Given that at this point the prohibition against using LLMs in user-to-user communication WP:LLMCOMM has become something of a norm, I think it would be sensible to make it an official guideline as part of the ongoing attempt to strengthen our LLM policy.

Rather than just promote the exact text of LLMCOMM, I've decided to try to create something which synthesises LLMCOMM, HATGPT and general advice about LLMs in user-to-user communication. My proposal as it currently stands is at User:Athanelar/Don't use LLMs to talk for you

My proposed guideline would forbid editors from using LLMs to generate or modify any text to be used in user-to-user communications. Please take a look at it and let me know if there's anything that should be added or modified, and if you agree with the proposed restrictions. I'd love to workshop this a bit and get it to a stage where it can be RfCed. Athanelar (talk) 13:33, 7 January 2026 (UTC)[reply]

My thoughts:

• Your proposal is much more strict than LLMCOMM (which is already enshrined in guidelines as WP:AITALK). It doesn't just synthesize LLMCOMM and HATGPT, which both allow exceptions for refining one's ideas; it goes beyond that and bans LLM use entirely for writing comments. This, combined with the Editors should not use an LLM to add content to Wikipedia phrasing of the proposed NEWLLM expansion, would effectively ban all use of LLMs anywhere on the English Wikipedia. This makes sense given your stated opinions on LLM policy, but I'm sure this is going to get significant opposition.
• Your proposal also goes beyond commenting to basically say that LLMs are useless for any Wikipedia editing at all, as indicated by the section about copyediting. This seems out of place for a guideline that is supposed to be about using LLMs for comments. Again, this makes sense given your stated anti-LLM sentiments, but I have seen it repeatedly demonstrated that such a sentiment is far from universal.
• I am concerned mainly because this guideline assumes bad faith from LLM-using editors. Most LLM-using editors are unaware of their limitations because of the massive hype surrounding them. My opinion is that instead of setting down harsh sanctions for LLM use, we should instead educate new users on why LLMs are bad and teach them to contribute to Wikipedia without them.
• Finally, a lot of editors are just worn out at this point from having so many LLM policy discussions in such a short period of time. Can we at least wait until the NEWLLM expansion proposal is over? SuperPianoMan9167 (talk) 14:23, 7 January 2026 (UTC)[reply]

I appreciate your feedback and your continued presence as a moderate force in these discussions.

1. I recognise my proposal is quite extreme. My goal was to shoot for 'best case' and compromise from there as necessary.
2. The subsection on copyediting exists to justify the restriction against using LLMs to refactor, modify, fix punctuation etc; because at best the LLMs are unfit for this task anyway, and at worst it provides a get out of jail free card for bad faith editors. The overall section is in fact expressly intended to demonstrate that LLMs simply are not any good at doing the things people might want them to do in discussions.
3. I have tried to avoid that by pointing out that I believe the motivation to use LLMs in these cases comes from a good place (concerns about one's abilities)
4. I understand; but I still have the passion and energy, and I hope others do too. We are in something of a race against the clock here; every month we wait before strengthening our policies is another month of steadily being invaded by this type of content.

Athanelar (talk) 14:43, 7 January 2026 (UTC)[reply]

Some quick notes:

• Missing the biggest reason not to use LLMs for your comments: it will make people more likely to dismiss your comments, not less.
• As usual I think we need to specifically name what tools we are talking about. People genuinely don't know things are AI that actually are, and if we can't convince them of that, we can at least say "don't use ____" in the guideline.
• 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 - Technically you could provide policies and guidelines in a prompt. Most people probably aren't doing that, but they could.
• There are probably better copyedit examples; the first one seems like splitting hairs, and the original sentence had the same problem with different punctuation. The one where an AI copyedit turned "did not support Donald Trump" to "withdrew her support for Donald Trump" comes to mind. Better yet would be a copyedit to a talk page comment, though that might be hard to come by without using AI yourself.

Gnomingstuff (talk) 15:07, 7 January 2026 (UTC)[reply]

Much appreciated, I'll replace the example in the copyedit section later and add the info from your first two bullets. Athanelar (talk) 17:53, 7 January 2026 (UTC)[reply]

Might be worth saying that people are much more likely to engage with a comment in broken english than an llm-generated one Kowal2701 (talk) 21:16, 7 January 2026 (UTC)[reply]

I agree with this. SuperPianoMan9167 (talk) 21:19, 7 January 2026 (UTC)[reply]

Is it true? I've seen plenty of broken English comments reverted without comment, or the response has been unhelpful. WhatamIdoing (talk) 03:20, 9 January 2026 (UTC)[reply]

"Editors are not permitted to use large language models to generate or modify any text for user-to-user communication" will have a disparate impact that discriminates against people with some kinds of disabilities, such as dyslexia. A blanket ban is therefore in conflict with WP:ACCESS and possibly with foundation:Wikimedia Foundation Universal Code of Conduct.

I think it is patronizing to tell people "You don't need it" when some of them actually do. I oppose telling English language learners to simply go away ("If your English is insufficient to communicate effectively, then once again, you unfortunately lack the required language ability to participate on the English Wikipedia, and you should instead participate on the relevant Wikipedia for your preferred language"), because (a) that's rude, and (b) sometimes we need them to bring information to us. If you don't speak English, but you are aware of a serious problem in an English Wikipedia article, I want you to use all reasonable methods to alert us to the problem.

Here's the Y goal:

• A: I don't know English very well, but the name on the picture in this article is wrong.
• B: Thanks for letting us know about this factual error. I'll fix it.

Here's the N anti-goal:

• A: I don't know English very well, but the name on the picture in this article is wrong.
• B: This is obvious AI slop. If you can't write in English without using a chatbot to translate, then just go away and correct the errors at the Wikipedia for your native language instead!
• A: But the error is at the English Wikipedia.
• B: I don't have to read your obvious machine-generated post!

Unfortunately, that's not a made-up example. WhatamIdoing (talk) 19:06, 7 January 2026 (UTC)[reply]

@WhatamIdoing, spelling out this scenario has helped me think through some of what I'm seeing in this discussion. I think that a weak point in LLMCOMM, CIR, and similar guidelines is that there are really at least three different broad categories of "editors" who have different needs and interests:

• People who genuinely want to help build an encyclopaedia and may be in this for the long term ("Wikipedians") – most of our policies and guidelines are written with these editors in mind
• People who have identified serious problems in specific articles (regardless of whether they're article subjects or have a COI, or are uninvolved) – if there are serious problems that need to be fixed, we need to fix them, and we should be thanking these helpful non-Wikipedians, not putting up barriers based on CIR or LLMCOMM
• People who are here for self-promotion, not to build an encyclopaedia – we have rules and procedures for dealing with these

Trying to address these different types of editors/visitors with one broad brush does a disservice to the well-intentioned non-Wikipedians, as WAID has illustrated. ClaudineChionh (she/her · talk · email · global) 21:22, 7 January 2026 (UTC)[reply]

Amazingly I don't think this is said anywhere in LLM PAGs or essays, but we should say somewhere that "Wikipedia does have a steep learning curve and it is very normal for a new editor to struggle. Some learn quicker than others, and people are obligated to be patient with new editors and help them improve." Basically, don't worry if you find it hard. I'd rather something like that replaced "You don't need it" Kowal2701 (talk) 21:23, 7 January 2026 (UTC)[reply]

"People are obligated to be patient with new editors and help them improve" = Wikipedia:Please do not bite the newcomers (a behavioral guideline). WhatamIdoing (talk) 03:27, 9 January 2026 (UTC)[reply]

Note that I have slightly rewritten the "You don't need it" section to focus a bit more on the encouragement, and also to soften the language around English proficiency. @WhatamIdoing @SuperPianoMan9167 et al, is this something more in line with your ideal spirit? Athanelar (talk) 21:34, 7 January 2026 (UTC)[reply]

Yes! I'm still somewhat opposed to the general premise, banning all use of LLMs in comments, but that section is much better now.

My ideal version of such a guideline would be:

• Generating comments with LLMs (outsourcing your thinking to a chatbot) is prohibited. You have to be able to come up with your own ideas.
• Modifying comments with LLMs, such as using them for formatting, is strongly discouraged. This is due to the risk of the LLM going beyond changing formatting and fundamentally changing the meaning of the comments.

SuperPianoMan9167 (talk) 21:43, 7 January 2026 (UTC)[reply]

I think this is a compromise I would absolutely be on board with, if others agree. Athanelar (talk) 21:46, 7 January 2026 (UTC)[reply]

I like this a lot. Gnomingstuff (talk) 22:21, 7 January 2026 (UTC)[reply]

I do also like this. Many editors say that they used LLMs "only for grammar" while having the kind of issues that only comes with LLM generation (for example, the same vague, nonspecific boilerplate reassurances that can be found almost word-for-word in at least half of the unblock requests I've seen), and others might genuinely not realize that the LLM has completely changed the meaning of their comment behind a facade of "grammar fixes". Chaotic Enby (talk · contribs) 23:04, 7 January 2026 (UTC)[reply]

Per the feedback given, I have changed the scope of the proposal. The proposal now:

• Forbids the use of LLMs to generate user-to-user communication, including to generate a starter or idea that a human then edits. (this clause is added to close the inevitable loophole that would arise from that)
• Strongly discourages the use of LLMs to review or edit human-written user-to-user communication, and explains that if doing so results in text which appears wholly LLM-generated, then it may be subject to the same remedies as for LLM-generated text

So, LLM-written and LLM-written, human-reviewed communications; not allowed.

Human-written, LLM-reviewed communications; strongly discouraged

Please re-review accordingly and let me know if there's anything else that needs changing or adding before we move to RfC. Athanelar (talk) 00:15, 8 January 2026 (UTC)[reply]

The sentence about people unwilling or unable to communicate/interpret/understand feedback etc. should be reworded to the following: People unable to communicate with other editors, interpret and apply policies and guidelines, understand and act upon feedback given to them etc. should ask for help at the teahouse. If you keep the current, the word incompatible should not be linked as the linked page is something on categories and redirects, not related to the linking sentence. In any case, I support the proposal. Nononsense101 (talk) 02:39, 8 January 2026 (UTC)[reply]

• Proposal: "LLM-generated communications in talk pages, deletion discussions, requests for comment and the like should be collapsed and excluded from assessments of consensus".
• Scenario:
  • A: Hi, I found a serious factual error in the article. Here's a source to show I'm not making this up.
  • B: Collapsed! GPTZero said this is AI generated. Come back when you can right in your own words.
  • A: But I didn't generate this idea; I only used it for copyediting.
  • B: Collapsed again! The rules say I should collapse your AI-generated nonsense, so I'm following the rules!

This is looking like the N anti-goal to me. WhatamIdoing (talk) 00:26, 9 January 2026 (UTC)[reply]

Nobody is arguing that we should treat text as AI generated just because GPTZero says so; this is a strawman. I even have another proposal specifically to address the identification of AI generated text, but that's for another time. Athanelar (talk) 00:39, 9 January 2026 (UTC)[reply]

Nobody (here) is arguing that we should trust GPTZero, and I suspect that everybody here has seen editors actually do that, and believe they are completely justified in doing that. WhatamIdoing (talk) 03:30, 9 January 2026 (UTC)[reply]

Sure, but if someone quoted my hypothetical guideline to justify collapsing an evidently good-faith, human-written edit request just because GPTZero said it's AI generated, I think any sensible editor seeing that would say it's not a reasonable application of the guideline.

You can't argue against a guideline by taking the worst possible way a person could misinterpret it. It constantly happens that editors accuse other editors of personal attacks because they get told their contribution was bad, does that mean WP:NPA isn't fit for purpose? Athanelar (talk) 03:48, 9 January 2026 (UTC)[reply]

For many editors, "GPTZero said it's AI generated" proves that it's not a "human-written edit request". If you don't want that to happen per your proposal, then you need to increase its already bloated (~1800 words) size even more, to tell editors not to believe GPTZero. WP:NPA might be a viable model for this, as it explains both what is and isn't a personal attack, and how to respond to differing scenarios.

I can, and have, since before some our editors were even born, argued against potentially harmful rules by taking the worst possibles way a person could misinterpret it, and then deciding whether that worst-case wikilawyer is both tolerable and likely. Thinking about how your wording might be misunderstood or twisted out of recognition is how you're supposed to write rules.

This has been known since at least the 18th century, when James Madison wrote in Federalist No. 10 that "It is in vain to say, that enlightened statesmen will be able to adjust these clashing interests, and render them all subservient to the public good. Enlightened statesmen will not always be at the helm: Nor, in many cases, can such an adjustment be made at all, without taking into view indirect and remote considerations, which will rarely prevail over the immediate interest which one party may find in disregarding the rights of another, or the good of the whole", and went on to propose a large federal republic as a way of keeping individual liberty (which is a necessary precondition for factionalism) and national diversity (which leads to factionalism through an us-versus-them mechanism) while reducing the opportunity for any one faction to seize power over the others.

I recommend Madison's work on factionalism is to anyone who wants a career in policy writing, but for now, spend a few minutes thinking about how we could adapt Madison's definition of a faction: "a number of Wikipedians...who are united and actuated by some common impulse of passion against AI...adverse to the rights of other Wikipedians (e.g., to have others focus on the content, instead of focusing on the tools used to write it), or to the permanent and aggregate interests of the community (e.g., to not WP:BITE newcomers or have hundreds of good-faith contributors told they're not welcome and not WP:COMPETENT)."

In the present century, we call this phenomenon things like misaligned incentives (e.g., editors would rather reject comments on a technicality than go to the trouble of correcting errors in articles or explaining why it isn't actually an error, but articles need to be corrected, and explanations help real humans), and we address it through processes like designing for evil (e.g., don't write "rules" that can be easily quoted out of context; don't optimize processes for dismissive or insulting responses) and use cases (e.g., How will this rule affect a person who doesn't speak English well? A WP:UPE? A person with dyslexia? An autistic person? A one-off or short-term editor?).

For example:

• Protect the English language learner by declaring AI-based machine translation to be acceptable.
• Ignore the UPE's AI use as small potatoes and block them for bigger problems.
• Educate anti-AI editors that both human- and AI-based detectors make mistakes, and that these mistakes are more likely to result in editors unintentionally discriminating against editors with communication disabilities.
• Remind editors to WP:Focus on content, which sometimes means saying "Thanks for reporting the error" instead of collapsing AI-generated comments.

These are only examples; the more important part is to think through each bit yourself. WhatamIdoing (talk) 05:10, 9 January 2026 (UTC)[reply]

I do understand your point, and am truly appreciative of the time and effort you're taking to make it. I still have two concerns with it;

1. The first is bloat; as you've indicated, words are precious in any policymaking effort and the longer people have to read to 'get to the point' the less chance they will. I'm concerned at how much weight should be added to cover things like "it's also possible to make mistakes without AI" that in any case should be assumed by any reasonable audience. It also feels redundant, i.e., AGF and BITE still apply even if I don't explicitly restate them. The existence of a guideline prohibiting AI-generated text is by no means a carte blanche to ignore those other, more fundamental principles.
2. Given that your primary cause for concern seems to be about collapsing AI-generated comments; well, that already exists as WP:HATGPT, all I'm doing is restating it here. However, on rereading that, I suppose I could (and will) add some language specifying that conversations should not be collapsed if their content proves otherwise extraordinarily useful, which should cover the edge cases you're concerned about, with super-useful AI users and overly anal-retentive wikilawyers.

Athanelar (talk) 10:13, 9 January 2026 (UTC)[reply]

@Athanelar, when I'm working on policy-type pages, the definitions in RFC 2119 are never far from my mind. Here are the most relevant bits:

• SHOULD This word, or the adjective "RECOMMENDED", mean that there may exist valid reasons in particular circumstances to ignore a particular item, but the full implications must be understood and carefully weighed before choosing a different course.
• MAY This word, or the adjective "OPTIONAL", mean that an item is truly optional.

And now let's compare what you wrote vs the text at HATGPT:

• LLM-generated communications in talk pages, deletion discussions, requests for comment and the like should be collapsed and excluded from assessments of consensus.
• Comments, nominations, and opening statements that are obviously generated (not merely refined) by a large language model or similar AI technology may be struck or collapsed...

In short, you've said that this "SHOULD" normally happen unless someone has carefully considered the situation and decided to make a special exception (e.g., for "extraordinarily useful" comments), and the existing guideline says that this "MAY" happen, but it's strictly optional and not ever required. Can you see the gap between what you're proposing and the existing guideline? If you genuinely believe that well, that already exists as WP:HATGPT, all I'm doing is restating it here, then I don't think we're speaking the same language. WhatamIdoing (talk) 01:31, 12 January 2026 (UTC)[reply]

Sure, nevertheless, the kind of person you're describing would do what you're saying regardless of whether it's 'should' or 'may', and my entire contention is as to whether it's realistic enough of a concern to affect anything, which I simply doubt that it is. People who fail to AGF won't get a free pass just because 'Athanelar's guideline says 'should' so that means I can collapse whatever I want"

I think the real solution to your concern is to establish a consensus standard for determining what is "AI-generated text" for the purpose of these types of guidelines, which I already tried before Athanelar (talk) 10:29, 12 January 2026 (UTC)[reply]

I do think that if someone reads AI-generated comment, and collapses it per your proposal, that they should "get a free pass", because they were actually following the guideline to the best of their good-faith ability.

As a minimum, I suggest that when you change "may" optionally to "should" normally, you don't present that as a non-change that is already enshrined in guidelines. This is a significant change; either own it as being a change, or don't propose it. WhatamIdoing (talk) 23:40, 12 January 2026 (UTC)[reply]

Who are these editors who are relying on GPTZero and nothing else? That doesn't describe anyone I'm aware of working on AI cleanup, and it doesn't describe most of what goes on at ANI (the people who bring in GPTZero or whatever tend to be uninvolved participants). Gnomingstuff (talk) 14:27, 9 January 2026 (UTC)[reply]

There was a discussion not long ago about LLMs creating spam; see here. As I said there, I think this is one way to look at it -- we will not be able to detect all uses of LLMs, but if our rules force LLMs to become hard to detect (because they have improved the usefulness of their posts) maybe that's the best outcome we can hope for. I can see why we want to ban LLMs for user communication, and for things like FAC and GAN reviews, but there is no guaranteed way to detect LLM-generated text. Plus I'd argue that in the right hands they are useful. I have used them myself to find problems in articles I have worked on, for example. TL;DR: I am not strongly opposed to a rule like the one suggested here, but I doubt it will be very useful. I don't have a better suggestion, though. Mike Christie (talk - contribs - library) 03:56, 8 January 2026 (UTC)[reply]

I wonder how long it will be before attempting a ban is just pointless, either because we can't detect it at all, or because the amount of time spent arguing over whether a comment is prohibited type of AI overtakes the cost of permitting it. WhatamIdoing (talk) 00:19, 9 January 2026 (UTC)[reply]

The amount of effort spent arguing whether something is AI-generated is already at times greater than the amount of effort spent determining whether the content of that something is actually problematic. Thryduulf (talk) 03:23, 9 January 2026 (UTC)[reply]

In my anti-goal scenario above, what's motivating B to ignore the reported error and focus on the communication style? Why is B failing at Postel's law of practical communication? Assume that B would respond in a practical manner if he was realistically able to. Is the problem within B himself (e.g., B is fixated on rule-following, to the point of not being able to recognize that the error report is more important than the method by which the error report was communicated)? Is the problem accumulated pain (this is the 47th one just today, and B's patience expired several comments ago)? Is the problem in our systems (e.g., if B can quickly dismiss a request as AI-generated, then harder work can avoided)? WhatamIdoing (talk) 05:28, 9 January 2026 (UTC)[reply]

I have to admit I've never really thought about that before. My gut reaction is to say it's mostly "accumulated pain" with a bit of over-focus in there too: Someone finds and fixes a problem, then they find another similar one and fix that. At some point they realise they've seen a quite a few of these and start looking for others to fix. This becomes an issue if they get overwhelmed by the scale of the issue and/or stop looking at the wider context to see whether it is actually a problem that needs to be fixed. Thryduulf (talk) 12:27, 9 January 2026 (UTC)[reply]

It's kind of a mix of:

• This is the 47th one today, and the majority of the last 46 people were either hostile or uncooperative.
• AI copyedits go much farther than "normal" copyedits do in terms of rewriting meaning -- they're more akin to line edits -- but AI companies do not always make it clear how much they're changing. So when someone hears "I just used it for copyediting," they're inclined to distrust that.
• A subtler one - there are at least some cultural differences in what "copyediting" entails and how transformative it is. I don't know all of these but I know American copyeditors and British subeditors have different expectations of the job.
• In reality, this kind of conversation does not usually begin with "Hi, I found a serious factual error in the article. Here's a source to show I'm not making this up," it begins with a wordy behemoth full of AI platitudes. But even those -- at least on article talk pages -- often don't result in that because many editors watching individual articles aren't aware that AI is even a thing (still). Where this conversation usually happens is someone asking a question, and receiving an AI reply.

Gnomingstuff (talk) 14:30, 9 January 2026 (UTC)[reply]

1. Plus there's another 103 to go, and it always feels like "I" am the only person doing this (because there's no way to know how many people have already checked the thing that I'm checking now).
2. How much an AI copyedit changes shouldn't be visible to the other talk page participants.
3. I love that link, and I hope someone gets her book and improves Copyediting with it. I wonder if British editors are more irritated by AI 'copyediting' for tone/voice reasons.
4. That's a testable claim. Here's the first five recently edited talk pages that contain a Template:Collapse AI top: Talk:Kristi Noem#Semi-protected edit request on 8 January 2026 (2), Talk:Primerica#Request for balanced description reflecting regulator sources + independent media sources, Talk:Scott Wiener#Wiener’s CLJC role warrants a stand-alone section, Talk:Pythagorean triple, and Talk:2026 Tamil Nadu Legislative Assembly election#Major cleanup required. In all five cases, it was either the first comment in the section or a later reply from the editor who started the section.

WhatamIdoing (talk) 19:27, 9 January 2026 (UTC)[reply]

Not sure these are great examples:

• Kristi Noem seems like a standard "why is Wikipedia mentioning this public scandal, it must be politically biased" type comment, so the good outcome was already off the table
• Primerica seems like a standard "why is Wikipedia mentioning this negative coverage of a company instead of promoting it?" type comment, which, same
• Scott Wiener comment hatted four months after it was posted; there was a substantive discussion before then
• Pythagorean triple thread began with a wordy platitude behemoth yet still was not hatted until several comments deep (after the LLM had already provided low-quality sources when asked about them)
• Hard to tell what's going on in the 2026 Tamil Nadu Legislative Assembly election thread but it looks like there was some backstory and at least one block prior to the comment

Gnomingstuff (talk) 03:25, 10 January 2026 (UTC)[reply]

I'm not saying that none of these deserved to be hatted. I'm saying that they're not evidence supporting the claim that what "usually happens is someone asking a question, and receiving an AI reply".

You can repeat the search yourself if you'd like, and pick a different sample set. It's sorted to have the most recently edited Talk: pages containing that template at the top (e.g., /Archive pages if an archive bot just did its daily run – it's most recent edit of any kind, not specifically the most recent addition of the template). WhatamIdoing (talk) 05:53, 10 January 2026 (UTC)[reply]

As a Polish native speaker, my English is strong, but it is not at a native level. I can easily understand written and spoken English, but expressing myself in English - especially writing comments - is much harder for me than it is for native speakers. Banning the use of machine translation tools (which increasingly rely on LLMs) to edit messages would be exclusionary and would push people like me out of discussions about building the encyclopedia, even when we can judge whether a translation faithfully conveys what we meant.

This is even more pronounced for non-native speakers with dyslexia. Without tools that help with grammar and punctuation, even strong substantive arguments can come across as weaker or less persuasive - not because the reasoning is bad, but because the English reads poorly. Grudzio240 (talk) 10:49, 12 January 2026 (UTC)[reply]

Myself and many others have expressed and will continue to express that a poorly-worded non-native comment will always read better and be stronger and more persuasive than a comment which reads like LLM generation. In the former we can at least be confident that the ideas and convictions presented are your own, whereas in the latter we have no means to differentiate you from any of the other people that generate some boilerplate slop and call it a day. Athanelar (talk) 10:55, 12 January 2026 (UTC)[reply]

I understand the concern: when a comment reads like it was generated, it’s harder to trust that the wording reflects the editor’s own thinking. That said, as machine translation and AI-assisted editing become more common - and harder to detect - “imperfect English” will increasingly become a marker that singles people out. In practice, that can discourage non-native speakers (and people with dyslexia) from participating, even when their underlying points are solid. I think the better approach is to focus on the substance and evidence, and allow limited language assistance (especially translation), while still discouraging using LLMs to generate arguments or positions.

Also, reactions to AI-assisted text vary a lot. Not everyone reacts negatively to AI-assisted wording, and I don’t think policy should be optimized for the most suspicious readers. If the content is clear and sourced, that should matter more than whether the phrasing “sounds too polished”. Grudzio240 (talk) 11:08, 12 January 2026 (UTC)[reply]

Regarding that should matter more than whether the phrasing “sounds too polished”, the fact is that this isn't the most glaring sign of AI writing. We've seen many editors write in a quite refined way, without being suspected of using LLM assistance, as LLMs will overuse specific kinds of sentence structures (e.g. WP:AISIGNS). This is very different from the pop-culture idea of "anyone who uses refined or precise language will sound like a LLM".

As these get associated with people using these tools to generate arguments completely divorced from policy, they get picked up as cues that make readers tune out from the substance of arguments, and end up hurting the non-native speakers they hope to help. As a non-native speaker myself, I might worry about flawed grammatical structures here and there, but I would much prefer that to other editors reading my answers with immediate suspicion due to obvious AI signs. Chaotic Enby (talk · contribs) 11:17, 12 January 2026 (UTC)[reply]

Defaults treating comments that have “AI SIGNS” as suspicious may undermine Wikipedia:Assume good faith. "Unless there is clear evidence to the contrary, assume that fellow editors are trying to improve the project, not harm it." We should start by evaluating the content; proceed to distrust only when the content or behavior indicates a real problem. Grudzio240 (talk) 11:38, 12 January 2026 (UTC)[reply]

The entire conceit of this guideline is that AI-generated comments are problematic in and of themselves. If the guideline said "ignore whether the comment is AI generated and just assess whether it violated any other policy or guideline" then it would be a pointless guideline. Obviously AI-generated comments which violate other PAGs are already forbidden -- because they violate other PAGs. The point of this is to forbid AI generating comments regardless of whether their content breaks any other PAG (obviously subject to the usual exception) Athanelar (talk) 11:42, 12 January 2026 (UTC)[reply]

I wonder how many editors agree that all AI-generated comments are problematic in and of themselves. WhatamIdoing (talk) 23:43, 12 January 2026 (UTC)[reply]

Well, the the close of the VPP discussion emphasized that “The word ‘generative’ is very, very important” and that “This consensus does not apply to comments where the reasoning is the editor's own, but an LLM has been used to refine their meaning… Editors who are non-fluent speakers, or have developmental or learning disabilities, are welcome … [and] this consensus should not be taken to deny them the option of using assistive technologies to improve their comments.” WP:AITALK was written on the basis of that discussion, and it seems to rest on exactly this distinction: LLMs should not be used to generate the substance of user-to-user communication (i.e., the arguments/positions themselves), but meaning-preserving assistance e.g. translation or limited copyediting where the editor’s reasoning remains their own) was explicitly not the target of that consensus. Grudzio240 (talk) 11:58, 13 January 2026 (UTC)[reply]

The core issue is that AI signs are heavily correlated with fully AI-generated arguments, themselves usually detached from policy. AGF is not a suicide pact, and editors used to the preponderance of flawed AI-generated arguments (compared to the few meaningful arguments where AI has only played a role in translation/refinement) might discount all as falling in the former category. This is magnified by many editors choosing to defend clearly abusive uses of AI (for example, adding hallucinated citations) as only using it to refine grammar or correct typos, even when that manifestly wasn't the case. Chaotic Enby (talk · contribs) 12:10, 12 January 2026 (UTC)[reply]

For approximately the gazillionth time, saying that text is likely to have been generated by AI says nothing about good faith or bad faith. It is pointing out characteristics of text. By this logic, adding a "copyedit" or "unreferenced" tag is assuming bad faith. Gnomingstuff (talk) 17:20, 12 January 2026 (UTC)[reply]

Yes, but it's just a short step from "I think he used AI" to Chaotic Enby's "heavily correlated with fully AI-generated arguments" to "This person is just wasting my time with fake arguments and has no interest in helping improve Wikipedia". WhatamIdoing (talk) 23:45, 12 January 2026 (UTC)[reply]

That is very much assuming bad faith in what I said – I'm not saying that we should discount comments on that basis, only that some editors will do it, and I was explaining the source of that distrust rather than defending it. Chaotic Enby (talk · contribs) 00:07, 13 January 2026 (UTC)[reply]

I agree with you: Some editors will do it. WhatamIdoing (talk) 00:28, 13 January 2026 (UTC)[reply]

I know this is not entirely what you are arguing, but this kind of LLM-nihilist stance I keep hearing like "why do you care how it was made if the content is policy compliant?" seems patently absurd to me. If the only thing we care about is the substance and not who made it or how we might as well use Grokipedia. It's rather like presenting someone with a dish of Ortolan and saying "if it tastes good, why concern yourself with the ethical implications of its production?" The ends simply do not justify the means, in my mind. Athanelar (talk) 11:33, 12 January 2026 (UTC)[reply]

I think there are huge differences in values that people have, from a cultural or philosophical perspective. Some people see AI as inherently evil and some just see it as a tool. If there was an end product of English Wikipedia, something we were actually trying to finalise one day, it seems silly to the faction that sees AI as a tool to make humans do the job of machines. You get the same thing either way, we're just making it harder on ourselves. They see no value in human labour over machine labour. The means don't need to be justified because they don't see anything wrong with them. People who prefer human output do. This is especially true in the larger context of modern society, where the system requires people to work even when there's no work to be done. If a machine is doing your job for you, that doesn't mean you don't have to work, it means you have to create a need for yourself. If robots are writing encyclopaedias, that's just another existing need filled. ~2026-24291-5 (talk) 12:13, 12 January 2026 (UTC)[reply]

You're missing the third camp of "AI is a tool, but a flawed one". Using AI as a tool to write an encyclopedia would work in theory, and might be a very real possibility in the future, but has shown its current limits, and regulating it is necessary to address those immediate concerns, rather than for more abstract philosophical reasons. Chaotic Enby (talk · contribs) 12:18, 12 January 2026 (UTC)[reply]

I would vote no on any guideline that purports to tell editors what technology they can and can't use to draft communications. It's none of our business. You don't like somebody's writing style? Too bad; don't read it. It doesn't matter if it was generated by ChatGPT or polished by Grammarly or if it's just bad writing: we can judge people for the content of their posts (e.g., WP:NPA, WP:NOTFORUM, WP:BLUDGEON, etc.), but not for the tools they use to draft that content. Also, if you've been active on Wikipedia for 3 months, maybe you don't try to write a new guideline that purports to tell everybody else what tools they can and can't use to communicate on Wikipedia. If most of your edits are about trying to fight against LLM use, you might be WP:RGW instead of WP:HERE. Levivich (talk) 17:59, 12 January 2026 (UTC)[reply]

We can and already do judge people for the tools they use to edit: that is why we have a bot policy, for example, or limitations on fast tools such as AWB. In these cases, the reason is the same as the proposed reasons to limit AI-generated writing. Namely, the potential for fast disruption at scale: someone can generate 50 proposals in a few minutes, leaving other editors in need of a disproportionate effort to address them all – or leave the unread ones to be accepted as silent consensus, as no one will take the time to analyze 50 different proposals in detail.

Additionally, it isn't necessarily helpful to say that if you've been active on Wikipedia for 3 months, maybe you don't try to write a new guideline, as newcomers can absolutely learn fast and have worthy insights – especially as you wish to judge others for the content of their posts. Chaotic Enby (talk · contribs) 18:44, 12 January 2026 (UTC)[reply]

Has anyone posted 50 proposals in a few minutes? Or even 10 long talk-page comments in a few minutes? WhatamIdoing (talk) 22:55, 12 January 2026 (UTC)[reply]

This isn't a new guideline. It's a refinement of WP:AITALK, which has existed for an entire year now. The RfC that produced the guideline was closed stating in part (boling mine): There is a strong consensus that comments that do not represent an actual person's thoughts are not useful in discussions. Thus, if a comment is written entirely by an LLM, it is (in principle) not appropriate. The main topic of debate was the enforceability of this principle..

So, if you wanted to "vote no," on that, you should have done that one year ago, which you didn't. Gnomingstuff (talk) 18:44, 12 January 2026 (UTC)[reply]

Sorry, but both of you missed what I was saying. Re CE: I didn't say to edit, I said to draft communications, and our existing bot policy already prohibits spam (as you point out). Re GS: WP:AITALK is about the content--the output, what gets published on this website--not about the method. AITALK doesn't say editors can't use AI to start or refine their posts, or to copyedit or fix grammar. Any proposed guideline that says anything like This prohibition includes the use of large language models to generate a 'starter' or 'idea' which is then reviewed or substantially modified by a human editor. or Editors are strongly discouraged from using large language models to copyedit, fix tone, correct punctuation, create markup, or in any way cosmetically adjust or refactor human-written text for user-to-user communication. would draw an oppose vote from me. Re both: how long before the community thinks repeated anti-LLM RFCs are a bigger problem than the use of LLMs on Wikipedia? Be judicious, mind the backlash, note the difference between the LLM proposals that have passed, and the ones that have failed. (Hint: the super-anti-AI proposals are the ones that have failed. The ones that allow use within reasonable boundaries have passed.) Levivich (talk) 18:52, 12 January 2026 (UTC)[reply]

The main problem with those RfCs is that the stricter proposals get shot down by editors wanting reasonable boundaries, and the more lenient proposals get shot by "all-or-nothing" editors. Given that, and the speed at which the technology advances, it isn't surprising that we are often discussing these issues – especially since recent proposals have been closed with consensus that the community wants some regulation but disagreed on the exact wording proposed. In that regards, the disruption doesn't come from the RfCs themselves, but from the inability of editors on both sides to compromise.

Additionally, we also regulate what someone may do to draft communications, with proxy editing being the best example – if we can disallow proposals coming from a banned user, we can disallow proposals coming from a tool that has repeatedly proven disruptive. Chaotic Enby (talk · contribs) 19:09, 12 January 2026 (UTC)[reply]

Re proxy edits, that is not regulating the technology used to draft communications. We don't tell people what word processor to use, or whether they can use a typewriter, or which spellchecker to use, etc. etc. This proposed guideline would be a first in that sense, and I believe is doomed for that reason.

As to the main problem with the RFCs, yes, I agree with you, but does this proposed guideline look like any kind of compromise? It's proposing rules that are stricter than the rules we have for mainspace (for Pete's sake!), and it's still trying to do the thing that the community has repeatedly said no to, which is to stop or "strongly discourage" all or almost all use of LLMs (as opposed to just "bad" use of LLMs). The drafter, in comments above, below, and elsewhere, is very transparent that the goal of the proposed guideline is to get people to stop using LLMs (as opposed to getting them to use LLMs correctly rather than incorrectly).

I'll say again the same the thing I said about the last doomed RFC: hey, go ahead and run it, maybe I'm wrong and it'll get consensus, or maybe the next one will :-P

But really, CE, you've been around long enough to know what's up, I think you know I'm right... the reason your proposal at WT:TRANSLATE is on its way to passing is because that was a good proposal that compromised and is obviously responsive to community concerns from other RFCs (btw, great job there!). This proposal is not like that, it's almost the opposite in its stubbornness.

And I know you've personally put in a lot of time and effort into trying to get a handle on increased LLM usage on Wikipedia, what with the WikiProject and all, and I hate to see those productive efforts get sunk because we (collectively) aren't being clear enough to the hard liners in saying: "No. Stop trying to stop everybody from using LLMs, it's counterproductive." Because right now, NEWLLM is still laughably short, and it's not getting any better, because we're wasting time on uncompromising proposals like this one, instead of on compromise proposals like the translation one. And, frankly, it's because people who have no experience building consensus are being allowed to drive the bus, and are driving it off the road, rather than deferring to people who do know how to build consensus (like you). Levivich (talk) 21:06, 12 January 2026 (UTC)[reply]

Yep, I think we agree on the broad strokes here. I still respectfully disagree that proxy edits are that far away from using ChatGPT to generate an argument from scratch (as in both cases, you're delegating the thoughts to someone/something else), but the crux of the issue isn't a specific policy detail, but the fact that compromises end up being overshadowed by more hardline proposals on which a consensus can't realistically be reached. Chaotic Enby (talk · contribs) 21:42, 12 January 2026 (UTC)[reply]

I am explicitly open to compromise here. I want people to propose compromises that they find acceptable. I have already changed my initial proposal in response to one such compromise. I know you know that, I just want to put it out there. Athanelar (talk) 21:52, 12 January 2026 (UTC)[reply]

You also explicitly said that you want an "extreme", maximalist rule. Demanding everything and then taking as much as you can get in practice isn't exactly the same as compromising. WhatamIdoing (talk) 23:13, 12 January 2026 (UTC)[reply]

The problem with allowing starters or ideas generated by AI is because first it permits an unfalsifiable loophole ("My comment isn't subject to this guideline because it's not AI generated, I just used AI to tell me what to say and then reworded it") and second, while the style of AI-generated posts is certainly problematic, another problem (as addressed in my guideline) is the content, and generating a starter with AI means the idea is still not yours but is rather the AI's, which is the whole thing this guideline aims to address.

If the AI tells someone to wikilawyer by citing a nonexistent policy or misapplying one that does exist, it doesn't matter if they do it in their own words or not.

So the point is to say that the ideas need to be your own, not just the presentation thereof.

As for how long before the community thinks repeated anti-LLM RFCs are a bigger problem than the use of LLMs on Wikipedia? to take a page from your own book in dismissing one's interlocutor; perhaps a person who is not active in the constant organised AI cleanup efforts doesn't have the best perspective on how much of a problem LLMs are.

I really encourage you to take some time and tackle one of the tracking subpages at WP:AINB some time. Take a look at this one of a user who generated 200+ articles on mainspace wholesale using AI with no review or verification and tell us again how the people trying to fight the fire are the real problem because they're getting everything wet in the process. Athanelar (talk) 19:10, 12 January 2026 (UTC)[reply]

Yeah, ironically, If most of your edits are about trying to fight against LLM use, you might be WP:RGW instead of WP:HERE is closer to actually assuming bad faith than anything people doing AI cleanup have been accused of. Gnomingstuff (talk) 21:04, 12 January 2026 (UTC)[reply]

...which is the whole thing this guideline aims to address. Yes, that's the problem, in my view: you are trying to address something Wikipedia has absolutely no business to address, which is what technology people use to communicate. As has been pointed out by others above, there is, first and foremost, the accessibility issues and the issues for non-native English speakers (like me btw). But beyond that, how a human being gets from a thought in their head, to a policy-compliant non-disruptive comment posted on Wikipedia, is none of our (the community's) business. It doesn't matter if they use a typewriter or what spellcheck or Grammarly or an LLM. If the output is not disruptive--if it's not bludgeoning or uncivil, etc.--we have no business telling an editor what technology they can and can't use to generate that output. (And btw if you think 200+ bad articles is a lot, lol, we've had people generate tens of thousands of bad articles, redirects, etc., without using LLM, and that's happened for the entire history of Wikipedia--we still never banned people from using scripts or bots, despite the fact that they've been abused by some, and with much worse consequences that what's being reported at AINB). Levivich (talk) 21:15, 12 January 2026 (UTC)[reply]

we still never banned people from using scripts or bots, But... we do? As pointed out before, we absolutely do restrict what technology people use to edit. You need express permission to operate a bot because of the potential for rapid, large-scale disruption.

You cannot seriously compare an LLM to a word processor or typewriter. Neither of those things is capable of wholesale generating a reply without any human thought behind it. Athanelar (talk) 21:21, 12 January 2026 (UTC)[reply]

WP:MEATBOT comes to mind. SuperPianoMan9167 (talk) 21:22, 12 January 2026 (UTC)[reply]

We don't require permission to use a script. You don't need permission to use the WP:API. What is regulated is the output--specifically, BOTPOL and MEATBOT prevent unauthorized bot-like editing regardless of whether a script is actually used or not. It's the effect, not the method, that's regulated (in fact, the effect is regulated the same way -- bot or meat -- regardless of the method!). And yes, I am absolutely comparing LLMs to the pen, the typewriter, the word processor, the spellchecker, the grammar checker, autocorrect, predictive text, etc. It's just the latest technological advance in writing tools. And LLMs are not capable of generating anything "without any human thought behind it"; they require prompts, which require human thought, and their training data is a bunch of human thought. Levivich (talk) 22:49, 12 January 2026 (UTC)[reply]

Sure, but that's like arguing that paying someone to do your homework is materially the same as if you did it yourself, because you still had to describe the task to somebody else and then they still came up with an answer. You must know you're splitting hairs by now. Athanelar (talk) 22:53, 12 January 2026 (UTC)[reply]

Maybe hiring a secretary to write a letter on your behalf would be a more relevant analogy: Bob Business tells his secretary to send a letter saying he accepts their offer to buy 1,000 widgets but wants to change the delivery date slightly. He glances over the letter, decides that it makes the points that he wanted to communicate, and signs it before mailing it.

Do you think the typical recipient of that letter would be offended to discover that Bob didn't choose every single word himself? Is the recipient likely to believe that the facts communicated did not represent Bob's own thoughts? WhatamIdoing (talk) 23:33, 12 January 2026 (UTC)[reply]

That analogy only makes sense if you assume AI never makes up new arguments, and that it is only ever used to clarify existing thoughts that have been communicated in the prompt, rather than something like "please write me an unblock request". In the latter case, the fact that the substance of the unblock request isn't an original thought (but only the request to write one) is problematic, as we can't evaluate whether or not the blocked user properly understands the issues. That specific case is very much not theoretical, as around half of unblock requests have strong signs of LLM writing. Chaotic Enby (talk · contribs) 23:42, 12 January 2026 (UTC)[reply]

That analogy makes lots of sense, if you've ever worked with (or been) a human secretary.

(You might be interested in reading about the students who used AI to apologize after getting caught using AI to cheat,[1][2] if you haven't already seen it.) WhatamIdoing (talk) 23:55, 12 January 2026 (UTC)[reply]

The problem is that this analogy is very far removed from the actual situations we're facing, and makes it harder to talk about them in precise terms. In one case, you're having a secretary playing a purely functional role of transmitting a message and helping convey thoughts to an interlocutor, possibly adding some context of their own. The key task is to transmit the information, and using a secretary (or AI) to do it makes sense. On the other hand, an unblock request aims to show that the blocked user has some level of understanding of the situation. If a secretary (or AI) writes the unblock request, with the blocked user having only told them "write me an unblock request", then the unblock request fails at its purpose. Chaotic Enby (talk · contribs) 00:12, 13 January 2026 (UTC)[reply]

But how do we know what the prompt was? If the prompt was "write me an unblock request" and that's it, then your point holds true. But what if the prompt was "write an unblock request that says [user's own understanding]"? Like, for example, "write an unblock request that says I lost my cool and said something I shouldn't have and in the future I'll be sure to walk away from the keyboard when things get too heated and also I'm going to avoid this topic area for a while"? Could you tell what the prompt was based on the output? I don't think so... Levivich (talk) 00:25, 13 January 2026 (UTC)[reply]

We don't know what the prompt was exactly, but we can get some strong indications when the user leaves unfilled phrasal templates, or apologizes for nonexistent issues completely unrelated to their behavior, or only writes generic, nonspecific commitments that could apply to literally any unblock request. In many of these cases (and, again, these are a large proportion of the unblock requests I'm seeing), I'd probably be even more worried if the prompt came from the user's own "understanding". Chaotic Enby (talk · contribs) 00:56, 13 January 2026 (UTC)[reply]

The unblock process might fail at its intended purpose, but that's entirely within the realm of normal secretary behavior. Have you never read tales like the https://www.snopes.com/fact-check/the-bedbug-letter/? Or heard stories about secretaries who make sure that the boss always remembers to buy a present for his wife's birthday, send flowers on their wedding anniversary, and so forth?

In the end, I think that it might make more sense for us to re-design the unblock process (to make it more AI-resistant) than to tell people they shouldn't use AI. Maybe a series of tickboxes, setting up a sort of semi-customizable contract? "▢ I agree that I won't put the word poop in any more articles" or "▢ I agree that I won't write long comments on talk pages" or whatever. WhatamIdoing (talk) 00:37, 13 January 2026 (UTC)[reply]

Regarding that latter point, you might want to take a look at the Wikipedia:Unblock wizard, which is a step in that direction. Chaotic Enby (talk · contribs) 00:50, 13 January 2026 (UTC)[reply]

Yes, I remember you starting that project. I think a wizard would be the right approach for a new, AI-resistant unblocking request model. WhatamIdoing (talk) 00:56, 13 January 2026 (UTC)[reply]

To note: last time someone generated tens of thousands of redirects, we had to create a whole new speedy deletion criterion for it. More generally, there have been many discussions on article creation at scale (the other WP:ACAS) and attempts at building a framework to regulate it. So, while I don't disagree that we can't control everything, the issue of disruption at scale isn't new to Wikipedia, and efforts to address it aren't new either. Chaotic Enby (talk · contribs) 21:45, 12 January 2026 (UTC)[reply]

Yeah, we did that this time, too, and kudos to the community, it got to WP:G15 much faster than it took to get to WP:X1. But you know what we didn't do about the redirects or sports articles? Prohibit, or try to prohibit, people from using scripts or templates or bots, etc. We never went after the technology that made that spam possible, we went after the editors who did the spamming, and made new tools to efficiently deal with the spam (csd's). And those were 100,000-page problems; whereas this is like thousands of articles? (How many G15s have there been so far? I see 46 in the logs in the last two days.) So like an order or two orders of magnitude less? And our response, or some folks' response, has been an order or two orders of magnitude stronger. Levivich (talk) 22:41, 12 January 2026 (UTC)[reply]

More to the point: We didn't try to "Prohibit, or try to prohibit" everyone else "from using scripts or templates or bots, etc." just because a few people abused those tools. WhatamIdoing (talk) 23:20, 12 January 2026 (UTC)[reply]

But I never said we should prohibit anything entirely, just have a framework to regulate it. Which is exactly what we've done with bots (through WP:BRFA), with mass creation of articles and redirects (through draftification and new page patrolling), etc. Chaotic Enby (talk · contribs) 23:22, 12 January 2026 (UTC)[reply]

You: "I never said we should prohibit anything entirely".

Proposal: "Editors are not permitted to use large language models to generate user-to-user communications" (emphasis in the original)

One of These Things (Is Not Like the Others). WhatamIdoing (talk) 00:10, 13 January 2026 (UTC)[reply]

As you might have noticed, I did not write the proposal. Chaotic Enby (talk · contribs) 00:12, 13 January 2026 (UTC)[reply]

The main worry I have with AI is that it is much more widely distributed. We don't have a few editors who can be blocked to get rid of the spamming, but tools that have been causing issues in the hands of a much broader range of editors, mostly because, sadly, many of them don't know how to use it responsibly. Banning the tool entirely is too harsh, blocking individual editors doesn't solve the underlying problem, meaning we're in this problem zone where it's hard to craft good policy.

G15 is for the most extreme, blatant cases, but Category:Articles containing suspected AI-generated texts contains nearly 5000 pages, while Category:AfC submissions declined as a large language model output adds another 4000, just from the last 6 months. With all the smaller tracking categories, plus the expired drafts, we're easily above 10,000 pages. Chaotic Enby (talk · contribs) 23:20, 12 January 2026 (UTC)[reply]

I agree that we're in a difficult place. I don't like the idea of Wikipedia appearing to be AI-generated (even if it's not). I don't like the idea of Wikipedia having the problems associated with AI-generated content (including, but not limited to, factual errors).

But if:

• We can't accurately detect/reject AI-generated content before it's posted
• Many people believe that it's normal, usual, and reasonable to use AI tools to create the content they need for Wikipedia
• The individual incentives to use AI (e.g., being able to post in a language you can barely read; being able to post an article quickly) exceed the expected costs (e.g., the UPE's throwaway account may get blocked)

then I think that having a rule, or even having an ✨Official™ Policy🌟, will not change anything (except maybe making our more rule-focused editors even angrier, which is not actually helpful). WhatamIdoing (talk) 00:27, 13 January 2026 (UTC)[reply]

@Gnomingstuff I think it would help readers if the summary also reflected an important nuance from the earlier RfC: it explicitly carved out cases where the reasoning is the editor’s own and an LLM is used only to refine meaning (e.g. for non-fluent speakers or users with disabilities). This consensus does not apply to comments where the reasoning is the editor's own, but an LLM has been used to refine their meaning... Editors who are non-fluent speakers, or have developmental or learning disabilities, are welcome ...

The current proposal seems materially more restrictive than that consensus, because it prohibits even “starter/idea” use and goes further by strongly discouraging copyediting/tone/formatting with LLMs:

Editors are strongly discouraged from using large language models to copyedit...

If the intent is to align with the earlier consensus, it may be worth explicitly stating that assistive uses that don’t outsource the editor’s reasoning (especially accessibility/translation-adjacent cases) are not what the guideline is trying to discourage. Grudzio240 (talk) 09:25, 13 January 2026 (UTC)[reply]

and i have one more concern about the “copyedit/tone/formatting” section: it reads as shifting the downside risk onto the editor in a way that can chill legitimate assistive use. The proposal first strongly discourages even cosmetic LLM assistance, and then says that editors who do so “should be understanding” if their LLM-reviewed comment “appears to be LLM-generated” and is therefore subject to collapsing/discounting/other remedies.

Editors who choose to do so despite this caution … should be understanding if their LLM-reviewed comment/complaint/nomination etc. appears to be LLM-generated and is subject to the remedies listed above.

That framing seems to pre-emptively validate adverse outcomes based on appearance (“looks LLM”) rather than on whether the editor’s reasoning is their own. If the intent is accessibility/meaning-preserving assistance to remain acceptable, it may be worth rewording this to avoid implying that a “looks LLM” judgment is presumptively correct, and explicitly protect meaning-preserving copyedits/formatting from being treated as fully LLM-generated. Grudzio240 (talk) 09:31, 13 January 2026 (UTC)[reply]

Updated based on the ongoing feedback. Relevant diffs: [3] [4] Please review and give feedback, particularly Levivich and WhatamIdoing Athanelar (talk) 15:42, 13 January 2026 (UTC)[reply]

Like the idea of the different prompt examples. That said, if someone is writing 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., it seems like they could just... say that, instead, without AI, and that doing so would be more likely to produce a positive outcome. Gnomingstuff (talk) 15:52, 13 January 2026 (UTC)[reply]

I have added the explanatory paragraph "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." Athanelar (talk) 16:28, 13 January 2026 (UTC)[reply]

I fully support it as is. And to address @Gnomingstuff:, it is an example of an acceptable prompt. Nononsense101 (talk) 16:20, 13 January 2026 (UTC)[reply]

I appreciate the effort but I'm probably not the best person to give feedback given I think (1) there shouldn't be a new guideline at all (Wikipedia needs fewer WP:PAGs, not more); (2) there shouldn't be a new guideline about "LLM communication" (as opposed to about LLM use in mainspace or LLM translation); (3) "Large language models are unsuited for and ineffective at accomplishing this, and as such using them to generate user-to-user communication is forbidden." is a deal breaker for me, in principle (I don't agree it's ineffective or unsuited or that it should be forbidden); (4) I do not support "a prohibition against outsourcing one's thought process to a large language model"; (5) I do not support "Editors are not permitted to use large language models to generate user-to-user communications"; (6) I do not agree with "It is always preferable to entirely avoid the use of LLMs and instead make the best effort you can on your own"; (7) the entire section "Large language models are not suitable for this task" is basically wrong, including "Large language models cannot perform logical reasoning" (false/misleading statement, they do perform some logical reasoning); and (8) I disagree with the entire section "Anything an LLM can do, you can do better". This is a guideline that says, in a nutshell, LLMs are bad and you shouldn't use them, and since I think LLMs are good, and people should use them, I don't think we're going to find a compromise text here. For me. But I'm just one person. Levivich (talk) 18:34, 13 January 2026 (UTC)[reply]

Understandable. So long as your disagreements are ideological and not "there's a fundamental contradiction" or the likes, that's still a good indication for me that I'm in the right direction. Much appreciated. Athanelar (talk) 18:44, 13 January 2026 (UTC)[reply]

I don't think this is an ideological disagreement. (Some proponents of a ban on LLMs may be operating from an ideological position; consider what Eric Hoffer about movements rising and spreading without a God "but never without belief in a devil". AI is the devil that they blame for many problems.) I do think that as someone hoping to have a successful WP:PROPOSAL, it's your job to seek information about what the sources of disagreement are, and to take those into account as much as possible, so that you can increase your proposal's chance of success (which I currently put at rather less than 50%, BTW). Feedback is a gift, as they say in software development.

For example:

1. Levivich expresses concerns about the proliferation of new guidelines. There have been several editors saying things like that recently. Do you really, really, really need a {{guideline}} tag on this? Maybe you should consider alternatives, like putting it in the project namespace and waiting a bit to see if/how editors use it.
2. He wonders whether a new guideline against "LLM communication" should be prioritized over AI problems in the mainspace. What are you going to say to editors who look at your proposal and say that it's weird to advocate for a total ban on the Talk: pages, when it's still 'legal' to use AI in the mainspace? You don't have to agree with him, but you should consider what he's telling you and think about whether you can re-write (or re-schedule) to defend against this potential complaint.
3. Your statement that "Large language models are unsuited for and ineffective at accomplishing this" is a claim of fact (getting us back to that ideology power word: opponents of LLMs are entitled to their own opinions, but not to their own facts). Are LLMs really unsuited and ineffective? Can you back that up with sources? Does it logically follow from "success depends on the ability of its participants to communicate" that a tool helping people communicate is always going to be ineffective at accomplishing our goals?
   • What if the use of AI in a particular instance is "Dear chatbot, please re-write the following profanity-laced tirade so that it is brief and polite, because I am way too angry to do this myself"? Does that interfere with the goal of "civil communication"? Or would that use of a chatbot actually improve compliance with our Wikipedia:Civility policy? Is it really true that "Anything an LLM can do, you can do better" – right now?
   • What if the use is a newbie who is pointing out a problem and who used an LLM to try to present their information as "professionally" as possible? What I'm seeing in my news feed is that Kids These Days™ aren't doing so well with reading and writing in school. Does trying to communicate clearly interfere with our goals of reaching consensus, resolving disagreements, and finding solutions?
   • What if the realistically available alternatives are also less than ideally effective? You've added a paragraph about dyslexia and English language learners (thank you), but how is the average editor supposed to know whether the person has a relevant limitation? For comparison, many years ago, we briefly had an editor whotypedallthewordstogetherlikethis and said that pressing the space bar was painful due to Repetitive strain injury, which he thought we should accept on talk pages as a reasonable accommodation for his disability. I never have been able to decide whether he had a surprisingly inflated sense of entitlement or if it was a piece of performance art, but we sent him on his way with a recommendation to look into speech-to-text software. Thinking back, I'd have preferred that he used an LLM to what he was doing. It would have been more effective at supporting communication than what he was doing. But: If he was here today, and used an LLM today, how would the other editors know that he had (in his opinion) a true medical reason for using an LLM? More importantly, if LLMs are effective for those groups of people, does that invalidate the factual claim that LLMs are "unsuited for and ineffective at" discussions?

You should go through the rest of Levivich's feedback and see whether there is any adjustment you can make that might reduce the likelihood that anyone else would vote against your proposal on the same grounds. Can you re-write it to be less strident? Less absolute?

Or take the opposite approach: Write an essay, and tell us how you really feel. Don't say "The substance of this guideline is a prohibition against outsourcing one's thought process to a large language model"; instead say something like "Whenever I see LLM-style comments on talk pages, I feel like I'm talking to a machine instead of a human. I worry that if you aren't writing in your own words, you won't read or understand my reply. I worry that if you're misunderstanding something, you won't care – you'll just tell the LLM 'she said I'm wrong; write a reply that explains why I'm right anyway'. That's not what I'm WP:HERE for." WhatamIdoing (talk) 20:51, 13 January 2026 (UTC)[reply]

This is all very good, and gives me something to work with for another round of improvements on this thing, so I appreciate it greatly. One thing I want to address specifically in this reply is the question of "why a guideline rather than an essay in WPspace?" and the answer is that while I absolutely do have a lot to say about LLMs on Wikipedia, I want to materially improve the situation by doing something about it, not just vent. The community norm is already against the use of LLMs in talk pages. People who use LLMs for that pretty much universally get told "hey, quit it" so I thought it would be sensible to make the unwritten rule written rather than having it exist in this nebulously-enforceable grey area. Athanelar (talk) 21:16, 13 January 2026 (UTC)[reply]

It's not an unwritten rule. It's written down at WP:AITALK, which is already a guideline. So – why another guideline? Why a separate guideline? WhatamIdoing (talk) 21:18, 13 January 2026 (UTC)[reply]

AITALK doesn't forbid the use of LLMs for discussions, it merely suggests that they may be hatted (which, by the way, if you didn't notice I also changed my 'should' to 'may'.) The only time LLM use in talk pages tends to escalate to sanctions is when a user persistently lies about it; which to be fair is common, but what I'm proposing is that any persistent (i.e., continuing after being notified and obviously subject to the limited carveouts) LLM usage for discussions should be considered disruptive. As my title here says, it's more WP:LLMCOMM than it is WP:AITALK. LLMCOMM begins with the sentence Editors should not use LLMs to write comments generatively. and my whole goal here is to basically turn that 'should' into a 'must' (while giving reasoning, addressing loopholes, and also synthesising AITALK into it to provide remedies/sanctions for the prohibited action) Athanelar (talk) 21:23, 13 January 2026 (UTC)[reply]

I have been looking at this with fresh eyes, and I think that the entire ==Large language models are not suitable for this task== section can be safely removed.

Overall, I feel like the bulk of the page is trying to persuade the reader to hold the Right™ view, instead of laying out our (proposed) rules.

The ===Boldness is encouraged and mistakes are easily fixed=== subsection is irrelevant. Boldness is encouraged in articles. Mistakes can be fixed in articles (though if you're listening to what people are saying about fixing poor translations and LLM-generated text, "easily" is not true). In the context of user-to-user communication, boldness has costs, and some mistakes are not fixable. Maybe a decade ago, we had an influx of Indian editors (a class?) who had some problems, and in a well-intentioned effort to be warm and friendly, they addressed other editors as "buddy" (e.g., "Can you help me with this, buddy?"). This irritated some editors to the point that there were complaints about the whole group being patronizing, rude, etc. As the sales teams say, you only have one chance to make a first impression. Even if you're just trying to fix grammar errors and simply typos, the Halo effect is real, and it is especially real in a community that takes pride in our brilliant prose (←the original name for Wikipedia:Featured articles). A well-written comment really does get a better reception here than broken English or error-filled posts.

Also, "using an LLM to communicate on your behalf on Wikipedia fails to demonstrate that you...have the required competence to communicate with other editors" might feel ableist to people with communication disorders. The link to Wikipedia:Not compatible with a collaborative project is misleading (it's about people who are arrogant, mean, or think they should be exempt from pesky restrictions like copyrights; it's not about people who are trying to cooperate but struggle to write in English).

I have been thinking about an essay along these lines:

How to encourage non-AI comments

There are practical steps experienced editors can take to encourage non-AI participation.

• Please do not bite the newcomers. People who use AI regularly are often surprised that this community rejects most LLM-style content. Gently inform newcomers about the community's preferences.
• Focus on content, not on the contributor or your perception of their skills. Don't tell newcomers that the Wikipedia:Competence is required essay says they have to be able to communicate in English. Kind and helpful responses to broken English, machine translation, non-English comments, typos, and other mistakes encourage people to participate freely. If people see that well-intentioned comments written in less-than-perfect English sometimes produce rude responses, they will be more motivated to use AI tools.
• Accept mistakes, apologies, corrections, and clarifications with grace. Ask for more information if you think the person's comment doesn't make sense. Ask for a short summary if it is particularly long.

 

but I'm not sure it would actually help. People who are most irritated by "AI slop" don't automatically all have the social and emotional skills to be patient with the people who are irritating them.

I've posted a much shorter (~20%) and softer version of this proposal in my sandbox. I tried to remove persuasive content and examples from the mainspace, as well as shortening the few explanations that I kept. I also added practical information for experienced editors (so we're permitting dyslexic editors to use LLMs, but you're permitted to HATGPT, so...let's at least not edit war?). Maybe the contrast between the two will be informative. WhatamIdoing (talk) 19:53, 14 January 2026 (UTC)[reply]

I much prefer WAID's version as it restricts itself to the point and doesn't preach or demonise anyone or anything. I would though reprhase the authorised uses section so as to focus on the uses rather than actions, advice or specific conditions. Perhaps something like:

The following uses are explicitly permitted:

• Careful copyediting: You may use an LLM to copyedit what you have written (for example to check your spelling and grammar), but you must always check the output as the tools sometimes change the meaning of a sentence.
• As an assistive technology: If you have a communication disorder, for example severe dyslexia, LLM tools are permitted as a useful assistive technology. You are not required to disclose any details about your disability.
• Translation. People with limited English, including those learning the language, may use AI-assisted machine translation tools (e.g., DeepL Translator) to post comments in English. Please consider posting both your original text plus the machine translation.

You are not required to state why you are using an LLM but in some cases doing so may help other editors understand you.

That last sentence in particular needs significant improvement though. Thryduulf (talk) 00:37, 15 January 2026 (UTC)[reply]

I do plan to synthesise some of WAID's into mine, but I still have major issues with the suggestions for how to handle some of these carveouts; because they provide any bad-faith editor (which, given the amount of people I see lie about using LLMs, is a lot) a get-out-of-jail free card. Or rather, it means we essentially can't enforce the guideline in good faith at all. We can't simultaneously say "you shouldn't generate comments with LLMs" and also say "but if you have certain exempting circumstances, you can essentially do whatever you want with LLMs with no disclosure whatsoever" because it makes it impossible for us to enforce against users using LLMs 'wrong' without inevitably catching, for example, a dyslexic editor who decides they want an LLM to compose their entire comment and so it sounds 100% AI generated. Athanelar (talk) 02:28, 15 January 2026 (UTC)[reply]

Yes, this is a problem. We can declare a total ban and thereby officially endorse discrimination against people with disabilities and English language learners into our guidelines.

Alternatively, we can permit reasonable accommodations and give editors no way to be certain that the person is using it truly qualifies for it. We can predict that we will have a number of emotional support peacocks in addition to people who don't know that it's banned, people who legitimately do fall into one of the reasonable exceptions, some rule-breaking jerks, and some people who believe that what they're doing is reasonable (in their eyes) and therefore the community's rule is unreasonable and shouldn't be enforced against them. (I'm pretty sure psychology has a name for the belief that rules don't apply to you unless you agree with/consent to them, but I don't remember what the word is.)

Plus, of course, no matter what we write, there would still be the problem of editors incorrectly hatting comments written by English language learners and autistic editors, because AI-generated text resembles some common ESL and autistic writing styles (e.g., simpler sentence structure).

Ultimately, to loosely adapt Blackstone's formulation, the question is: Would you rather discriminate against autistic people and people with limited English proficiency, or would you rather give the benefit of doubt to lazy liars? WhatamIdoing (talk) 02:59, 15 January 2026 (UTC)[reply]

I support revision 3 as is, without any changes that would further weaken its language. Having seen how LLM use is currently being handled by the community at other venues, including article talk pages, content-related noticeboards, and WP:ANI, my impression is that the discussion here is not representative of the community sentiment toward LLM use as a conduct issue, which is much more negative than is being portrayed here. A request for comment will invite input from the editors who spend more time resolving issues resulting from LLM use but do not closely follow all of the relevant village pump discussions. — Newslinger talk 05:11, 15 January 2026 (UTC)[reply]

you'd think so, but there are large amounts of people who don't seem to be aware they happened (plural) Gnomingstuff (talk) 06:49, 15 January 2026 (UTC)[reply]