The followingis aproposed Wikipedia guideline against the use of LLMs in user-to-user commmunications. The proposal may still be in development, underdiscussion, or in the process of gatheringconsensus for adoption.

This page in a nutshell: Other editors want to talk toyou, not a chatbot speaking on your behalf.

Thisuser page or sectionis undergoing significant expansion or restructuring. You are welcome to assist in its construction by editing it as well. If thisuser pagehas not been edited in several days, please remove this template. If you are actively editing this article or section, you can replace this template with{{in use|5 minutes}}.This page waslast edited byAthanelar(talk |contribs) 33 days ago.(Update timer)

Wikipedia is a collaborative project; its success depends on the ability of its participants to communicate with one another. On a daily basis there are conduct and content disputes alike which require effective,civil communication in order to come to consensus, resolve disagreements, and find solutions. Large language models more often than not serve as a barrier to this process rather than an assistant of it,therefore using them to generate user-to-user communication is forbidden.

The substance of this guideline is a prohibition against outsourcing one's thought process to a large language model. The goal is to ensure that all user-to-user communication is based on thoughts and ideas actually presented by the human editors on both ends of said communication, rather than one or both ends actually being a chatbot communicating through a human proxy. An editor is expected to bring their own ideas to a discussion, not merely to repeat the ideas generated for them by an LLM. The following guidance should be taken with this goal in mind.

Editorsare not permitted to use large language models to generate user-to-user communications, including but not limited to talk page comments, noticeboard complaints, and comments or nominations in deletion discussions.

This prohibitionincludes the use of LLM-generated text which is then reviewed, reworded or otherwise modified by the human editor, but where the fundamental idea or argument is ultimately still from the LLM output.

LLM-generated communications in talk pages, deletion discussions,requests for comment and the like may becollapsed and excluded from assessments of consensus. LLM-generated noticeboard complaints may be procedurally closed and the complaining editor instructed to resubmit the complaint in their own words. Unblock requests formulated with an LLM may be procedurally declined by any reviewing admin without further justification. Repeated violations of this guidance are considereddisruptive editing and may result in sanctions.

LLMs can be usefulassistive technology for those with certain disabilities or limitations that affect their ability to contribute to Wikipedia, such asdyslexia or limited English proficiency. This guidelinedoes not aim to restrict the use of LLMs in this capacity, provided that this usage does not violate the prohibitions above and that the caution provided below is given due consideration.

Particular caution must be taken if using large language models to translate, copyedit, fix tone, correct punctuation, create markup, or in any way cosmetically adjust or refactor human-written text for user-to-user communication. LLMs cannot be trusted to stick to the scope of this task, as outlined in the section below. The use of LLMs in this manner may result in human-written text being overly 'corrected' (ie. largely rewritten) in a way which substantively changes the meaning of the communication from what was intended by the human editor, which is obviously problematic if the intent was to clarify one's meaning or present one's ideas in a better way. Thisincludes tools such as Microsoft Word (which includesMicrosoft Copilot LLM functionality as of 2025) andGrammarly (which is now 'Grammarly AI', and also includes LLM functionality).

Editors who choose to do so despite this caution are expected to thoroughly review the copyedited/translated/formatted/refactored text for any unintentional or excessive changes.

It isalways preferable to entirely avoid the use of LLMs and instead make the best effort you can on your own. As explained in a section below, it is almost certain that any task an LLM can do for you here on Wikipedia is a task you could do better when you learn how, and other editors would typically much rather engage with a well-intentioned but imperfect (in terms of grammar, structure etc) comment/discussion rather than the same AI-generated boilerplate they have seen countless times before.

A type of prompt which isnot permitted under this guideline would be to prompt the LLM something likeI have been blocked on Wikipedia for disruptive editing. Write me an unblock request citing all relevant policies and guidelines. and directly copy the output. The reasoning behind the unblock request in this example would be entirely generated by the LLM, which does not demonstrate the blocked user's understanding of their misconduct or how they will avoid it in future.

A type of prompt whichis permitted would beI have been blocked on Wikipedia for disruptive editing. Write an unblock request which explains that I understand that edit warring and insulting other editors was disruptive, and that in the future I plan to avoid editing disputes which frustrate me in that way to prevent a repeat of my conduct, and that I am willing to accept a voluntary 1RR restriction if it will help with my unblock. Cite all relevant policies and guidelines This outsources thewording to the LLM, but not the substance. However, the output of this prompt still should not be copy-pasted wholesale, and should be thoroughly reviewed to ensure it actually reflects the sentiment the editor is trying to communicate. In particular, any cited policies and guidelines should be thoroughly read through to ensure the LLM does not misquote or misapply them, which is very common.

These are examples of a prompt that would result in an obviously unacceptable output and a prompt that would result in a likely acceptable one, to act as guidance for editors who might use LLMs. They should not be taken as a standard to measure against, nor is the prompt given necessarily always going to correlate with the acceptability of the output. Whether or not the output falls afoul of this guideline depends entirely on whether it demonstrates that it reflects actual thought and effort on the part of the editor and is not simply boilerplate.

Large language models are specialised for a single task; generating human-like language output. They do this throughmachine learning; they take in a large amount of human-written training data, and based on that they develop an idea of how a human would respond to any given prompt. For instance, the question "What is two plus two?" is much more likely to be responded to by a human with the answer "Four." than with the answer "Zvsugwnfonsyabajnsbgueub," and so the language model is able to determine that the former is an example of realistic human language and the latter is not.

Based on the above, a person who asks an LLM "what is two plus two?" and receives the answer "four" may be inclined to believe that the LLM is able to perform mathematical operations. Strictly speaking, it is not; the LLM may be able to give you a correct answer to a mathematical question, but that is only because its training data contains sufficient samples of humans giving those kinds of answers, and so it is able to determine how a human would respond to that question.

Similarly, in an editorial dispute over conduct or content, a large language model is completely unable to determine what a sensible or ideal outcome would be. There is no means by which an LLM can determine that Outcome X can logically follow Circumstance A modified by Circumstance B. It can only determine that sentence/paragraph B (its output) is what a human would likely respond to sentence/paragraph A (its input prompt) based on its training data. If you're trying to argue thatExampleBadArticle should be deleted, or thatExampleGoodArticle should be kept, or thatUser:ExampleBadUser should be sanctioned for misconduct, an LLM is not able to formulate an actually logical argument in favour of that outcome, it is only able to generate a human-sounding argument, which is much different.

Due to their lack of ability to reason logically, LLMs cannot read, interpret, and apply Wikipedia policies and guidelines for you.

First and foremost, they have a habit of hallucinating policies and guidelines that don't even exist. Their training data shows that in arguments on Wikipedia users have a habit of saying things like "this source isn't verifiable, which is bad because ofWP:VERIFIABILITY" but they have no way of knowing which policies and guidelines actually exist except for those quoted in their training data, so in the quest to generate human-sounding output they will come up with arguments to the effect of "This article shouldn't be deleted, perWP:DONTDELETE"

Similarly, they may quote policies/guidelines thatdo exist, but wildly misapply them because the name of the policy (or at least the policy's shortcut) tangentially applies to the situation. One editor in their AI-generated unblock request argued that their block was invalid becauseWP:COMMUNICATE states that an admin must warn a user before blocking them. That essay of course says no such thing.

While there is a lot of Wikipedia text (including these kinds of debates) in the training data for these LLMs, they are not specifically trained in generating convincing-sounding arguments based on Wikipedia policies and guidelines, and considering they have no way to actually read and interpret them, these kinds of failures are inevitable.

While copyediting and reformatting is often cited as an inoffensive use of LLMs, it is not without its own dangers. Inthis diff for example, a rewording by a copyediting LLM completely changes the meaning of a statement, with serious resulting implications;

The original makes the plain claim that the individual never supported Trump, while the 'copyedit' completely changes the claim to imply that Silvers supported Trump at one time, but stopped.

Inthis diff a careless use of LLM copyediting results in a thesaurus-like rewording of a direct quotation.

Inanother case, an editor blocked for chronic LLM misuse and repeated additions of hallucinated sources explained that they had asked an LLM to format an infobox, and in the process it had hallucinated an ISBN number without them noticing. Whether we should take this explanation at face value is another matter; if we do, then it means that LLMs cannot be trusted to copyedit text and create formatting without making other, more problematic changes, as has also been shown in the other examples above.

If you're planning to use an LLM to communicate for you, it may be either because you're nervous about doing it wrong, or because you feel you aren't capable of doing it yourself. Wikipedia has a steep learning curve and it is normal to feel overwhelmed, especially by complicated tasks like article creation. However, with a little effort, you can very quickly reach a level of understanding necessary to do anything you want to do here—and as outlined above, you'll almost certainly be able to do it better than an LLM can.

Nobody is born knowing how to edit Wikipedia, to formulate an ANI complaint or to respond in a talk page discussion. These are things we learn to do by reading applicable policies, guidelines, and essays, by copying others, and moreover by simplydoing it. Wikipedia encourages us tobe bold and attempt things. If we're unclear in a talk page then other editors can ask us for clarification. If we malform an ANI complaint somebody will correct us. There's no need to rely on an LLM to try to get things perfect on the first try, and other editors areobligated to be patient and kind to those that are learning.

WP:Competence is required states that an editor must havethe ability to communicate with other editors and abide by consensus.

It is increasingly normal in society to rely on AI to accomplish tasks a person cannot or is unwilling to do themselves; creating visual art, writing poetry, completing a homework assignment, etc. But just like how completing a homework assignment with ChatGPT doesn't mean you actually understood it, using an LLM to communicate on your behalf on Wikipedia fails to demonstrate that you, the human being behind the screen, have the required competence to communicate with other editors.

People who are unwilling or unable to communicate with other editors, interpret and apply policies and guidelines, understand and act upon feedback given to them etc., and require an LLM to do so on their behalf, areincompatible with a collaborative, consensus-based project such as Wikipedia.

This is the English Wikipedia, and participation here both in article-space and elsewhereshould normally be in English. Consequently, asWP:CIR puts it, editors must havethe ability to read and write English well enough to avoid introducing incomprehensible text into articles and to communicate effectively.

If your English isn't perfect, that is completely fine. So long as you can communicate effectively, then nothing else matters. There's no need to use an LLM to try to speak in perfect English, because in the process it strips your voice away from your words, making the communication impersonal and failing to properly represent your intentions. Other editors would much rather engage with a well-intentioned comment written in less-than-perfect English than read the same boilerplate AI-generatedslop we've seen a hundred times.

On the other hand, if your English is entirely insufficient to communicate effectively, then in any case you may prefer to participate onthe relevant Wikipedia for your preferred language. If there is a pressing issue with an article on the English Wikipedia that you wish to comment on, you could always use thelocal embassy system to find a Wikipedian who speaks your language and can present your comment in English.

• Wikipedia:Large language models - general guidance on LLMs in the context of Wikipedia
• Wikipedia:Writing articles with large language models - Guidance against using LLMs to generate new articles

• Wikipedia proposals