Wikipedia talk:Writing articles with large language models

I've noticed that my mental AI-detector has gotten a lot weaker and a lot fainter over the past month or two. Admittedly the data on this isn't great, but I did notice that in my WP:AfC reviews, AI went from being one of my top decline categories in August 2026 to being completely absent from my October 2026 declines. I wrote WP:BYHUMANS in April 2026, and I still agree with the essay, but I also have seem enough AI accusation discussions that I also wrote WP:STILLSLOP, which effectively argues an opposite position from WP:BYHUMANS, and which I also still agree with. I will probably never know exactly how much AI is flying under my radar, but I have a feeling it's ugly. That being said, I have intentionally and knowingly accepted AI generated drafts, then stubbified them in order to delete any AI generation artifacts. So would it be a good idea to align WP:NOLLM closer to WP:BYHUMANS? Could something akin to WP:STILLSLOP be incorporated, arguing that AI is bad because the content is bad, not necessarily because it's AI? guninvalid (talk) 08:27, 17 October 2026 (UTC)Reply

It feels like this policy gets revisited daily. I just added two discussion {{please see}} below. Is your point that more reasons-for-ban should be added to the guideline? Dw31415 (talk) 12:46, 17 October 2026 (UTC)Reply

I think @Guninvalid's point is that "being LLM-generated" should not be a reason in an of itself to delete content. If it's LLM-generated and unreferenced it should be dealt with the same as human-written unreferenced text, if it's got references that don't verify the associated text then it should be dealt the the same way whether it was LLM-generated or human-generated. This is something I've long argued for, but the voices who wish to treat LLMs as a matter of principal have always been louder. Thryduulf (talk) 12:53, 17 October 2026 (UTC)Reply

Thanks, responded below. Incidentally, I’m moving away from allowing fixing “blue with envy” because I’m tired of rehashing. Better to have a simple guideline. Dw31415 (talk) 13:30, 17 October 2026 (UTC)Reply

That's not quite what I'm arguing, though I'm a little hazy on my own plans ideas as well. In my opinion, being LLM generated should be disqualifying enough. That being said, I also agree that trying to prove LLM is often a lost cause, and that notability or NPOV are often better ways to argue for removing LLM content. guninvalid (talk) 14:06, 17 October 2026 (UTC)Reply

I’d be happy to see it simplified to “every word of a Wikipedia article should be written by a human”. I bet that would even be wiki-lawyered: the generated text was originally written by a human. Maybe “Every word you add to a Wikipedia article should be written by you”. Dw31415 (talk) 13:15, 17 October 2026 (UTC)Reply

Every word you add to a Wikipedia article should be written by you would disallow the addition of quotes, titles of works (whether they are the article subject or a source), picture captions not written by the revision author, translated content, and arguably also grammar- and spell-checked material and content merged from other articles. It also completely ignores the fact that bots have been contributing to Wikipedia since the very beginning and that some of the earliest articles were imported from the 1911 Britannica (which wasn't written by those adding to Wikipedia). It's catchy but not even remotely close to practical. Thryduulf (talk) 15:08, 17 October 2026 (UTC)Reply

You and your inconvenient good points. Dw31415 (talk) 16:27, 17 October 2026 (UTC)Reply

To join in being a little pedantic, there are many bots on Wikipedia which do copyediting, clean up links or revive dead ones, automatically revert obvious vandalism, etc. The distinction I would say is that they are deterministic machines whose editing is the responsibility of the Wikipedian maintaining them, compared to large language models which are pseudorandom and at the moment are treated as if nobody is responsible, and further that the plagiaristic people who typically make use of LLMs oft fail to take up any responsibility. I think it's misleading to claim to be "human-only" or "anti-AI" when you really just dislike LLMs and LDMs (elaborated in my essay). Altoids0 (talk) 19:40, 29 October 2026 (UTC)Reply

Sorry, what is being suggested? NicheSports (talk) 16:29, 18 October 2026 (UTC)Reply

Changing the guideline to ban AI-generated content simply by virtue of being AI-generated, as opposed to the current version, which bans AI-generated content due to frequent violations of content policies. In other words, making the LLM ban purely ideological. SuperPianoMan9167 (talk) 16:47, 18 October 2026 (UTC)Reply

As @Guninvalid has said, he’s not sure. If I were to attempt to represent the suggestion: Remove all exemptions for LLM use in creating article text. It’s import that the encyclopedia is written by humans. Added content can be reverted if it was generated by LLM or lacks evidence of sufficient of human editorial enrichment. (Emphasis on the “or”). Dw31415 (talk) 16:48, 18 October 2026 (UTC)Reply

If anything AI has gotten somewhat easy to detect these days due to WP:AIDISCLAIMER popping up out of fucking nowhere to be a giant klaxon for 20,000-byte wads of AI content.

(have seen at least one source suggest that most AI chatbots' output is not becoming linguistically closer to human writing, with the exception of Claude; the source is not the strongest though) Gnomingstuff (talk) 05:01, 22 October 2026 (UTC)Reply

  Late notice, but an RfC was opened 15 days ago. You are belatedly invited to join the conversation. Dw31415 (talk) 12:36, 17 October 2026 (UTC)Reply

 You are invited to join the discussion at Wikipedia:Village pump (proposals) § Bot proposal: automatic addition of infobox for company articles. Dw31415 (talk) 12:42, 17 October 2026 (UTC)Reply

That discussion needs to be snow closed before the nomination gets themselves into more hot water, it doesn't need more input. Katzrockso (talk) 15:41, 17 October 2026 (UTC)Reply

Can LLMs be used to suggest a synonym for a word or short phrase within a sentence (that an editor manually wrote) under the "This is limited to spelling, punctuation, capitalisation, grammar, and other simple mistakes." clause (assuming that "Editors must ensure that any text copyedited in this manner is still supported by the sources cited." is followed)? Noleander (talk) 01:18, 30 October 2026 (UTC)Reply

No (unfortunately in my view) that’s not permitted by the guideline text or previous discussion. In other discussions, there’s been a healthy dose of “if you can’t tell, it’s not a problem”. Out of curiosity, what’s the word? I should also say we’re happier to have your organic, hand-written text. If you send me a diff (or reply here) I’ll gladly take a whack. (The simple correction discussed was fixing an idiom like “blue with envy”. Dw31415 (talk) 01:27, 30 October 2026 (UTC)Reply

Thanks for the info: I read that Talk page discussion about the “blue with envy” idiom, and I understand the result. I don't have a specific word in mind; my question was based on the fact that in years past I've used web sites like https://www.thesaurus.com for synonyms (e.g. to avoid repeating the same word in consecutive sentences). It sounds like I'll need to make sure that in the future I'm actually looking at a static thesaurus web site, and not an LLM list of synonyms. Noleander (talk) 01:49, 30 October 2026 (UTC)Reply

My read of WP:NOLLM is that using LLMs to find a synonym (literally one word) is fine, because that is not using LLMs to generate or rewrite article content. I did that once, when I was looking for a word on the tip of my tongue. Pinging @Kowal2701 for their thoughts NicheSports (talk) 01:59, 30 October 2026 (UTC)Reply

lots of people IAR and use LLMs for light copyediting (which'd include this), that's fine imo, the guideline is moreso written for irresponsible people where a bright, clear line is necessary Kowal2701 (talk, contribs) 14:18, 30 October 2026 (UTC)Reply

I've tried a few tests asking free llms variations of "improve spelling and grammar", and the responses included wording changes and phrasing changes, so in practice it will be hard to handle synonyms as a distinct category. CMD (talk) 23:09, 30 October 2026 (UTC)Reply

Seconding Kowal and CMD here. Using an LLM as a thesaurus for individual words isn't an issue by itself, but the changes that the LLM might suggest slowly grow in scope (up to cases of "light copyediting" rewriting entire paragraphs), and we have to put a clear line somewhere. Chaotic Enby (in solidarity · talk · contribs) 15:11, 1 November 2026 (UTC)Reply

WP:BEANS issue IMO, we cannot detect a lot of this but we at least can not encourage it Gnomingstuff (talk) 22:28, 1 November 2026 (UTC)Reply

I think it's time to upgrade this from a guideline to a full-fledged policy. We already treat it as such. We block people for violating it. We de-sysop people for violating it. At this point, calling it a guideline just makes it appear to be optional when in practice we treat it as mandatory. If it quacks like a duck ... RoySmith (talk) 12:25, 1 November 2026 (UTC)Reply

I would support this. I think the practical difference between a policy, guideline and a well-respected essay is a little blurry, but others might feel differently. Making this page a policy will prevent those people from trying to violate it just because it's "only" a guideline. Given how strongly this guideline was supported back in March, I believe this has the wide acceptance needed to promote it to a policy. Chess enjoyer (talk) 13:38, 1 November 2026 (UTC)Reply

I'd also support it. Technically, it's already treated as such (and often informally referred as our "AI policy"). In general, I'm wary of putting too much of a hierarchy between policies and guidelines: core aspects like reliable sources and notability are technically "only" guidelines, and we even have an explanatory essay rebutting most of the purported differences between policies and guidelines – in my opinion, what matters most is whether it is supported by a broad consensus, and I wouldn't be against entirely deprecating the distinction between policies and guidelines in favor of a more explicit category of "these are the ground rules the community has agreed to", which would also be much more newcomer-friendly. Chaotic Enby (in solidarity · talk · contribs) 15:21, 1 November 2026 (UTC)Reply

+1 Support -- DoubleGrazing (talk) 19:17, 1 November 2026 (UTC)Reply

If and when we do upgrade this from guideline to policy, we should probably send out a mass talk page message to every unblocked editor about the change, (partly per my comment at WP:AINB regarding keeping users informed). Some1 (talk) 20:10, 1 November 2026 (UTC)Reply

Personally I think whether or not this guidance should be publicized in some manner should be decided separately from whether or not it gets the label of a policy. In particular, if there is consensus to publicize the guidance, I don't think it should be done now, and then again if it gets relabelled as a policy. isaacl (talk) 00:00, 2 November 2026 (UTC)Reply

agreed, because there are still (still!) people saying things like "well I think AI use is OK." Imagine this happening with any other guideline. "We shouldn't worry about citing reliable sources." "I think it's fine to create hoaxes." "Disruptive editing is great actually and we shouldn't punish people for it." They would be laughed out of the room. Gnomingstuff (talk) 15:10, 2 November 2026 (UTC)Reply

Given the discussion here, I had changed it to policy, but @Levivich reverted me saying making this change requires an RfC. SuperPianoMan9167 (talk) 20:18, 1 November 2026 (UTC)Reply

I would have reverted you if I had seen it first. We need more than 5 editors to promote a guideline to a policy. Chess enjoyer (talk) 20:21, 1 November 2026 (UTC)Reply

Exactly, per WP:PROPOSAL. There's an academic paper about this: . Levivich (talk) 20:25, 1 November 2026 (UTC)Reply

Well, no harm done, as I've rolled back all the edits that I made elsewhere to reflect the update. Obviously it goes without saying that I would also support making this a policy. SuperPianoMan9167 (talk) 20:27, 1 November 2026 (UTC)Reply

Just for the record, yeah, just going ahead and changing it to policy seems to go well beyond WP:BOLD. My plan was to talk it up here and if it gained traction after a few days (as appears to be happening), propose it at WP:VPP. RoySmith (talk) 20:43, 1 November 2026 (UTC)Reply

Count me as No. 6. I support changing it to a policy. Anything to help keep people from finding grey areas and excuses to keep using LLMs. Masterhatch (talk) 20:38, 1 November 2026 (UTC)Reply

This should definitely be shaped as a full-fledged RfC, if we want to achieve this result, probably with a courtesy link at Wikipedia:Village pump (policy) (I'd say CENT is overkill for what's ultimately a semantic difference, although others might disagree). I took this as more of an informal discussion on the issue, but if we want to do it formally, we should do it properly. Chaotic Enby (in solidarity · talk · contribs) 20:41, 1 November 2026 (UTC)Reply

I agree that an RFC is the best way to make this happen. When should we start it? Chess enjoyer (talk) 20:48, 1 November 2026 (UTC)Reply

Once it’s all set up and ready to go? Some1 (talk) 21:00, 1 November 2026 (UTC)Reply

Definitely WP:CENT and VPP for any new policies. Policies are a Big Deal(tm) decision that should have the broadest possible participation. Levivich (talk) 22:33, 1 November 2026 (UTC)Reply

There’s some momentum gathering for an RfC to remove the exemption for translation (see Wikipedia talk:WikiProject AI Cleanup#LLM exemption for translation. Maybe we let that play out first? Dw31415 (talk) 21:52, 1 November 2026 (UTC)Reply

Probably for the best to hold them sequentially. Chaotic Enby (in solidarity · talk · contribs) 21:55, 1 November 2026 (UTC)Reply

Hmmm, maybe, but I don't want to commit to that. If that whole process is going to be over in a week, then for sure I'm willing to wait that long. If it's going to drag on for months and months, not so much. Let's see how things go. RoySmith (talk) 23:09, 1 November 2026 (UTC)Reply

would also support, obviously Gnomingstuff (talk) 22:29, 1 November 2026 (UTC)Reply

As is? No. It only incentivizes people to hide it (which can make the problem even worse by making people send the text through one of those humanizer tools, causing even more hallucinations). Paid editing has similar problems and is allowed through disclosure, the blanket ban on LLM text seems to come from the fact that most people misuse the tool and not from a problem with the means through which the text was made (which is what the guidelines currently addresses); aren't there already policies for spamming, non-existing sources, fake quotes and whatnot? What is achieved with the prohibition of LLM text itself? How does it improve wikipedia? I can see how banning text made by torturing a person (torturing someone to make text for a wikipedia article, not quoting text from the transcription of a torture) would be a sensible policy, most would agree the means through which the text was made is relevant in such a case and not having that on wikipedia would be a good thing, but I fail to see how prohibiting (instead of requiring disclosure) text made by an LLM improves wikipedia.

Also, it will be very hard to find people opposing this, as they would likely be using LLMs themselves and fear that posting here in support of it would bring unwanted attention to themselves, I myself mistakenly openly disclosed the use LLMs after reading the llm disclosure essay thinking it was a guideline, so now I just don't disclose it, hard to see how this benefits anyone. ~2026-58506-33 (talk) 21:41, 2 November 2026 (UTC)Reply

The prohibition against sockpuppetry also encourages people to try to hide it when they sock, but that's no reason to allow block evasion. Like you say, people often misuse LLMs. Other than the idealogical objections, we disallow them the because the time it takes to find and fix LLM hallucinations and whatnot is much more than the time it takes to generate bad content with an LLM. Maybe if LLM users properly reviewed their contributions before publishing them we wouldn't need this, but this does not happen. They very frequently violate other policies while using an LLM. Chess enjoyer (talk) 22:37, 2 November 2026 (UTC)Reply

Sucketpuppetry isn't blanket banned (WP:SOCKLEGIT). "we disallow them the because the time it takes to find and fix LLM hallucinations and whatnot is much more than the time it takes to generate bad content with an LLM", using LLM agents to automatically edit already falls under WP:BOTUSE & WP:MEATBOT, hallucinations under WP:VERIF, and so on, so the only thing that could possibly be addressed by this guideline isn't: the ethics of using LLMs (probably because the case for it is much weaker). "Maybe if LLM users properly reviewed their contributions before publishing them we wouldn't need this, but this does not happen. They very frequently violate other policies while using an LLM" Survivorship bias, the rigorous users who know how to operate a LLM go under the radar while the people who copypaste the first reply of chatgpt into an article make everyone mad leading to a ban on the tool instead of its misuse.

Maybe its easier to do it like that, would also have been easier to blanket ban sockpuppetry and paid editing, but for whatever reason we didn't. ~2026-58506-33 (talk) 23:18, 2 November 2026 (UTC)Reply

Also forgot to add how hysterical this becomes when another user has to ask if he can use a LLM to find synonym suggestions, this seems legitimately insane, why would someone even think it's any different than using any random website that lists synonyms? You'd have to verify if the synonim is appropriate yourself anyway, how can the means through which the synonim was derived be of any relevance? Alas ~2026-58506-33 (talk) 23:26, 2 November 2026 (UTC)Reply

Why is it so agonizing to you to follow a request by other people that you change your behavior in this one limited activity (editing Wikipedia), and thus demonstrate basic respect for other human beings, rather than dig in your heels and go "na na na na na hahah you can't make me, I don't care what you think, I do what I want"? Gnomingstuff (talk) 19:12, 3 November 2026 (UTC)Reply

Why is it so agonizing to you to follow a request by other people that you change your behavior in this one limited activity (editing Wikipedia)

"Why can't you just be quiet and do as I say" oh, wow. Because its frustrating, not agonizing, it lacks justification, I'm not doing something just because a mob of people say so without being able to properly articulate why, who just complies with whatever anyone says? As written this guideline would prevent me from improving wikipedia, so I ignore it.

and thus demonstrate basic respect for other human beings

That doesn't show respect, it shows submission, respect is when someone explains you why the thing you are doing shouldn't be done and you abide, and none of those LLM guidelines or discussion tell wikipedia users why they shouldn't use a LLM, they just say that most people don't use them properly, and what does that have to do with me?

dig in your heels and go "na na na na na hahah you can't make me, I don't care what you think, I do what I want"

I'm not saying no one can make me, I'm not snarky about it, just pointing out how the guidelines don't address anything, I care about what others think, insofar as they think relevant and thought provoking ideas, "llm bad" doesn't address the problem and isn't engaging, it's just though-terminating cliché. I already said, the problems of sloppy LLM use are already addressed by current wikipedia policies, making this guideline pointless, actually, its not pointless, its detrimental because it creates a very interesting scenario: a perfectly factually correct and well cited and formatted wikipedia article would need to be removed because of this, so my question is: **how does removing said article improves wikipedia**? Wikipedia won't become more complete, more accurate, so what exactly is gained here? I hear you thinking "oh but the vast majority of LLM submissions aren't like that" and I've already addressed that in the previous comment, survivorship bias plus the fact that wikipedia already has policies that allow you to remove, warn or ban whoever is misusing LLMs by spamming, doing disruptive edits, adding false information, so the question stands, what does this guideline, as a blanket ban instead of requiring disclosure, addresses? ~2026-58506-33 (talk) 22:16, 3 November 2026 (UTC)Reply

Everyone editing English Wikipedia is expected to abide by our policies and guidelines, even if they disagree with some or all them. If you do not care about what others think, then this project is not for you. CMD (talk) 00:17, 4 November 2026 (UTC)Reply

{{talk quote block|I care about what others think, insofar as they think relevant and thought provoking ideas}}

I'm not interested in tribalism and through-terminating gotchas, but in arguments and reasonable discussion, from the way people here talk about LLMs it's very clear they have a very basic understanding of how to operate them and most likely have only interacted with LLMs in their browser and never built a whole scaffold of tools that allows a LLM to automatically search, fetch, quote, fact-check, lint and copyvio check it's output with a single instruction, it's not without a reason that the complains are about Gemini and Chatgpt and not articles built by agent swarms. ~2026-58506-33 (talk) 01:03, 4 November 2026 (UTC)Reply

@~2026-58506-33I'll take this bait because I haven't been jaded from talking about this for years.

it lacks justification, I'm not doing something just because a mob of people say so without being able to properly articulate why

You can read through the talk page archives on this page, which contain, among other things, people articulating why this guideline is the way it is and why it ought to be the way it is. That's the great thing about new Wikipedia guidelines.

who just complies with whatever anyone says

Well, NOLLM is not "whatever anyone says". It is a guideline backed by community consensus. If you want to partake in volunteering in a community‚ you are expected to follow its rules. If you don't like the rules, you can try to change them—you won't be the first and you won't be the last—or you can leave and go do something else in your free time.

As written this guideline would prevent me from improving wikipedia, so I ignore it.

It seems like you can write several paragraphs of coherent, grammatical English with little effort. If you can do that, you can improve Wikipedia without using a LLM. Not only are you doing that, you are also communicating on an project-space talk page, making reasoned arguments, and competently citing Wikipedia policy. There is no shortage of improvements on Wikipedia you can make here.

That doesn't show respect, it shows submission

Yes, deference to WP:CONSENSUS in a community governed by consensus-based decision-making is basic respect. Being able to follow instructions is an essential part of partaking in a collaborative enterprise.

respect is when someone explains you why the thing you are doing shouldn't be done and you abide

That is respect, yes, but that also goes far beyond respect. One way to respect people's time is by reading and finding out on your own why you should or shouldn't do something.

none of those LLM guidelines or discussion tell wikipedia users why they shouldn't use a LLM

Yes, they do. NOLLM tells you LLM-generated text often violates several of Wikipedia's core content policies, and can be produced at a rate that makes review by other editors infeasible. The fact that you are not personally convinced is a separate question.

they just say that most people don't use them properly, and what does that have to do with me

If your use of LLMs violates community guidelines, then you are not using them properly. Regardless, no guideline or policy has anything to do with you personally. They are here to facilitate a collaborative project. Why do you have to go through airport security when you know you're not carrying anything they're looking for?

I'm not saying no one can make me, I'm not snarky about it, just pointing out how the guidelines don't address anything

This is a point that has been litigated before. Explaining why you should follow a guideline is not part of the purpose of a guideline.

"llm bad" doesn't address the problem and isn't engaging, it's just though-terminating cliché. I already said, the problems of sloppy LLM use are already addressed by current wikipedia policies, making this guideline pointless

Yes, it is a thought-terminating cliche. Every policy, guideline, formalized procedure, and everything in between is a thought-terminating cliche. The utility of making an explicit rule for something is that you save yourself the effort of justifying every decision you make from first principles. By your account, you could make a single policy of "make good encyclopedia" that makes every other policy and guideline obsolete. We have WP:NOR and WP:FRINGE so we don't have to evaluate every person's unpublished manuscript by their academic merit just in case we miss out on documenting the next great scientific breakthrough. We have WP:Notability so we do not need to endlessly litigate the long-term encyclopedic significance of each individual competitive cyclist and New Jersey state legislator. In the same vein, we have NOLLM so we don't have to inspect each chunk of LLM output for hallucinations. There are plenty of editors I would trust to publish their original research on Wikipedia, and plenty of subjects that I think merit their own page despite not satisfying any notability guidelines, and plenty LLM-generated prose already on Wikipedia likely better than what any human will ever write about the same subject.

actually, its not pointless, its detrimental because it creates a very interesting scenario: a perfectly factually correct and well cited and formatted wikipedia article would need to be removed because of this, so my question is: **how does removing said article improves wikipedia

In the scenario where we know ahead of time that an article is perfectly factually correct and well cited, you are right. It would indeed be detrimental to remove the article. However, the reality is that you don't know this ahead of time. To make sure, you have to go out of your way to check each claim against the cited source, and you have verify that each source is reliable enough to support the claim. This not even considering that a Wikipedia article needs to satisfy more requirements than being factually correct and well cited: it needs to abide by WP:NPOV and be free of WP:Plagiarism and WP:Close paraphrasing. Very frequently, you will spend a long time in verifying things, and you will come to the conclusion that an article has too many hallucinations to be salvageable. In that case, you will have wasted your time and effort, as the article will need to be deleted and rewritten from scratch anyway, and by invoking NOLLM you probably could have come to that conclusion much sooner. In the case that you meticulously inspect every aspect of the article and conclude that it has few or no relevant issues, you will still have invested a large amount of effort and time that you will never get back. If instead you had presumed the article unworthy of inclusion and moved on, you could have spent that time helping the project in other valuable ways, perhaps even contributing more quality content in the same time than the amount of content you might have lost by discarding the LLM-generated content.

Wikipedia won't become more complete, more accurate, so what exactly is gained here? I hear you thinking "oh but the vast majority of LLM submissions aren't like that" and I've already addressed that in the previous comment, survivorship bias plus the fact that wikipedia already has policies that allow you to remove, warn or ban whoever is misusing LLMs by spamming, doing disruptive edits, adding false information, so the question stands, what does this guideline, as a blanket ban instead of requiring disclosure, addresses

The answer is editor labor. Policies and guidelines help apply editor labor to the subjects that need it the most. Although editors provide labor for free, editor labor is Wikipedia's most valuable and most needed resource, one which needs to be rationed carefully. Expending vast amounts of volunteer editor labor into reviewing LLM content is an inefficient and irresponsible way to build an encyclopedia, for similar reasons to how expending vounteer editor labor into reviewing original research is an inefficient and irresponsible way to build an encyclopedia. Disclosure doesn't help either, for the same reason that we still disallow disclosed original research—the disclosure does nothing to improve the quality of the underlying content.

I hope I have given you some relevant and thought provoking ideas. Kisseran (talk) 01:00, 4 November 2026 (UTC)Reply

I'll take this bait

There's no bait to be taken, this guideline has to be rearticulated to target bad llm use, not ban every single word a LLM produces from being published in the website.

You can read through the talk page archives on this page, which contain, among other things, people articulating why this guideline is the way it is and why it ought to be the way it is.

I read at least one discussion on the Village pump (policy) from a month ago, the "Recent changes to WP:LLM are too restrictive and actively harm some parts of Wikipedia", and my points stand, the arguments come from people who don't know how to use the tool complaining about other people who also don't know how to use it, "llms hallucinate", well, humans too, don't we have policies for people who constantly add false information? And for those who do in en masse? They already cover pasting garbage from LLMs into wikipedia.

Well, NOLLM is not "whatever anyone says". It is a guideline backed by community consensus.

A consensus arrived from...? From my admittedly short view on the history of this guideline what saw were quite a few people distraught by the fact that they weren't able to participate in the discussion and prevent this guideline from becoming what it is now.

If you don't like the rules, you can try to change them.

I would love to engage in an earnest discussion about this, but there hasn't been one here, what there is is a large crowd of people who categorically oppose LLMs regardless of anything, and by default they are much more comfortable being vocal about it since they have no skin in game.

And as I said previously, since LLMs are not allowed, most who support it and oppose this won't bring attention to themselves by publicly defending it.

...you can improve Wikipedia without using a LLM.

I can, I could also do it blindfolded, but it wouldn't last long since it would be exceedingly cumbersome, so I view the LLM who can write much better and faster than me, allowing me to reserve myself to the much simpler task of proof reading, which I am much better at than writing, so I don't see much of a point in putting my effort into writing itself.

Yes, deference to WP:CONSENSUS in a community governed by consensus-based decision-making is basic respect. Being able to follow instructions is an essential part of partaking in a collaborative enterprise.

I know the policies and guidelines didn't pop out of nowhere, the part about submission and respect wasn't about guidelines, but in this specific case I believe it applies because as I've said this guideline was implemented terribly, I believe this comment(and whole page) here is quite relevant.

If your use of LLMs violates community guidelines, then you are not using them properly.

Writing a wikipedia article that follows all policies and guidelines (verifiability, notability, NPOV etc) isn't propper usage only because of WP:LLM, which brings me back to "how does this improve wikipedia?"

They are here to facilitate a collaborative project. Why do you have to go through airport security when you know you're not carrying anything they're looking for?

I would believe so if the policy was about disclosure and to be less lenient on people who misuse LLMs, not to outright prohibiting it.

Yes, it is a thought-terminating cliche. Every policy, guideline, formalized procedure, and everything in between is a thought-terminating cliche.

The complaint about thought-terminating cliches wasn't directed at the guidelines or policies, but at the people in the discussions about this guideline, they do not articulate well their point, they just go back to mob mentality and "llm bad"; as I said if the policy/arguments were along the lines of "LLMs are easy to abuse so we will be less lenient on people who infringe the rules with them" I would be 100% behind that, but they aren't.

In the same vein, we have NOLLM so we don't have to inspect each chunk of LLM output for hallucinations

Same point I stated above, then it should've been about being harsher on people who do it and giving people who have a good history a permission to use it.

The answer is editor labor.

Sure, but as is it also hinders the people who can use it 👇

Expending vast amounts of volunteer editor labor into reviewing LLM content is an inefficient and irresponsible way to build an encyclopedia.

I completely agree, but then shouldn't the ban be for unregistered, new, or otherwise problematic users, instead of everyone?

Disclosure doesn't help either, for the same reason that we still disallow disclosed original research.

I think this is categorically different, original research is just an enforcement of verifiability, disclosing original research doesn't change the fact that it goes against the core policy of verifiability, LLM disclosure (like paid editing) is about transparency, not a carte blanche to regurgitate LLM text, hence my point about why it shouldn't ban everyone but just people who do not clear the bar on the skills to use it.

I better understand the rationale now, its a blunt tool to tackle a monumental issue, still, the issue persists: it should not be a policy without some serious discussion on the creation of mechanisms to allow users to use this tool to improve wikipedia; I'm much better at reading than writing, and wasting time trying to humanize perfectly correct text is not cool.

Thank you for your time, I truly appreciate it. 🙏 ~2026-58506-33 (talk) 02:29, 4 November 2026 (UTC)Reply

I'm not active in this part of Wikipedia but the way I understand it, the fact Wikipedia is a collaboration project maintained by volunteers is the one argument having the most weight that will always outweighs others. Your suggestion that there should be an exception will eventually require more, not less, volunteer time in managing.

Suppose there's a consensus about an exception. Over time, naturally there'll come a time when somebody suggests a whole new system to manage that exception is necessary. That system may address questions like, who determines whether a user is sufficiently competent to use an LLM, who evaluates whether LLM-assisted contribs align with Wikipedia's rules etc. Such a system will grow with time for sure.

The way I see it, the community has decided to deal with the problem as early as it can, and not let it grow to the point of it being intolerable and unmanageable, or to the point where building new systems are necessary. Part of the objection to this is also ideological, that LLM use should be prohibited no matter what, even if it means affecting otherwise high-quality articles that tick all the right guidelines, policies etc. AdaCiccone (talk) 13:01, 4 November 2026 (UTC)Reply

@~2026-58799-35 Other people have thoroughly argued why ignoring the current guideline is bad, but you do suggest some alternatives to the current guideline.

as I said if the policy/arguments were along the lines of "LLMs are easy to abuse so we will be less lenient on people who infringe the rules with them" I would be 100% behind that

This is a good idea, in fact it is so much of a good idea that it is current practice, but it is not enough. For one, this is because detecting LLM abuse is disproportionately more difficult than LLM abuse itself. Again, if we have perfect knowledge of who will abuse LLMs and who won't abuse LLMs, the guideline would be unnecessary. Secondarily, the blocking of an editor is not in itself a basis for removing content, which means that removing the low-quality AI slop is far more difficult without NOLLM. I don't think anyone disagrees that there's a lot of AI slop on Wikipedia that has no sound reason to be on Wikipedia. If you're suggesting that we relax NOLLM but strengthen LLMPROD, then sure, that sounds logically coherent. It would certainly be a better solution in a world where all LLM abuse were caught immediately, but we don't live in such a world. It all comes down to editor labor.

it should've been about being harsher on people who do it and giving people who have a good history a permission to use it

This doesn't solve the problem of editor labor though. Now you have to set up a system to evaluate whether an editor has a good enough track record to be given permission to use LLMs. Could such a system result in more "good" LLM content that improves Wikipedia? In theory, yes. At the same time, it would be balanced out by the extra effort involved in maintaining such a system that could have been spent on content work anyway. As much as you might have the expertise and self-restraint to use LLMs responsibly (I don't know if you do, but I'll suppose you do), the number of people who could actually be trusted to use LLMs more than currently allowed is very small. Every person who applies to this hypothetical permission needs to not only have the relevant expertise, but demonstrate this convincingly to someone who has the expertise to judge this, who in turn must be trusted by the community at large to have that expertise. Applications to this permission will be riddled by people who lack the appropriate self-awareness, and there will be endless debates over who has or lacks the competence to use AI and who has or lacks the competence to judge others' competence. These processes will inevitably cause friction, and people will get blocked or quit Wikipedia in the process. Many people who might qualify for the permission won't be inclined to apply, maybe on ideological grounds, maybe because they enjoy working parts of the Wiki where LLMs are either already allowed (e.g. coding new tools), maybe because they simply don't want to use AI after using AI all day at work. All of this overhead will still be there, even if you successfully institute your perfect version of this policy.

it also hinders the people who can use it

Well, yes, the convenience of some (or indeed many) people is something you have to sacrifice when operating a collaborative, volunteer project. There are many people I would personally trust to unilaterally delete any page, but policy still requires deletion processes.

shouldn't the ban be for unregistered, new, or otherwise problematic users, instead of everyone?

Well, it could be. That would be an improvement from imposing no restrictions on LLM use, but time and time again at AINB we see otherwise experienced, respected editors misuse LLMs, often mistakenly thinking that their AI use is responsible, while in reality they had been subtly introducing hallucination-infested prose to hundreds or thousands of articles. The reality is that there's no reasonable metric on Wikipedia you can use to evaluate if someone has the necessary competence and judgment to use LLMs more intensely than currently permitted, and there most certainly isn't any way to convince the community to trust that you have that competence and judgment.

original research is just an enforcement of verifiability, disclosing original research doesn't change the fact that it goes against the core policy of verifiability

So why is it the case that verifiability is a core policy? Core content policies are pragmatically justified like any other policy and guideline. There's certainly nothing inherently unencyclopedic about original research (at least as defined on Wiki). Plenty of paper encyclopedias do not provide rigorous references and citations, and plenty of original ideas or syntheses thereof have been published in paper encyclopedias (especially specialist ones). An important difference is that those paper encyclopedias are often written by respected experts in the relevant field, who are paid for their work, and putting at least some of their valuable reputation on the line that the facts are accurate. On Wikipedia, you can't (or at least shouldn't) just take any anonymous editor's word for something, and you especially can't when there are editorial disputes to resolve. An academic submitting an entry for Britannica doesn't need to defend their work from being removed or altered by anyone out there who cares enough about it. Verifiability isn't an inherently desirable of encyclopedias, but accuracy is. Verifiability is in turn a very good heuristic to enforce accuracy: verifiability is much easier to determine than accuracy, so we save much more time in determining verifiability when compared to the miniscule value to be gained from judging the accuracy of an unverifiable claim. Original research in turn makes verifiability easier to enforce: you save all the effort of arguing if, e.g. two sources come together to verify a novel claim.

LLM use is a similar heuristic. Some people could be out there using highly sophisticated agent swarms with frontier models and intricately engineered prompts with rigorous human oversight. That, however you want to put it, is a tiny, negligible minority of people that would comprise a tiny minority of LLM content. Like I said above, there is a lot of valuable content that a prohibition on original research would exclude from Wikipedia, and plenty of editors who could be trusted to improve Wikipedia by being exempt from the ban on OR, yet it is not the case that Wikipedia would be benefit from carving out policy exceptions for those editors.

If you want to make a difference, I think the most meaningful way to start is by convincing people that your usecase for LLMs is a valid one. Write a proper article on a complex, nuanced topic using your workflow, and document every step of the process. If the workflow doesn't consistently produce content at an exceptionally high standard, then it's unlikely to be persuasive. Kisseran (talk) 20:04, 4 November 2026 (UTC)Reply

If you want to make a difference, I think the most meaningful way to start is by convincing people that your usecase for LLMs is a valid one. Write a proper article on a complex, nuanced topic using your workflow, and document every step of the process. If the workflow doesn't consistently produce content at an exceptionally high standard, then it's unlikely to be persuasive.

That's precisely what I wanted the talk around LLMs in wikipedia to revolve around, ways to mitigate the problem by teaching people how to properly use it, as I see it just pointing to the misuse of the tool without instructing people on how to use it is pointless, I will see what I can come up with, there should definitely be a project about developing and improving tools for LLM usage in wikipedia, but who would advertise themselves as using LLMs in the current scenario? I myself have it set up so every written line/paragraph is directly linked to a page in a document I have saved, the inline <ref> citations formatting are all saved in with the books/articles ISBN, DOI, authors etc, double checked using the info printed in the files themselves checked against online tools like isbnsearch, amazon, etc, with the LLM presenting to me conflicts or gaps (funnily enough through this pipeline of mine I ended up learning that the native citation tool on wikipedia is very unreliable) so the data in the <ref> tags is always the same one pulled from the db, the only variable being the cited pages, after writing the draft it runs a copyvio check and rephrases as needed, then I run multiple blank slate fact-checking rounds where the model is explicitly instructed to find flaws in the text, at most two or three depending on the results, those rounds tell me how sure it is of what it wrote, if it keeps correcting itself on some or multiple points then this indicates that either the source has problems (it could contradict itself somewhere or be contradicted by another source) or that the model is incapable of this task, which is where I would have to check myself, but if it finds less and less problems then it converged and all I'm left with is doing the proof reading, converging doesn't mean its correct of course, otherwise I wouldn't need to proof read, its just a data point on the cohesiveness of the model and sources, my setup is somewhat haphazardly made, but it gets the job done; another option is to have people familiarize themselves with the proper use of LLMs is to use them for fact checking articles, people frequently stretch the truth, not out of malice, but because they think they understood what they saw, this happens a lot when citing statistics and is a good exercise on statistical literacy.

I acknowledge that probably (certainly) less than a fraction of a percent of users do it like that, but I reiterate that this should mean wikipedia ought to have a place for people to develop and improve these tools, like how we have WP:JAVASCRIPT, WP:GADGET, WP:BOTS and on, but as I said, it doesn't seem like this can happen in parallel with this guideline, its an awkward position, and the community seems much more keen on shutting it down than navigating around it, leave the ban in place, make it a guideline, police, law, commandment of god or whatever, seems necessary for how things currently are, but this should happen in tandem with efforts to educate people on it to diminish the problem, that's where I believe the community efforts should be aimed at since these tools are here to stay and become more and more ubiquitous as the days passes. ~2026-58799-35 (talk) 21:45, 4 November 2026 (UTC)Reply

It's also worth saying that Wikipedia has no obligation to justify its policies and guidelines to you. You can choose to follow them or choose not to follow them, and if you choose not to follow them then the appropriate sanctions would follow. But you don't get to say "I'm only going to follow the ones that make sense to me." A lot of people think rules like WP:WEIGHT and WP:FRINGE shouldn't apply, for example (notably Larry Sanger, who recently and very dramatically got himself CBANNED from Wikipedia for his tendentious campaigning against those and others).

The community has decided by overwhelming consensus that LLM-generated text isn't welcome on Wikipedia. It's difficult to provide any one explanation for why we did that, because all kinds of people supported that restriction for all kinds of different reasons, some pragmatic, some ideological. The only part that matters is that the restriction exists. You can follow it or you can not follow it, but nobody owes you an explanation to get you to follow it.

I'm not going to try to justify it myself because that's already been done here by somebody else. But you say As written this guideline would prevent me from improving wikipedia, so I ignore it and that's just not how it works. You don't get to ignore a guideline because you don't think it's been satisfactorily explained to you. Athanelar (talk) 15:40, 4 November 2026 (UTC)Reply

It's also worth saying that Wikipedia has no obligation to justify its policies and guidelines to you.

It has no obligation to anything, especially not towards me, but it should have some accountability to itself, I don't think you believe any arbitrary thing can become a policy merely because people agreed so, without some kind of practical or ethical justification, no?

A lot of people think rules like WP:WEIGHT and WP:FRINGE shouldn't apply

A lot of people believe there is bearded man in the sky, not relevant, if many people fail to see why those policies exist that's immaterial to me, those policies have very obvious practical justifications and anyone's failure to see them isn't of my concern.

The only part that matters is that the restriction exists.

Yes, and before turning the terrible guideline into policy it should be revised, maybe start by only allowing only registered editors with a million edits and 20 year old accounts if it must be so, I believe it's not that complicated to create and automatically add such a flag to accounts and they certainly know how to write wikipedia articles and do their due diligence in checking the text a LLM outputs, see how it goes, is such a mechanism unfeasible and would create a big unjustified burden in the maintainers of wikipedia? I assure you if you ask experienced editors who know how to operate a computer and LLMs they will be very keen on demonstrating the community how powerful this tool is and how it warrants the necessity of implementing a way to allow people to use it.

"As written this guideline would prevent me from improving wikipedia, so I ignore it" and that's just not how it works. You don't get to ignore a guideline because you don't think it's been satisfactorily explained to you.

I don't ignore guidelines because they weren't explained to me, I do because this one in particular doesn't apply to what I do, see WP:TRI, WP:WIARM: "Don't follow written instructions mindlessly, but rather, consider how the encyclopedia is improved or damaged by each edit" & "The spirit of the rule trumps the letter of the rule.", this rule exists to tackle people who mindlessly copypaste garbage from browser chatbots, not to prevent people from adding any text that was outputted by a LLM (such as factually correct information with inline citations that were verified by a human—what I do), hence why it needs serious revision before becoming a policy. ~2026-58799-35 (talk) 16:36, 4 November 2026 (UTC)Reply

I don't think you believe any arbitrary thing can become a policy merely because people agreed so I do absolutely believe that is how it works in a consensus-based system, yes.

Convincing people to come to consensus about something usually requires some sort of reason to compel them to do so, but still, the reason why the thing is banned is because there was consensus to ban it, not because of some kind of logical reasoning. It may be that the consensus was informed by a general agreement about a particular logical reasoning, but still, the ban is justified by the consensus, not by the reasoning.

But, like I said, the AI thing is unique in that a lot of different people threw their support behind it for a lot of different reasons. There were some people in those discussions who were of the belief that we should ban AI use for now and then lift that ban later on when AI gets better at not breaking Wikipedia's rules. There are other people, myself included, who are of the ideological belief that we should ban AI generated text wholly and for good, simply for the reason that we're philosophically opposed to it. The answer to the question "why is LLM generated text banned on Wikipedia" is "Because the community decided that should be the case." The question "Why did the community decide that should be the case?" doesn't have any one answer. The closest you're going to get is "because we decided that LLM output generally isn't conducive to good Wikipedia content, and LLMs can produce that output at such high pace that it's impossible for us to keep up with checking whether each individual piece of it is appropriate on a case-by-case basis, so it's easier to just get rid of it as soon as we detect it."

those policies have very obvious practical justifications and anyone's failure to see them isn't of my concern

Some people would say the same about you and this policy. That's my whole point. All kinds of people disagree with lots of our PAGs for lots of reasons. We have absolutely no concern whatsoever for convincing those people. The PAGs are what they are, if you want to participate here you're signing a social contract to be bound by the community's consensus in exchange for getting to be part of that consensus. I think our stance on paid editing needs a lot of change, but the community disagrees with me. It is what it is.

and they certainly know how to write wikipedia articles and do their due diligence in checking the text a LLM outputs

You'd think that, but we've already had a couple of high profile cases of very experienced editors, even admins, running into trouble because they were using LLMs and they got caught out for major source-to-text discrepancies and that sort of thing; see WP:Administrator recall/Fuzheado for example. Massively veteran editor in very good standing, using a custom model rather than an off-the-shelf chatbot, exactly the kind of person who you're saying should be given the "LLM pass", and they were still introducing broken citations and source-to-text discrepancies, as described at the AINB discussion linked at the top of the admin recall discussion. Trust me, you are not the first person and will not be the last to suggest that we should permit experienced editors to use LLMs because they can surely be trusted to verify the output; but time and time again we see that is not the case, and even very experienced editors let very stupid mistakes slip through the cracks.

this rule exists to tackle people who mindlessly copypaste garbage from browser chatbots, not to prevent people from adding any text that was outputted by a LLM (such as factually correct information with inline citations that were verified by a human—what I do),

This is just not correct. The rule exists for exactly what it says in the first sentence; to strictly prohibit the use of large language models to generate or rewrite article content. Again, you're not the first nor the last to make this "spirit of the law" argument regarding WP:NOLLM; there's a whole essay about it, WP:YESNOLLM. Athanelar (talk) 17:01, 4 November 2026 (UTC)Reply

Yeah, I support it. Per WP:PAG#Role, this page has always been far more of a "policy" (a standard[s] all users should normally follow) than it is a "guideline" (a set[s] of best practices) anyway. Athanelar (talk) 15:34, 4 November 2026 (UTC)Reply

I'm not going to do a formal close of this (although if somebody wants to, I won't object), but I have taken the above discussion as broad agreement so I've gone ahead and posted this at WP:VPP. Please continue the discussion there. RoySmith (talk) 16:11, 4 November 2026 (UTC)Reply

See here. This is a discussion of a couple of edits I made a week or two ago, when I hadn't read as much as I have now about NOLLM and the reasons for it. I'd be interested in getting some opinions from here over there -- both "were those edits forbidden by NOLLM?" and "What should be done about it now, if anything?" I am finding the boundary of NOLLM hard to define and it seems I'm not the only one. Mike Christie (talk - contribs - library) 19:33, 6 November 2026 (UTC)Reply

To wit, the question under discussion there is what kind of copyedit-like edits are permissive - for example, having a LLM write a piece of math format? Or having a LLM provide a list of sentences with a proposed rewrite for each? Permitted or not? Jo-Jo Eumerus (talk) 19:55, 6 November 2026 (UTC)Reply

We should keep the discussion centralized in one place. I’ll go there and leave it to others to decide if it should move here. Dw31415 (talk) 20:12, 6 November 2026 (UTC)Reply

Viz Kowal2701's edit, I am kinda unsure on consensus or purpose of that provision (enforceability is another question altogether). I figure that there might be a consensus of sorts (or something that was misread as indicating a consensus) in Wikipedia talk:Writing articles with large language models/Archive 4#Refining basic copy editing since the RfC above was closed procedurally. Is the purpose mainly that to "close a loophole", as framed above? Style and flow improvements often result in minor SYNTH issues even when done by humans, and from looking at WP:AINB it doesn't seem like there is a notable tendency of "flow improvements" to be used as an excuse or to contribute to the backlog. (note to self: I'll have to remove that flow suggestion prompt from my user talk page if that is what is meant) Jo-Jo Eumerus (talk) 16:39, 11 November 2026 (UTC)Reply

it's more to make the current provision clearer, the real issue is using LLMs for expansive copyediting (as opposed to light copyediting), so this is just a clear place to draw the line somewhere around there. The guideline's moreso targeted at editors predisposed to irresponsible LLM use, I think it's perfectly fine for responsible editors to IAR a bit on this, and many do. Issue is, there's not really a way to incorporate that into the guideline that won't lead to massive amounts of wikilaywering (even atm there's a lot of it, but imo the exception's worth the tradeoff since it's easy to just say "no, that clearly isn't the case"). There was a similar discussion above started by @Noleander (Wikipedia talk:Writing articles with large language models#Synonym suggestions permitted?). If we get to the point where we're not getting flooded, imo the exemption can probably be loosened/have more nuance added Kowal2701 (talk, contribs) 17:27, 11 November 2026 (UTC)Reply

You know, maybe I am missing something but I don't have the impression that "expansive copyediting" is any more an issue than "light copyediting". OK, no more an issue than doing the same thing manually would be - we have had problems with copyediting causing SYNTH issues, it's just that these problems usually occur in the form of a thousand small edits rather than a spree that catches attention to the specific editors - so it's not quite accurate to single out LLMs as specifically problematic. The problems really start if you are creating large quantities of text (things like paragraphs or multiple sentences, and of course entire articles) by LLM. Viz IAR, I am a bit uneasy with there being a de facto exception that isn't mentioned. Jo-Jo Eumerus (talk) 17:49, 11 November 2026 (UTC)Reply

Yeah, we have had a couple large cases where people were just using LLMs to copyedit articles they hadn't authored (Wikipedia:AI noticeboard/FlutterDash344, Wikipedia:AI_noticeboard/Archive_3#Three_years_of_bad_AI_copyedits_by_User:Kofi_Meija), but most often it's people using an LLM to expansively copyedit their own work (WP:ESLAI touches on it), which (frankly) most often seems to make it worse (hard to tell because of availability bias/survivorship bias). Agreed having a de facto exception isn't ideal, hopefully it's just in the short term (PAGs usually lag behind anyway since they just document best practices). Some AI policies on other wikis have the lead be basically what we have now, and then a later section giving loads of caveats/criteria for responsible use (eg. fr:Wikipédia:Intelligence artificielle générative, m:AIP has some others), could do something like that for this + generating tables etc. (see Wikipedia:AI noticeboard/Tiny Particle) but idk Kowal2701 (talk, contribs) 18:05, 11 November 2026 (UTC)Reply