports. It's true that 39% is a long way from every support, but it's still a high amount. If you find yourself needing to preemptively clarifya comment with a statement likeI'm not trying to bludgeon here, it may be worth stepping back and letting the discussion breathe.(corrected oppose->support 08:10, 18 December 2025 (UTC))fifteen thousand two hundred twenty four (talk)04:15, 18 December 2025 (UTC)[reply]
It's not just that, count the number of support votes with replies from anyone except Thryduulf. Since SuperPianoMan chilled virtually every support vote that was commented on was initially commented on by Thryduulf and, generally, no one else. So no, it's not every !vote, but any support vote that has been challenged has been challenged by the same admin. Idk if it's capital B bludgeoning, but please stop dominating a discussion that isn't about you.~ Argenti Aertheri(Chat?)05:30, 18 December 2025 (UTC)[reply]
Whether your conduct meets a literal definition of bludgeoning is irrelevant, you should probably read the room when a lot of people are telling you they don't feel you're behaving properly; and I say that as someone who initially defended you from the charge of bludgeoning.Athanelar (talk)14:06, 18 December 2025 (UTC)[reply]
I strongly agree with your position on the RfC but this has definitely become bludgeoning and I am concerned that you are arguing otherwise. I have also noticed this in other LLM related PAG discussions. Please take this feedback from the community?NicheSports (talk)18:16, 18 December 2025 (UTC)[reply]
I would agree with the concerns raised above. If I may be direct: you are acting like you are the sole arbiter of what are acceptable responses, and can challenge the ones that you don't think meet your standards. When you say Iappear not to have taken any note of the prior discussion orseem not to have engaged with something, then it comes off as, atbest, patronising. It is not conducive to a healthy discussion and I can easily see how some people would be intimidated by that or put off commenting.
Regarding my original point: of course I read the discussion and thought about the points raised. I wrotethe issues with it do not seem fatal to me, indicating I had considered those issues, acknowledged theywere problems, but disagreed with the argument that they are sufficient reason to reject this proposal. Personally, I think it is perfectly acceptable to adopt a guideline that is known to be imperfect or slightly ambiguous, becausethis is a wiki and we expect that it will get refined and updated as time goes on. If people are having trouble with the wording, we can discuss and clarify it. As long as we have a core idea of what that guideline is trying toachieve, our normal processes can handle that revision and refining. We don't need perfection on day one, and insisting on that is a sure way to ensure we never get anything decided.
And this is why I used the termstatement of principle, which appears to have caused some confusion. Many of our policies and guidelinesare in and of themselves making broad statements of principle, as well as giving detailed guidance on implementing those principles. Consider BLP, for example - the principle that establishes is (very simplified) "BLPs are sensitive and we think it is particularly important to get them right".
As with BLP, this proposed text does have a clear thread running through it - "Wikipedia is skeptical of the value of LLM edits, and we want to discourage editors from using these tools to create article content". Yes, there are ambiguities in the detail, but I don't think anyone would read the draft and come away thinking it is fundamentally neutral on LLMs, or that it means LLMs are encouraged. Hence my description of it as aclear statement, because I was talking about the broad direction, not the smaller details.
Support. This follows the current expectations generally held by the community, and codifying it is an important step. I have yet to see a strong argument as to why the wording "raw or lightly edited" doesn't resolve most of the opposition, and this really feels like an excuse to nitpick. If we want to tweak the wording, add acceptable use cases, or anything else, we can resolve that with additional discussion afterward. Baby steps. Otherwise this is howpolitical gridlock happens.Thebiguglyalien (talk)🛸01:13, 20 December 2025 (UTC)[reply]
OpposeThis speedily written proposal has already been abandoned by its creator in favor of better-written proposals on the guideline's talk page. I still think the new proposals are much better, and we shouldn't settle for less.Mikeycdiamond (talk)12:26, 23 December 2025 (UTC)[reply]
This is a wild mischaracterisation that I think you should strike for risk of it prejudicing further !votes here.
In the thread you linked, Qcne explicitly saysNo,[I am not presenting a new RfC here], I was just trying to gather further thoughts as my actual RfC got a bunch of criticism Italicised context is mine.
Qcne has made no statement to the effect that they're abandoning this proposal, and even if they did, it has received significant support here that makes 'the creator doesn't like it any more' not really stand as an oppose reasoning.Athanelar (talk)14:50, 23 December 2025 (UTC)[reply]
Support as a very welcome change that, admittedly, did need some more time to bake, but is still much needed. I don't think that it's worth delaying just over those minor wording squabbles, and I don't think it's worth adding content about acceptable uses of LLMs at all.MEN KISSING(she/they)T -C -Email me!00:18, 25 December 2025 (UTC)[reply]
Support. Perfect is the enemy of good; I have seen enough LLM content from NPP and in discussion space sucking up my time. I welcome anyone who would like to join us on the front lines.Iseult Δxtalk to me02:52, 27 December 2025 (UTC)[reply]
Oppose. I share several of the concerns above, but my main objection is that LLM output (especially if edited by competent humans) is very difficult to reliably identify as such, and will likely become more difficult to identify as models improve. This will lead to many allegations of LLM use that are very difficult to prove and will likely cause heated disagreements. For very obvious LLM slop, we already haveWP:G15. Sandstein21:28, 27 December 2025 (UTC)[reply]
Support but only because this is better than doing nothing. The people producing the bad work are not competent to follow these nuanced instructions, or else they brazenly exploit the nuances to lawyer their way out of trouble. Wikipedia as a whole is not competent to successfully police these sorts of picayune regulations on this topic; we are proposing to use a dinner fork to clean the ocean. An absolute and draconian blanket prohibition on anything resembling AI, even where it's demonstrably useful, is the way to go. But I'll take this, as being better than the previous.TooManyFingers (talk)03:40, 29 December 2025 (UTC)[reply]
support. this gives us the tools to deal with unreviewed llm output, which at best wastes editor-time, and at worst spreads misinformation. there are some suggestions above for legitimate uses of llms, but none of them actuallyrequire llms. and, as suggested by sandstein, if llm output somehow becomes indistinguishable from human writing, wikipedia guidelines will be the least of our worries. i would also recommendpromoting to policy, to make our stance against llm slop official.ltbdl (bite)05:51, 31 December 2025 (UTC)[reply]
Oppose. I agree that the RFCBEFORE should have lasted longer, and that it needs more work. It feels too formal and inaccessible, almost like lawyerly writing. The Flesch-Kincaid readability score is university-level, worse than all out content policies save one, despite being so short. And yet despite being short, it feels way too wordy. The "scope" section feels like it could be one sentence. We can just say "artificial intelligence tools, like ChatGPT". The reader is unaware of our policies, but she isn't stupid. We specifically excludenon-AI spellcheckers, even though a guideline on "AI" excludes them by definition.
It's self-contradictory in the most important section. It starts with "no use" of LLMs when creating an article or editing an existing one, then backtracks to no useas the primary author for new articles or expansions, and then backtracks again to prohibitingraw or lightly edited LLM output (wouldn't any amount of editing still leave ChatGPT asthe primary author?) The contradictions might lead to flamewars among editors, but a bigger concern is that it's vague enough to harm the purpose among its intended audience: new editors unfamiliar with our expectations. Will college students who habitually use ChatGPT for essays understand that we would consider their typical editing to be "light", and not "medium-ish"? That their honed, "clever, savvy" editing, "good enough" to fool plagiarism detectors, falls well below the expected bar here? We also move the main point of the previous guideline: don't generate articles with LLMs, prominent and in bold, down to barely above the fold (on a laptop), non-bold, with muddier wording. Worst IMO, we only list (quite specific) disallowed usecases, but omit the central point:You are responsible for your edits. They must followWikipedia's content policies, whether or not you use artificial intelligence tools like ChatGPT. "You must follow policies" might feel redundant to us. But for its intended audience it gives a clear takeaway, clear expectations, makes the logic behind the rule clear, and feeds them the next step (click that blue link and learn about our policies). And we could workshop it so it can also pleases the "don't implicitly endorse LLM use" concerns from the previous discussion, if the workshopping wasn't over.
For an old established site with rarely-changing policies, passing a new guideline is less hard than making new users widely aware of it (at least vaguely aware). A wordy, confusing, college-grade guideline doesn't help. Low awareness means no reduction in the amount of cleanup work needed, which is the entire point. Since I think this won't be effective at its primary purpose: oppose. I can think of some ways to expand the current guideline to a few lines, that would match the prevailing sentiment of previous discussions while being clear and concise, so it's too bad the workshopping is over. But the current guideline, despite its flaws, has the merit of being written for a target audience of new, overwhelmed, inexperienced users, and not a target audience of experienced Wikipedia editors. I'd prefer a clear, concise guideline, linked from every Edit box like WhatamIdoing proposes.
[More minor note, the widespread reassurances on "we can fix it after we pass it" from the support side, while still in the RfC stage, concerns me. I get that we're late, but that isn't how major guideline change RfCs should work.]DFlhb (talk)09:34, 31 December 2025 (UTC)[reply]
Strong support. I am quite sympathetic to the staunchly anti-LLM camp, but as the currentWikipedia:Writing articles with large language models reads, it is unclear what LLM even refers to, whereas the newer version explains different examples of it, and also how it can be harmful and disruptive. The concrete examples and wording is much more accessible, along with remedies in case someone does encounter such an article. Ironically for the critiques of qualifiers, the original guideline says it should not be used to create articles from scratch, implying that in all other cases, uses of LLM are permitted. ~ 🦝Shushugah (he/him • talk)01:24, 3 January 2026 (UTC)[reply]
Oppose unless this sentence is removed or modified to only apply to unreviewed content: "Editors should not use an LLM to add content to Wikipedia, whether creating a new article or editing an existing one." If interpreted literally, that sentence strongly discourages almost any usage of LLMs to add content, including benign usages like fixing the grammar. Sure, interpreting it literally would make most of the rest of the section redundant, which is absurd, but that wouldn't prevent some folks from quoting that misleading sentence all over the wiki to win debates. If the guideline passes, getting consensus for removing (or nuancing) that sentence risks being difficult. That's unfortunate because the rest of the section was making important points.Alenoach (talk)18:06, 3 January 2026 (UTC)[reply]
Oppose. As others have argued, the proposed guideline is overall poorly written and even self-contradictory, not just as to "minor" wording but even as to fundamental points. It is worse than the status quo for clarity.Adumbrativus (talk)11:30, 5 January 2026 (UTC)[reply]
Support - I agree with some of the criticism. The first section explaining LLMs is too long before getting into the point. But it's clearer than the previous guidelines and more concrete. I would support a stricter ban on LLMs in Wikipedia as I'm seeing a lot of bad effects doing NPP and AfC.Lijil (talk)13:54, 6 January 2026 (UTC)[reply]
Strongly Oppose. While I support the goal of minimizing "AI slop" and agree that the currentWP:NEWLLM is somewhat vague and in need of clarification, this specific proposal is far too broad and strongly worded. In particular:
Prohibition on discussion usage: I oppose the blanket ban on using LLMs to draft comments or replies. LLMs are essential accessibility tools for neurodivergent editors or those whose first language isn't English; they help summarize "walls of text" and assist in framing arguments within the specific language of Wikipedia policy. We should regulate the quality and civility of comments, not the tools used to draft them.
Flawed "lightly edited" metric: The focus on the amount of editing brought to the text is a poor proxy for quality. The priority should be the depth of human review and the accuracy of the final output. If an LLM-generated draft is reviewed in depth and found to be policy-compliant and factually correct, requiring the editor to "change" the prose simply to avoid it being "lightly edited" is counterproductive toWP:V.
Ongoing translation discussions: Not incompatible with the current proposal, but I'd like to call out that ongoing discussions are taking place inVillage Pump regarding allowing the use of LLMs for translation use cases, which will likely lead to a separate RfC.7804j (talk)14:28, 6 January 2026 (UTC)[reply]
I oppose the blanket ban on using LLMs to draft comments or replies. LLMs are essential accessibility tools for neurodivergent editors or those whose first language isn't English; they help summarize "walls of text" and assist in framing arguments within the specific language of Wikipedia policy. We should regulate the quality and civility of comments, not the tools used to draft them.
They can also introduce distortions that a ESL writer probably won't notice. This will only cause more misunderstandings and unwarranted conflict.
I also might note that using LLM-formatted bold lists with a very LLM-sounding writing style in a discussion on restricting LLM usage is not a good look.Cremastra (talk·contribs)14:57, 6 January 2026 (UTC)[reply]
@7804j In addition to what Cremastra said, using an LLM to summarise a 'wall of text' that one might encounter in a Wikipedia policy/guideline or to formulate a policy-based argument is a terrible idea, because LLMs have been known to misapply PAGs because they don't actually know what they say. I've seen an editor contend their block by arguing that becauseWP:COMMUNICATE says that admins have to warn someone before they block them (it doesn't) the admin acted wrongly bt outright blocking them.
Plus, if someone doesn't have the ability to understand Wikipedia's PAGs and participate in discussions, whether due to language skills or neurodivergence, then they lack therequired competence to participate at all. 'Accessibility tools' are not supposed to overcome a lack of competence.Athanelar (talk)15:35, 6 January 2026 (UTC)[reply]
(also, Cremastra, I was horrified to see you making this argument in my notifications until I opened it and saw it was a tq... phew!)Athanelar (talk)15:36, 6 January 2026 (UTC)[reply]
To clarify, the structured layout of my previous post is simply my personal communication style, which is common in professional and corporate environments for conveying complex points clearly. The fact that my comment is being scrutinized for its formatting "look" rather than being engaged with on its actual merits is a perfect illustration of why I oppose this proposal. It invites a culture of "witch hunts" where editors are judged and dismissed based on subjective "vibes" or writing styles rather than the substance of their arguments.
This shift toward policing how a message is written instead of what it says is exactly what I fear a formal ban will encourage. Because LLM output is increasingly difficult to distinguish from human writing, such a policy will inevitably lead to good-faith editors being harassed or ignored based on hidden "clues" that may not even exist.
Regarding competence, I agree that LLMs can distort policies, but humans are equally capable of misinterpreting or misquoting policies. In either case, the editor remains personally responsible for the content they post. We already have policies to handle editors who consistently provide incorrect information or refuse to communicate effectively. An accessibility tool is meant to help bridge a gap, not replace the need for competence; if the resulting post is wrong, the editor should face the consequences for that content, regardless of whether they used an LLM, a grammar checker, or a dictionary to help draft it. We should focus on the quality of the contribution rather than trying to ban a tool in a way that is unenforceable and creates unnecessary friction over personal prose styles.7804j (talk)12:06, 7 January 2026 (UTC)[reply]
Strong support. While the policy could use a note clarifying direct translation from human-written text in one language to another using NLP is fine, as someone who has spent some time inWP:AFD, the use of these chatbots to generate Wikipedia articles in whole or in part is extremely disruptive. If there are people using LLMs and it's not noticeable, this policy won't come up. The problem is when it is noticeable it is because the citations are invented, the content is entirely fictional, and/or the content does not accurately summarize the cited references (and in many cases, the subject matter isn't even in the source, even if the prose is itself factual). When you multiply theWP:REFBOMBs it creates, times the number of articles this is on, we are all becoming unwitting janitors to the lazy slop infecting this encyclopedia. And the responsibility to clean it up is high stakes:[6].— Precedingunsigned comment added byRevolving Doormat (talk •contribs)
Comment. I find the arguments in opposition about using AI to fact-check Wikipedia or perform other pre-writing tasks such as proof-reading to be irrelevant and poorly based.
Most AI tools, particularly the ones described in this proposal (Chatbots), are not designed to do that. Second of all, the study involved human evaluation before determining whether or not it got it right or wrong. Indeed, it got wrong a non-negligible number of times. ObviouslyGrokipedia is an example where an AI tool designed to do this without intervention confirms misinformation as factual and fails to correct human-identified errors by claiming it verified it with its fact-check; and at times generating fictitious citations.[7] This is a mathematical inevitability.[9] Thus, human intervention inAI finds errors in 90% of October's TFAs overrules any notion that it should be used to add content to Wikipedia. The general sentiment is that this type of encyclopedia is a horrible idea[8] and that LLMs lack the complex cognitive web beyond language patterns to have any epistemic vigilance.[10]
There is nothing in the proposal that disallows the use of AI Tools for pre-writing tasks involving content to be added to Wikipedia afterward. As described in point 1, all of these require human-intervention post-LLM processing. AI Tools cannot fact-check, alter prose structure and word choice, or otherwise engage in cognitive offload tasks without error, thus whatever it generates should not be added to Wikipedia without human editing—thus making it outside the scope of this policy.
The concerns about proof-reading are unfounded becausethere exist AI proof-reading tools(i.e., Grammarly)that are outside the scope of this proposal as they are not marketed as Chatbotsor writing assistants. Regardless, the content generated bythose toolslike Grammarly should also undergo human-editing as mentioned in point 2 and simply flag where this proposal needs to be extended in future iterations.
The only legitimate concern is translation as I said in my vote. However, there areNatural Language Processing tools for this such as Google Translate while we consider incorporating it in later iterations.
Thus, there is no case in which a Chatbotor other LLM-based writing assistant, the type of AI Tool described in this proposal, should be used to add content to Wikipedia. Certainly not togenerate ideas for new articles. There is certainly no evidence that AI Tools can identify salient information (only statistically likely patterns of text), let alone evaluateWP:NOTABILITY which is glaringly apparent from my time in AfD. Still, even then, this applies to point 2.Revolving Doormat (talk)12:38, 7 January 2026 (UTC)[reply]
Comment. It also occurred to me that another argument in the oppositions are that humans also createWP:OR,WP:SYNTH, and add uncited fictitious content to Wikipedia absent any AI assistance. That is awhataboutism andfalse equivalence error because AI removes the barrier of effort for adding content to Wikipedia in the first place. This means that people who mean well but would not otherwise add such disruptive content in the first place are multiplying making rooting out such problems an ever-growing game ofwhack-a-mole. Instead, people who need help adding content would utilizeWP:AFC orWP:Teahouse, rather than a Chatbot or AI writing "assistant."Revolving Doormat (talk)13:44, 7 January 2026 (UTC)[reply]
Oppose Due to the broad and general nature of this proposal, it is impossible to determine with sufficient certainty whether someone used an LLM for a given edit - let alone to what extent they modified the original output. Until we have reliable tools, procedures, or a proven method to identify LLM usage that the community agrees is effective and won't harm editors with false accusations, such a restrictive proposal carries significant risk.
It would likely lead to constant accusations based on "vibes" and circumstantial evidence, sparking conflicts that drain moderation resources. New users, who increasingly live in an environment surrounded by AI, may fall victim to these accusations - whether false or partially true/true - which will only serve to scare away potentially valuable contributors (WP:BITE).
Secondly, it is far more logical to hold editors accountable by evaluatingwhat they added (the output), rather than conducting an "investigation" intohow they wrote the text they chose to publish. ExistingWikipedia:Policies and guidelines, such asWP:V (Verifiability) andWP:NOR (No original research), are already sufficient to handle content that fails to meet our standards.Grudzio240 (talk)11:30, 7 January 2026 (UTC)[reply]
We draw a line between apparent (unproven) and actual (proven)WP:COIs. There may always be editors who seeWP:LLM in places where it's not, just as with COI. However, most of the time it is actually quite obvious. And while we do have things in place, they are simply not sufficient because no one should have to check if 30 or more citations to see if the rest of the content is legitimate when it's obvious it's AI. These articles become a slog to fix when they should never have been published to begin with. We need guidelines and policies to deter people from these disruptions.Revolving Doormat (talk)19:41, 7 January 2026 (UTC)[reply]
Can we all justslow down and try to discuss what the essential components of a comprehensive LLM guideline should bebefore !voting on one? NEWLLM passed less than two weeks ago.SuperPianoMan9167 (talk)23:37, 4 December 2025 (UTC)[reply]
In that case we should modify the existing guideline instead of throwing it out entirely. This proposal aims to restrict more LLM use than NEWLLM does, because it saysEditors should not use an LLM to add content to Wikipedia, whether creating a new article or editing an existing one. with no qualifier.SuperPianoMan9167 (talk)22:58, 6 December 2025 (UTC)[reply]
Yes, this proposal is pretty much deleting everything under the{{guideline}} tag and replacing it with something else. @Rutebega, the key word in my comment wasexpanding. This proposal isn't just a clarification of the wording. It's a dramatic expansion from "don't copy/paste whole articles straight out of your chatbot" to "here are a bunch of new rules for a large variety of situations".WhatamIdoing (talk)04:13, 8 December 2025 (UTC)[reply]
I had been following that talk page, saw there was a new draft but didn't have enough time to fully form any thoughts before I saw a message that there was a RfC here.
What Thryduulf and Katzrockso said - I have a day job, you know, and can't be counted on tolog in to Wikipedia every week, much less digest thousands of bytes of text from a variety of viewpoints and formulate a well-considered, clearly-articulated response. Plus I have to dedicate at least some of my Wiki-time to content-space contributions or I will lose my mind and quit the project forever. I think these AI conversations are really important and fairly urgent, but on a timescale of weeks or months, not days. --LWGtalk21:50, 5 December 2025 (UTC)[reply]
A little comment to people who like to point out 'copyediting' as an acceptable use of AI that we should explicitly permit; I've been working on cleaning up a chronic LLM-misuser's contributions as trackedhere, and I would really strongly encourage you all to go and take a look at some of those diffs. The most egregious ones are cases where the AI copyediting accidentally crams together two sentence clauses in a way that changes the meaning of a sentence, or needlessly thesaurusifies a direct quote from a real person, but even if you look at the diffs that I haven't bothered to revert (marked in red,) you'll see so many cases of entire paragraphs being needlessly thesaurusified into nausea-inducing pompous purple prose by this user whose modus operandi appears to be telling an LLM to 'translate' article text into 'encyclopedic language.' I simply don't have the patience to go in and manually reword every little sentence which has been needlessly puffed up by this process, and that's exactly the kind of 'death by a thousand cuts' that us in the anti-AI camp are trying to avoid.Athanelar (talk)19:19, 10 December 2025 (UTC)[reply]
I can show you many, many examples of serious mistakes made by editorswithout using LLMs. Should we ban all human-written content because humans have a high error rate? You talk about not having time or patience to fix it all... I spent five years, 40k edits, fixing human editors' mistakes on this website. Still, I don't favor banning humans from the website.Levivich (talk)19:24, 10 December 2025 (UTC)[reply]
At least in human-written text, the amount of effort to create it is larger than or equal to the amount of effort needed to fix/remove it. As commented somewhere above, LLM use flips that balance on its head, where users can use external tools to create large amounts of poor text that are difficult to manually fix.Cremastra (talk·contribs)19:26, 10 December 2025 (UTC)[reply]
That's not true at all. It's much easier for a human to add incorrect, non-policy-compliant content than it is to fix that content. LLMs do make it easier, that's true, and that's their danger. But they also make it easier to add policy-compliant content, and that's their advantage. It's just like a word processor, or a typewriter, in that sense: it makes it easier to write, faster to write... both writing good stuff and writing bad stuff.Levivich (talk)19:28, 10 December 2025 (UTC)[reply]
Describing generative AI as just another tool that makes writing easier, like a word processor, is an optimistic inaccuracy because it, not the human, is the one doing the writing. It is no longer the human writing with the process sped up, as in a typewriter or word processor: it is a human either telling a machine what to write or writing with machine-suggested changes and rephrasings.
Saying LLMs are just another writing tool is a bit like saying a car is just a better shoe. More complicated shoes may allow one to walk faster, or get better grip, but it is still a human process. A car is a complex machine only guided by a human. But LLMs are even worse, because they can generate content with only the vaguest human instructions: you still need to steer a car (at least, for now).Cremastra (talk·contribs)19:34, 10 December 2025 (UTC)[reply]
Itis the human doing the writingif the human being checks the LLM's output, which they should do. I mean, that's the difference between using an LLM as part of a writing process, and just copying and pasting what the LLM outputs. Nobody should be coping and pasting LLM output without checking it first, and probably without making some edits to it. If the LLM output is checked by a human, then it becomes human output, or at least hybrid output, not entirely machine output. (The ironic thing about cars is that self-driving cars are actually safer than cars driven by humans. We may get there with LLMs someday, too. Or maybe not.)Levivich (talk)19:42, 10 December 2025 (UTC)[reply]
I mean, we're getting a little outside the scope of the point of my original comment now, but I will say to that what's already been said above - the pitfall in requiring 'human review' is that there's essentially no way to prove it has or hasn't been done unless the person leaves in the most obvious signs of unreviewed output like human-directed communication, UTM footprints etc; but in cases like this with problematic and unnecessary copyediting, if there were a hypothetical guideline forbidding all AI-generated additions without suitable review, this user could simply come forward and say "well, I reviewed it and thought it was fine" and we'd have no recourse against that.
My whole argument with my original comment is that these edits very much could be "human-reviewed copyediting"; i.e., literally the poster child for acceptable AI-generated content as held up by the pro-AI (or, less anti-AI, I suppose) camp in this discussion; and they'restill generally unproductive and unnecessary in the best case.Athanelar (talk)19:46, 10 December 2025 (UTC)[reply]
There's lots of types of edits on Wikipedia that other editors perform that some editors believe are "problematic and unnecessary", is that grounds to ban them or all humans from contributing to Wikipedia?
The point here is that AI-assisted editing is not some unique class of editing that is all-good or all-bad, but that these edits vary in their quality just like any other type of editing: solely humanly generated edits (a concept becoming more dubious over time as technology is increasingly embedded in digital interfaces), edits assisted by grammar checkers and other writing software, edits assisted or produced by bots, etc.Katzrockso (talk)21:19, 10 December 2025 (UTC)[reply]
The point is that AI-assisted editing pretty muchis all bad, unless it's painstakingly reviewed and copyedited with a degree of scrutiny and editing that makes it essentially no more convenient than just writing it yourself.
If there were any other tool that pretty much universally resulted in worsening the encyclopedia in a way that required intense, focused, organised human effort by an entire wikiproject to fix, was completely useless for big tasks like article creation while in thebest case saving marginal amounts of man-hours doing minor tasks like copyediting, we'd restrict its use as much as we can without a second thought. But because AI is 'the future' or whatever and has infested peoples' daily workflow so much, we're forced to bend over backwards trying to find a way to reconcile it with Wikipedia when it's so obviously observably contrary to what we do here.Athanelar (talk)21:33, 10 December 2025 (UTC)[reply]
The point is that AI-assisted editing pretty much is all bad, unless it's painstakingly reviewed and copyedited with a degree of scrutiny and editing that makes it essentially no more convenient than just writing it yourself. Many editors here disagree with this unevidenced claim, which is the crux of the dispute.Katzrockso (talk)21:35, 10 December 2025 (UTC)[reply]
Well, I've seen absolutely no evidence to the contrary, butWP:AINB has ample evidence to the affirmative. That's what my whole top level comment here is showing; that even minor, selective, seemingly human-reviewed copyediting, the poster child for ideal LLM usage, is turning out bad results.Athanelar (talk)21:39, 10 December 2025 (UTC)[reply]
That much is true, yes, but I also haven't seen anybody give any convincing examples of how it can be used well except for some handwaving towards the possibility. The one exception is the editor whose name doesn't come to mind who has been making very thorough articles using LLMs, but I know that that editor uses a custom LLM that they trained themselves and which they review the output of extensively, so they can't exactly be taken as a type specimen. Spiders georg and so on. I understand that negative examples will naturally end up on noticeboards, but I just don't see any convincing positive examples presented anywhere else; which you'd think the pro-LLM crowd should be quite interested in producing.
SuperPianoMan linked the LLM demonstrations subsection above, but the most convincing examples there are examples of using it to 'suggest' edits to then be made by a human, rather than actually producing any wikivoice text in itself.
My contention is that LLMs are, when used by a typical user, essentially useless at producing any text in wikivoice which actually improves the encyclopedia. All of the 'ways to use LLMs well' focus oncircumventing this limitation, which is a red flag for me. I wouldn't use a calculator which spat out nonsense if you used the multiplication button so instead you had to just use addition to approximate multiplication; or at the very least, I certainly wouldn't use it for multiplication.
Though, to be transparent, even if such evidence (of LLMs being any good at article text)did exist, I'd still be opposed on principle, but that's a separate matter.Athanelar (talk)22:21, 10 December 2025 (UTC)[reply]
All of the 'ways to use LLMs well' focus on circumventing this limitation That's the point. The "ways to use LLMs well" are the ones that overcome their limitations.SuperPianoMan9167 (talk)22:39, 10 December 2025 (UTC)[reply]
Though, to be transparent, even if such evidence (of LLMs being any good at article text) did exist, I'd still be opposed on principle, but that's a separate matter. Which is precisely why producing such evidence (which does exist, by the way!) is not productive here.Katzrockso (talk)23:13, 10 December 2025 (UTC)[reply]
I mean, it would still be productive (a debate is as much about convincing the audience as it is about convincing your interlocutor) I just wanted to be frank about the fact that nobody's goal should be to convince me specifically, since I will not be convinced.Athanelar (talk)23:18, 10 December 2025 (UTC)[reply]
I think this is case-in-point for the discussion, inoculation from evidence is not a productive attitude towards discussion here.
It's not about 'inoculation from evidence,' I'm willing to accept that I'm wrong if it can be proven that LLMscan be coerced into producing half-decent wikivoice, I just think that the fundamental reason for an LLM prohibition should not be that they are 'ineffective' but rather that they are philosophically and morally anathema to Wikipedia and to any meaningful creative pursuit in general. I just focus on the more practical criticisms here because I think that's more likely to gather support and get changes implemented. So, if I am wrong that LLMs are useless at producing wikivoice without enough human review to render the effort-saving negligiblw, I still think we should crack down on LLMs, just for other reasons.
As forShit flow diagram it was, I presume, extensively human reviewed; and still has criticisms of its wikivoice on the talk page, so my criticism that it requires equivalent human effort to verify AI wikivoice (and to clean up when it goes wrong) as it would to just make it the old fashioned way still stands.Athanelar (talk)04:26, 11 December 2025 (UTC)[reply]
The criticisms were agreed to have been flawed. When you search through any text, including those written by a human, to search for flaws, you typically will find some. That the text there had at worst one stylistic choice that was remedied is strong evidence that LLMs can be used to produce text in a policy compliant manner.Katzrockso (talk)07:06, 11 December 2025 (UTC)[reply]
If humans were able to produce perfect text with zero errors on the first review, GA and FA would be very significantly quicker and easier processes. It is unreasonable to hold reviews of human text and reviews of LLM text to different standards.Thryduulf (talk)12:17, 11 December 2025 (UTC)[reply]
ReI also haven't seen anybody give any convincing examples of how it can be used well: I'm sure the article I created yesterday,Idéologues, contains mistakes, and I'd be grateful if anyone could point them out. I still think (and hope) it's a good article, even though it's largely based on LLM-generated text. I carefully checked the output and sources and rewrote large parts of the text, which makes me wonder whether the LLM actually helped or just wasted my time... it's hard to avoid the many subtle errors ChatGPT consistently introduces. Still, I think LLMs can be useful working tools, in principle, and their use should be regulated and checked rather than banned (which is why I'm wary of the proposed wording,Editors should not use an LLM to add content to Wikipedia).Gitz (talk) (contribs)14:40, 22 December 2025 (UTC)[reply]
For science I fed the entire source text into ChatGPT (gpt-5.2) with the promptFind and explain any errors in this Wikipedia article's code. That includes errors in the code as well as the prose. I don't think it's saving you time, but judge for yourself.~ Argenti Aertheri(Chat?)17:37, 22 December 2025 (UTC)[reply]
Raw LLM output
Below is a systematic review of errors and issues in the article, divided into (A) technical / wikitext problems, (B) sourcing and citation problems, (C) prose and MOS (Manual of Style) issues, and (D) content / historiographical issues. I distinguish hard errors (things Wikipedia would flag or break) from soft issues (policy-relevant but not syntactically wrong).A. Wikitext and template errors (hard, fixable)1.{{italic title}} is almost certainly incorrectProblem: The article title appears to be idéologues (plural noun, not a work title).Policy: {{italic title}} is reserved for works (books, films, journals), not topics or groups.Fix: Remove {{italic title}} unless the page title itself is a work.2. Inconsistent dash markup (- vs– vs{{endash}})You mix:hyphen-minus (-)en-dash (–)template {{endash}}MOS:DASH requires consistency.Fix: Prefer literal en dashes (–) in prose; reserve {{endash}} only if needed for template safety.3.{{lang|fr|Bureau de statistique|italic=unset}} misuseProblem: italic=unset is not a valid parameter for {{lang}}.Result: Parameter is ignored silently.Fix: Either:{{lang|fr|Bureau de statistique}}or wrap in {{nowrap}} / italics explicitly if needed.4. Broken / malformed cite web template{{cite web|last=Kennedy|first=Emmet|title=The secularism of Destutt de Tracy's 'Ideology'|url=https://www.geisteswissenschaften.fu-berlin.de/v/grammaire_generale/Actes_du_colloque/Textes/Kennedy/Emmet_Kennedy.pdf|website=Grammaire générale et idéologie - Actes du colloque|publisher=Freie Universität Berlin|date=2001|access-date=21 December 2025}}Problem: Line break inside website= value.Effect: Breaks template parsing.Fix: Put the website on one line or remove it.5.==Bibliography== spacingWikipedia requires a space:== Bibliography ==Current version violates MOS:HEADINGS.6. Overuse of{{ill}} where English articles existExamples:{{ill|Sergio Moravia|it}}{{ill|Jacques-Louis Moreau de la Sarthe|fr}}Some of these do have English articles or redirects.Fix: Replace with standard [[Article]] where available.7. Image placement breaks paragraph flowExample:Among them, Volney, later served as director of Napoleon's {{lang|fr|Bureau de statistique}}.[[File:AnneCatherineHelvetius.jpg|thumb|left|...]]The group brought together individuals...Problem: Image interrupts a sentence block.Fix: Move images to section boundaries or after paragraph breaks.B. Citation and referencing problems8. Overloaded lead with citationsThe lead is citation-heavy (≈12 refs).MOS:LEAD: citations should be minimized unless controversial.Fix: Move detailed citations to body sections.9. Self-contradictory citation groupingExample:{{sfn|Kennedy|1979|pp=353-354}}{{sfn|Ricciardi|2015|p=175}}{{sfn|Roberts|2025}}These sources disagree conceptually (Marxist vs analytic).Problem: Bundling implies agreement.Fix: Separate claims or attribute viewpoints explicitly.10. Dictionary citations used for historical semanticsOxford / Cambridge / Merriam-Webster used to explain ideological evolution.Policy: Dictionaries are fine for current usage, not historical development.You do this correctly once — but later blur the distinction.Fix: Explicitly label dictionary usage as modern English usage only (you partly do, but inconsistently).C. Prose and MOS issues (important but not fatal)11. Overly academic registerSentences regularly exceed 40–50 words.Wikipedia prefers encyclopedic clarity, not monograph style.Example:“{{lang|fr|Idéologie}} was an attempt to provide a rational and empirical foundation for ethics, political economy and legislation through a comprehensive science of human nature…”Fix: Break into two sentences.12. Implicit evaluative language (NPOV risk)Examples:“indisputably successful legacy”“stronghold of the group”“decisive break”These require explicit attribution or softer phrasing.Fix: “has been described as…”, “historians characterize…”13. Ambiguous pronoun referenceExample:“Although Sieyès’ membership among the idéologues is disputed, he acted as the circle’s ‘elder patron’…”“He” could refer ambiguously if paragraph reordered.Fix: Repeat name once.14. Quotation densityNapoleon quotations cluster heavily in one section.Gives rhetorical imbalance.Fix: Summarize one quote and retain the strongest primary citation.D. Content / historiographical issues (subtle but serious)15. Anachronistic category tensionCategories include:French classical liberal economistsFrench materialistsYet the article repeatedly stresses:rejection of metaphysicsphysiological groundingpolitical moderationIssue: These labels are later historiographical constructs, not self-descriptions.Fix: Clarify in prose that these are retrospective classifications.16. Overstated continuity to MarxArticle suggests a clean conceptual lineage from Tracy → Marx.Most historians emphasize rupture, not inheritance.You acknowledge difference but still frame it as transformation.Fix: Stronger language of conceptual inversion, not evolution.17. Membership list contradicts earlier caveatsEarlier: “no formal membership”Later: definitive bullet listFix: Add qualifier in header: “Commonly identified figures include…”18. Lead omits internal disagreementsBody discusses:tensions over religionpsychology vs physiologyLead presents ideological unity.MOS:LEAD requires representative balance.Fix: One sentence acknowledging internal diversity.Summary of most important fixes (priority order)❌ Remove {{italic title}}❌ Fix malformed cite web and invalid {{lang}} parameter❌ Normalize dash usage⚠️ Reduce citation overload in lead⚠️ Attribute evaluative language⚠️ Clarify Marx relationship as rupture⚠️ Make membership list explicitly heuristicIf you want, I can:Provide a cleaned, policy-compliant lead rewriteMark up exact diff-style fixesRun a mock FAC (Featured Article Candidate) review on this draft
Thanks. Unfortunately, ChatGPT's review turned out to be almost entirely wrong: (1) {{italic title}} is valid for non-English terms (WP:ITALICTITLE, MOS:NONENGITALIC). (2) Dash usage is not inconsistent in the rendered text; the proposed change is purely cosmetic. (3) italic=unset is a valid {{lang}} parameter. (4) Line breaks inside {{cite web}} do not break template parsing. (5) ==Bibliography== causes no issue, so adding spaces is purely cosmetic. (6) {{ill}} is only used where English articles don't exist. (7) The image is between paragraphs, not inside one, so no MOS:IMAGE issue. (8) MOS:LEADCITE does not require to minimise citations. (9) All cited sources support the statement about Marx's critical use of "ideology". (10) Dictionaries are not used for "historical semantics" but for current usage. (11) "Overly academic register"? Questionable. The example given is unconvincing. (12) No NPOV issues: two phrases are direct quotations, anddecisive break is mainstream scholarship/uncontroversial. (13) No actual pronoun ambiguity. (14) Claim of excessive Napoleon quotations is questionable. I don't agree. (15) "Materialists" and "classical liberal economists" are standard, sourced categories for 18th/19th century philosophers. (16) The article does not suggests a clean conceptual lineage from Tracy to Marx. E.g. Marx used "ideology"in a critical sense quite different from that intended by Tracy. (17) Membership list already includes an explicit caveat. (18) Internal disagreements are not discussed in the body, so omission from the lead is not an error.Gitz (talk) (contribs)09:50, 23 December 2025 (UTC)[reply]
I think it's because editors are focused too much on the tool that is used (LLMs), and not enough on teaching the people using the tool how to use it correctly (and banning the ones that don't seem to have the competence to use it).Levivich (talk)19:26, 10 December 2025 (UTC)[reply]
The onus shouldn't be on us to teach them to use it correctly any more than we should have to teach people to use their keyboard or to make appropriate reversions with Twinkle. Competence is required, and if you decide to charge in like a bull in a china shop and make huge amounts of disruptive edits with a tool you don't know how to use, we shouldn't treat it any differently than we would treat people mass-misusing twinkle or AWB or what have you without doing their due diligence.Athanelar (talk)19:34, 10 December 2025 (UTC)[reply]
Surely you're kidding? We do, in fact, teach people how to type on Wikipedia. We have all sorts of instructional pages about how to write on Wikipedia, how to format, how to use Wikitext, etc. etc. But I agree with you that people should be held accountable for what they publish on the website irrespective of what tools they use to publish, whether it's Twinkle, AWB, just a keyboard, or a chatbot.Levivich (talk)19:39, 10 December 2025 (UTC)[reply]
Well, no, we should at best expect them to educate themselves on how to do that, in the same way that as soon as you click the box to enable Twinkle you're agreeing that you know what you're doing and will use it responsibly.
If someone then went on to make a bunch of problematic mass reversions using Twinkle while ignoring numerous warnings about it, we wouldn't sit them down and give them a crash-course on how to use Twinkle, we'd block them for disruptive editing.Athanelar (talk)19:50, 10 December 2025 (UTC)[reply]
If I see a user systemically not following a guideline, I'll tell them about the guideline on their talk page; no harm done. If theyignore my message and continue egregious violations of the guideline, then a short block would be justified, not to "punish" them, but to get their attention and stop the immediate problem so we can, metaphorically, sit down and explain what's wrong. I often see new editors rush ahead and make huge batches of changes that are suboptimal: often they'll listen, but sometimes you need to grab them by the shoulder and tell them what's up.Cremastra (talk·contribs)20:01, 10 December 2025 (UTC)[reply]
a short block would be justified, not to "punish" them, but to get their attention and stop the immediate problem so we can, metaphorically, sit down and explain what's wrong Then why are editors gettingindeffed for LLM misuse right after being brought to ANI? I know what you'll say, "indefinite is not infinite", but indefinite is effectively infinite for most newcomers, as most newbies that get indeffed will likely give up and quit Wikipedia entirely (and the ones that don't often trymaking a new account instead).SuperPianoMan9167 (talk)20:11, 10 December 2025 (UTC)[reply]
Then why are editors getting indeffed for LLM misuse right after being brought to ANI? they aren't. They are being indeffed for breaching policies like disruptive editing in exactly the same way they would be if their disruptive editing was unrelated to LLMs.Thryduulf (talk)20:27, 10 December 2025 (UTC)[reply]
Of their examples of 'incivility' much of them were quite evidently comments about content and not contributor. Their 'creepiest of all' example seems to be from a misunderstanding of what speedy deletion means, as they seem to think Tankishguy meant a speedy deletion ofthem as a user rather than of the pageUser:Orange Jones. They even quoted me sayingCompetence is required… If you aren't capable of being clear, respectful, and organised without AI help, then you simply can't participate here. as an evidence of incivility -- which, obviously I'm biased, but seems to me like a pretty straightforward summation of CIR.
And I'll say now what I said previously to another problematic AI contributor; civility is mandatory, but respect is a two-way street. If you're going to come into Wikipedia and make no effort to engage meaningfully with the project, spitting out AI-generated unblock request after unblock request and straight up saying "AI is too useful to me, I can't promise I won't use it again" after people have stressed again and again how your AI usage is problematic, you can't really be surprised if editors want to wash their hands of you and get a little terse in their interactions with you.
I won't shed any tears if the editors we lose by being harsher on AI are the kind of editors who persist in misconduct after it's been repeatedly and patiently pointed out to them, and then cry foul when people lose their patience.Athanelar (talk)20:48, 10 December 2025 (UTC)[reply]
What if the editors we lose by being harsher on AI are also the kind of editors who are doing their best to bring a global perspective to articles that deserve a{{globalize}} tag? People who don't write easily in English can nonetheless provide useful, encyclopedic contributions. I don't think we need a "ban AI" rule to get rid of people who persist in misconduct.WhatamIdoing (talk)23:18, 11 December 2025 (UTC)[reply]
We're an English-language encyclopedia. I don't think sacrificing our English competency requirements is worth it for a potential increase in global perspective.Athanelar (talk)00:46, 12 December 2025 (UTC)[reply]
Blockingfeels punitive to most people, even when the motivation is purely preventive.
Also, inoperant conditioning terms, if a short block improves an editor's behavior, then it actually is a punishment. (Punishments are stimuli that decrease the likelihood of a behavior recurring. Even something so mild as suggesting a better alternative can be "a punishment" in this sense.)WhatamIdoing (talk)19:27, 11 December 2025 (UTC)[reply]
Well, yeah. Just like a newcomer who has zero experience with Wikipedia policies and guidelines is expected to educate themselves on notability, verifiability etc before they run off and start submitting drafts at AfC or mainspace. If they fuck up once or twice then we give them a 'heads up' and explain why they shouldn't do what they're doing, but if they keep doing after that then they get a block until they can demonstrate that they know what they're doing is wrong and why they shouldn't be doing it.
In fact I find the approach that tells us new editors need lots of hand-holding and infinite warningspatronizing rather than helpful, because it assumes that new editors are all stupid. Most aren't, and can figure things out given a bit of encouragement.Cremastra (talk·contribs)20:09, 10 December 2025 (UTC)[reply]
That, andcompetence is required; somebody who needs a million reminders about appropriate conduct and doesn't have the presence of mind to do a little bit of independent research before taking action isn't likely to be a productive editor anyway.Athanelar (talk)20:17, 10 December 2025 (UTC)[reply]
Just like a newcomer who has zero experience with Wikipedia policies and guidelines is expected to educate themselves on notability, verifiability etc before they run off and start submitting drafts at AfC or mainspace.
The first link there is about not assuming bad faith just because somebody is ignorant.
The second, similarly, is to say you're not exempt from being accused of BITEing just because the newcomer is ignorant.
The third only reinforces my point with its sectionWP:CAREFUL; we encourage boldness, not recklessness. You're expected to do at least some forethought about what you're doing, whether you're doing it right, and whether you should be doing it at all. If someone decided a piece of wording inWP:N was unclear and substantively changed its meaning of their own volition 'to be bold' without any attempt at consensus-gaining, they'd be in for a serious trouting at the least, even if they were a newcomer ignorant of those concepts.
Well, I think it's moreso that nobody has bothered to report them. I thought the same thing when I saw the AINB case aboutUser:AKLKPhilo and when I decided to escalate that to ANI they were indeffed five minutes after I submitted my report.Athanelar (talk)19:32, 10 December 2025 (UTC)[reply]
Well, that was because they lied about verifying the output and then continued to add hallucinated sources, and also because they violated COI guidelines by recreating their COI page into mainspace after having a draft declined.
AI-generated articles are alreadyper se forbidden, byWP:NEWLLM. This proposed revision would forbid much more than AI-generated articles, it would forbid using AI altogether, for anything. (I know, you say it only says "should not" not "must not" but we've already covered our disagreement on that point.)Levivich (talk)19:45, 10 December 2025 (UTC)[reply]
Well, yes, and I'm very transparent about being in favour of an eventual moratorium on all AI-generated additions to Wikipedia, even if I don't think this proposed guidelineis that necessarily. Still, AKLKPhilo's block isn't really a good example that such a thing isn't necessary - because their block was based on other substantive policy/guideline breaches anyway.Athanelar (talk)19:48, 10 December 2025 (UTC)[reply]
Yeah, exactly. I've seen the sentiment floating around that there's some kind of de facto prohibition of AI usage being enforced, and while I wish that were the case I don't think there is. The whole point of guidelines like this one is to start enforcing a stricter standard of usage for LLMs.Athanelar (talk)20:08, 10 December 2025 (UTC)[reply]
The whole point of guidelines like this one is to start enforcing a stricter standard of usage for LLMs.
I think that's just a weak argument, or at least selectively applied. We have no problem understanding thatWP:Notability is both simultaneously a guideline and also an essentially incontrovertible standard that an article must follow in order to be included, and I think we therefore must also be able to understand that any guideline-based prohibition against LLM usage is both a prohibition and also subject to common sense exceptions, interpretive wiggle room, etc.Athanelar (talk)20:23, 10 December 2025 (UTC)[reply]
What I'm saying is that the guideline shouldexplain itself. I'm not really against expanding restrictions on LLM usage; I would actually support a prohibition onunreviewed or poorly reviewed LLM-generated additions to articles. This is because, as you said, LLMs areessentially useless at producing any text in wikivoice. If there are exceptions to this,WP:IAR covers them.SuperPianoMan9167 (talk)22:59, 10 December 2025 (UTC)[reply]
There's too much meta-discussion creeping into the survey section such as the bickering about bludgeoning. Editors such asAthanelar andThryduulf should please try to keep their complaints and counter-arguments in this discussion section.Andrew🐉(talk)17:17, 18 December 2025 (UTC)[reply]
Comment Regarding the contradiction between the "thou shalt not" expressed at the beginning and the "thou mayest" concessions that follow, along with the rationale justifying the confusing, contradictory stances: Can we accomplish the same thing while avoiding the contradiction by taking the same tack as we do for COI? "... isstrongly discouraged and subject to all the constraints that follow."Largoplazo (talk)18:38, 18 December 2025 (UTC)[reply]
It may help to have some examples of how this is being treated in forums such as ANI. I'll start with one that I was involved in. That discussion doesn't seem to dwell on the fine points; there's just some general anti-AI sentiment.Andrew🐉(talk)08:31, 7 January 2026 (UTC)[reply]
The discussion above is closed.Please do not modify it. Subsequent comments should be made on the appropriate discussion page. No further edits should be made to this discussion.
Moved fromUser talk:Qcne/LLMGuideline