Wikipedia talk:Writing articles with large language models

I'm confused about this policy

On the Wikiproject AI cleanup page (Wikipedia:WikiProject AI Cleanup) it said that the goal of the project isn't to ban the use of AI in articles but to verify it's outputs match up with Wikipedia's other policies. But here, it says that you can't use them to write articles at all. So can you use AI to write Wikipedia articles as long as you check the article to make sure it's accurate? Electron230 (talk) 20:35, 22 March 2026 (UTC)

This policy is more recent than the creation of that Wikiproject. As of March 19 2026, Wikipedia's policy is what you see here: the use of LLMs to generate or rewrite article content is prohibited, save for very narrow exceptions for copy-edited and translations that comply with WP:LLMTRANSLATE. I would add that you can use LLMs in your research process, but everything added to an article must be carefully verified and written by you. ~ le 🌸 valyn (talk) 20:59, 22 March 2026 (UTC)

I still find it confusing that the 2 pages say different things. Electron230 (talk) 21:17, 22 March 2026 (UTC)

The WikiProject has updated its description to match the new guideline here. I hope that helps clarify. ~ le 🌸 valyn (talk) 05:07, 23 March 2026 (UTC)

So does this mean your article can be taken down simply because it's largely written by AI, even if its accurate, unbiased and follows most of Wikipedia's other guidelines? Electron230 (talk) 17:35, 23 March 2026 (UTC)

If it was created after this guideline passed, yes. We wouldn’t have passed this guideline if LLMs frequently produced accurate, unbiased, policy-compliant articles. ~ le 🌸 valyn (talk) 17:56, 23 March 2026 (UTC)

What if you checked and edited it before publishing? Electron230 (talk) 19:59, 23 March 2026 (UTC)

The page Wikipedia:Writing articles with large language models, which we are discussing, states the community's expectations, which are the use of LLMs to generate or rewrite article content is prohibited, save for the exceptions given below. Generating an article and then checking and editing it is not one of the exceptions listed, so no, it is not permitted. If you are curious about why the community has settled on that expectation, I encourage you to read the lengthy discussions in the archives of this talk page. -- LWG ^{talk (VOPOV)} 20:40, 23 March 2026 (UTC)

But what about Wikipedia:Ignore all rules? Electron230 (talk) 21:05, 6 June 2026 (UTC)

If you can convince the community to allow for an exception, yes. For example, Wikipedia:Articles for deletion/Uhr closed with unanimous consensus to keep an LLM-generated disambiguation page (which is technically not an article). Given the recent attitudes in the community about LLM-generated content, however, support for such an exception is technically possible, albeit very unlikely. SuperPianoMan9167 (talk) 22:23, 6 June 2026 (UTC)

What about it? Invoking WP:IAR is a cop-out. I've been here almost 20 years, most of that as an administrator, and I haven't had to resort to that more than once in my memory, and that one time was overridden upon review anyway. You don't invoke IAR for personal convenience. You resort to it in extenuating unique circumstances. Being unable to function on Wikipedia without using LLM-generated output cannot conceivably qualify as one of those circumstances. ~Anachronist (who / me) (talk) 23:21, 6 June 2026 (UTC)

According to a news source I read, AI companies are paying the Wikimedia Foundation to train their AIs on Wikipedia articles. Therefore, it would be stupid to allow AI-generated content to go into Wikipedia articles, because AI models being trained on AI output leads to a death spiral into uselessness. Prohibiting AI-generated content is necessary to ensure that AI models continue to improve. ~Anachronist (who / me) (talk) 02:57, 24 March 2026 (UTC)

Actually I think there is a flaw in that argument. In a proper LLM-assisted editing session, the human editor asks the LLM questions, checks sources and/or applies judgement as necessary (which might involve multiple rounds of asking for clarifications), writes drafts of paragraphs and presents them to the LLM, rejects some LLM suggestions (possibly telling the LLM why), and incorporates others that they agree in terms of both content and presentation. This back-and-forth can arguably result in more human-generated content going into the LLM's system than a purely human-based edit. Bbbbbbbbba (talk) 02:30, 24 July 2026 (UTC)

There is no flaw in the argument. What you describe is how LLM should be used to write a Wikipedia article. In what you describe, LLM isn't being used as the primary author or even a co-author, LLM is being used as an assistant, similar to the way a Supreme Court judge has assistants to do legal research and write briefs. In the end the judge writes the opinion, not the assistants. Nothing wrong with that. However, what we're seeing on Wikipedia isn't AI-assisted research or AI assistance in reviewing human-generated prose, but purely AI-generated content being posted to Wikipedia. The lazy editor simply prompts the AI "write a Wikipedia article on subject X" and pastes it in. Therein lies the stupidity in having LLM output on Wikipedia being used to train LLM models. ~Anachronist (who / me) (talk) 03:25, 24 July 2026 (UTC)

The condensed explanation for the policy is that LLMs can “change the meaning of the text such that it is not supported by the sources cited.”

My experience is that humans try to do this all the time on Wikipedia. My concern is that Wikipedia does not have adequate tools and processes to prevent edits that change the meaning of the text such that it is not supported by the sources cited.

Already LLMs are sometimes able to generate text that is indistinguishable from text written by humans. If the tools and processes work, we should have nothing to worry about edits that contradict the meaning of sources, regardless of how text was formed before added to Wikipedia.

Can we trust the tools and processes or not? Johnfromberkeley (talk) 15:10, 27 March 2026 (UTC)

The tools and processes have worked decently well over the last 25 years, but they only work if editors keep volunteering their time to make them work. A big part of why LLM text was banned is that it was consuming an unreasonable amount of volunteer hours compared to the benefit it provided. -- LWG ^{talk (VOPOV)} 16:14, 27 March 2026 (UTC)

Also if it was created before NOLLM, that just means it can't lead to sanctions for the user but LLM-generated text is not "grandfathered" in. ‑‑gurkubondinn 09:55, 25 August 2026 (UTC)

Yes, because we don't want AI content here. It's very simple: don't use AI to write new articles. Period. This is a project by people, for people, and is not the place to test ChatGPT prompts. Cremastra (talk · contribs) 16:05, 27 March 2026 (UTC)

Is that accurate? There is no mention of that rationale in the policy statement.

If that’s accurate, that rationale should be added to the policy statement. Johnfromberkeley (talk) 16:36, 27 March 2026 (UTC)

What's accurate is this reporting, confirming that AI companies are paying WMF to train their AIs on Wikipedia (and all other Wikimedia projects). Therefore, as I said above, it's stupid to allow AI content to creep into articles, lest the AIs start training themselves on their own slop. That's the rationale that should be added to the policy. ~Anachronist (who / me) (talk) 02:28, 28 March 2026 (UTC)

It’s becoming more and more apparent that Wikimedia’s tools and processes are not adequate for evaluating the quality of text submitted to Wikipedia based on the characteristics of the text alone.

That’s not a complaint, praise or criticism. Just an observation.

Quality text written by LLMs that is grounded in citations will “pass” undetected, and hopefully subpar text written by humans and or machines will be corrected. Johnfromberkeley (talk) 06:38, 28 March 2026 (UTC)

Oh, certainly. I've used LLMs to generate text in which individual sentences are excellent summaries of the cited sources. Where an LLM breaks down is from citing paywalled content, or when several sources all say similar things, or several sources provide trivial mentions. In those cases, the LLM resorts to vapid assertions like "X has received coverage in major media outlets", merely saying that coverage exists without describing the actual coverage. Nowadays that's the primary indicator I use to determine whether text is LLM generated. ~Anachronist (who / me) (talk) 06:52, 28 March 2026 (UTC)

Also, do you think there are many Wikipedians getting away with using AI in a way that’s against this guideline, but we simply can’t tell? Electron230 (talk) 03:17, 10 June 2026 (UTC)

Yes. As an experiment, I asked an LLM to write about some subject as if it was a Wikipedia article. I criticized its sourcing first, and it gave me a rewrite. I criticized any synthesis I found, and unsourced assertions, and it gave me a rewrite. I criticized the AI formatting "tells" and it gave me a rewrite. And so on.

Going several rounds like this, you can end up with an acceptable article that's 100% LLM-generated but you never wrote a word of it, simply by asking the AI to rewrite things to remove the WP:AISIGNS. It takes a while, but the only involvement you had was to check the AI's work and demand corrections, until it's impossible for you to detect signs of AI writing. ~Anachronist (who / me) (talk) 15:41, 10 June 2026 (UTC)

That's why I try to not point out every single AI sign or error I find, so there's a way for me to check if an editor actually did a rewrite/fix. InfernoHues (talk) 17:05, 10 June 2026 (UTC)

that's another thing I've also realized. Suppose the very problem that the list of AI signs are literally public. A bad actor could simply tell it to look up the page of AI signs and then quite literally avoid all the AI signs, what would cleanup people do then? (from what I've heard.) it's already arduous and annoying to prove someone has used it enough to take action PlutoidRepublic (talk) 17:42, 2 August 2026 (UTC)

Well luckily, LLMs can't follow instructions deterministically (hence why so many will happily write an article that "follows all Wikipedia policies and guidelines" even though such an instruction is contradictory), so it's likely some AI signs will still slip through. Though I'm sure pointing the LLM to AISIGNS would help it mask to some extent. I wonder if there's a way to block AISIGNS in Wikipedia's robots.txt, and whether that would do anything. Justin Kunimune (talk) 11:32, 3 August 2026 (UTC)

Yes, there is. I have added {{NOINDEX}} to that page. Not all crawlers respect robots.txt, though. ~Anachronist (who / me) (talk) 14:59, 3 August 2026 (UTC)

From what I read, that's not even how no index is supposed to work. And even if it was somehow excluded in every crawler on Earth completely complied. The user could just paste in the full wikitext or just Hand the model the direct url, I fear fellow editors aren't aware of what is available in models and are trying silver bullets without checking if they work PlutoidRepublic (talk) 15:36, 3 August 2026 (UTC)

Not all crawlers respect robots.txt, though. If there are any kinds of crawlers known for not respecting robots.txt, it's crawlers gathering AI training data. Also, I think that removing the page from indexing will remove a lot of the page's usefulness in exchange for very little benefit, as there are many non-Wikipedians that use it to help detect AI writing, and removing the page from indexing very likely won't stop it from being included in model training data. SuperPianoMan9167 (talk) 19:50, 3 August 2026 (UTC)

Yeah, that explanation makes me wonder why the editor even NOINDEX'd it. PlutoidRepublic (talk) 21:48, 3 August 2026 (UTC)

You can't block it that way. Training is different. You can just instruct it in a prompt to *look it up* and it can do that from any website in the world except obvious exclusions like adult content for no reason, I can just say "hey go read this website and check it" and just paste it a link. I would strongly recommend testing it out for yourself instead of relying on others explanations or attestations PlutoidRepublic (talk) 15:12, 3 August 2026 (UTC)

According to the docs, NOINDEX "may remove the page from some external search engines". It's better than nothing but more likely to be equivalent to nothing for AI purposes. Frequently, however, I ask an AI to access a page (like on my own blog) and it responds that it cannot access it, so I have to paste in the text to get answers to my queries that are relevant. I don't know how well any AI can access pages in the Wikipedia namespace. ~Anachronist (who / me) (talk) 23:31, 3 August 2026 (UTC)

From my testing, it can access it very well. I have seen a few examples of absurdly strong robots.txt making it very tough, though; for example, the BBC sets up its robots.txt PlutoidRepublic (talk) 13:00, 6 August 2026 (UTC)

In my experience LLMs make the same mistake humans do in interpreting AISIGNS, which is that they extrapolate too much and pretend Y is a sign because it sort of seems like X on some arbitrary axis. Gnomingstuff (talk) 17:05, 12 August 2026 (UTC)

That can also happen, it is more human now, where it can overcorrect just as much as under-correct, but prompts matter, if the prompt is crap, you get crap, I've always seen it as garbage in, garbage out PlutoidRepublic (talk) 02:59, 17 August 2026 (UTC)

Regarding the first rule...

The first rule says: "Editors are permitted to use LLMs to suggest corrections to their own writing, and to incorporate them after human review. This is limited to spelling, punctuation, capitalisation, grammar, and other simple mistakes. Editors must ensure that any text copyedited in this manner is still supported by the sources cited."

I think the spelling, punctuation, capitalisation, grammar, and other simple mistakes should be elaborated, by that I mean accompanied by a clear example. I forgot what page it is but I once saw someone laying this out using a before-and-after example, explaining that that is the only acceptable use of LLM outside of translation (second rule). The before has typos and the after one shows the corrected text and it's shown like when you open a diff page. But I'm 100% certain it's not a history page, it's in a discussion page like this.

The reason for this is just to make that rule much clearer, which greatly reduces the possibility of editors interpreting it another way in their heads. I haven't looked into the discussion history of this, so apologies if this is a repeated suggestion. AdaCiccone (talk) 07:14, 21 August 2026 (UTC)

We had a long discussion about this in § Refining basic copy editing that resulted in the current version. I suggest adding some examples in Wikipedia:LLMRESP or asking about a specific example. Dw31415 (talk) 11:57, 21 August 2026 (UTC)

Had no idea of that page, thank you. Anyway, I looked everywhere for that specific diff-like example but couldn't find it, but I found how to generate that here in this talk, so I made one myself:

−

yestreday me and my freind was going to the stroe because we need some of milk, but we forget are wallets at home.

+

Yesterday, my friend and I were going to the store because we needed some milk, but we forgot our wallets at home.

This is the example that I was talking about. Maybe we can put this in WP:LLMRESP? And then put an efn tag in WP:NOLLM that redirects to this example under "Copyediting". That is if there's clear consensus that this is what the rule means by 'basic copyediting'. AdaCiccone (talk) 04:05, 23 August 2026 (UTC)

A few hypothetical examples:

Original: John Doe was author and craetor of lots of screenplays movies and TV shows and he won 3 Oscars.

Acceptable: John Doe was the author and creator of lots of screenplays, movies, and TV shows and he won 3 Oscars. (Inserts missing article, fixes typo, fixes commas in list, fixes comma splice)

Technically breaks the rule, but most people would let it slide: John Doe authored several screenplays, movies, and television productions, winning three Academy Awards. (Style edits and revising informal tone in AI-typical ways, some of which are improvements and some of which are not)

Over the line: John Doe's career spanned a diverse array of screenplays, movies, and TV shows. He won three Academy Awards, reflecting his acclaim in the industry. (injection of typical AI slop) Gnomingstuff (talk) 14:00, 21 August 2026 (UTC)

could stick that in an {{efn}} Kowal2701 (talk, contribs) 15:00, 21 August 2026 (UTC)

kind of hesitant about encouraging people to break the rule; #2 is not necessarily a good copyedit (for one, AI loves the word "authored" over "wrote," and the sentence also subtly changes the meaning since it implies he was the screenwriter on the movies/TV, which is not in the original text) Gnomingstuff (talk) 18:15, 21 August 2026 (UTC)

more examples (this time working backward from an AI sentence because it is actually really hard to write genuine AI prose from scratch):

Original: She founded PR company BlahBlahBlah located Manchester. The agency represent clients in entertainment, sport, business, and podcasting fields She also works as media representative and crisis commnication consultant.

Acceptable: She founded the PR company BlahBlahBlah, located in Manchester. The agency represents clients in the entertainment, sport, business, and podcasting fields. She also works as a media representative and crisis communication consultant. (Missing article, missing comma, missing word, subject/verb agreement, missing period, typo, word usage)

Not acceptable: She established BlahBlahBlah, a Manchester-based PR and talent management agency, representing clients across entertainment, social media, business, and podcasting. She has worked with high-profile figures in entertainment and sport, providing strategic media representation and crisis communications support. Gnomingstuff (talk) 18:36, 21 August 2026 (UTC)

yeah that looks good, can use {{xt}} for 'Acceptable' and {{!xt}} for 'Not acceptable' (like the MOS does) and put the changes in bold Kowal2701 (talk, contribs) 18:46, 21 August 2026 (UTC)

LLM Paste Check - wording for the default translations of the UI

 You are invited to join the discussion at meta:Talk:Artificial intelligence/Policies by project § LLM Paste Check - Question about the default wording to show in the global feature. Briefly: We'd appreciate any multi-lingual/multi-wiki perspectives on what UI text to use, in this upcoming part of the Edit Check feature. The feature can (based on local config) show a message to any editors that copy & paste text from a few commonly used LLMs directly into an article (within VE), after they've done that. The local versions of the UI's text/link can (and should) be overridden for more specific local nuances/desires, but we need an ideal global baseline version for other wikis that are using the basic/fallback version. Thanks. Quiddity (WMF) (talk) 21:33, 21 August 2026 (UTC)

Discussion at Wikipedia:Village pump (policy) § RfC: AI use for generating citations within articles

 You are invited to join the discussion at Wikipedia:Village pump (policy) § RfC: AI use for generating citations within articles. —Myceteae‍🍄‍🟫 (talk) 17:31, 28 August 2026 (UTC)

Proposal to add an additional exemption to this guideline. —Myceteae‍🍄‍🟫 (talk) 17:33, 28 August 2026 (UTC)

Eye needed

Possible LLM seasoning at Tiger Woods and other pages also? Feels a lil too well-seasoned. Lots of pages getting very detailed accounts, reads like a news diary. One edit summary even says "mis-paste". NicoR8 (talk) 12:05, 3 September 2026 (UTC)

I see significant problems, but can't take care of this right now. Compare the language of September 2023: "I do not want to repeat the mistakes that was made" or February 2026: "Rules also says no revert can be done for 24 hours" with recent polished wording in both articles and edit summaries: 1, 2, 3. Then there is the rate of changes, often only a few minutes apart from article to article, and the... distinctive range of topic matter being edited. I would suggest raising this at an admin noticeboard. Dekimasuよ! 12:51, 3 September 2026 (UTC)

Hey @Dekimasu on their talk page, they have said they have used AI and also says "The volume was more than I could check properly"! I think, at minimum, their edits should be reverted. NicoR8 (talk) 11:43, 4 September 2026 (UTC)

I still won't be able to deal with this volume anytime soon (probably 150-200 problematic edits, many of which are not the latest revisions), so given that the problem has already been acknowledged by the editor, it may be easier to take this to Wikipedia:AI noticeboard for cleanup. Best, Dekimasuよ! 05:46, 5 September 2026 (UTC)

The start date of my edits in question is August 27. 200 mainspace edits across 84 articles or pages.

Roughly 81 of those edits are cosmetic changes - broken ref names, category additions, hyphen replacement, "," instead of "'" etc. I am willing to spend time to identify which is which and give in broad categories.

Only 2 pages created - Perversion of Justice and 2026 Maldives speedboat capsizing. I checked the sources and citations in those pages and they are all valid. No hallucinations or made up sources. Perversion of Justice is a redirect.

The drafting problem from my end has stopped. Voicecalls (talk) 17:34, 5 September 2026 (UTC)

Proposed reword

Looking over this again to make some small edits, and I really think we should just drop the opening clause, like so;

−

Text generated by large language models (LLMs)[a] often violates several of Wikipedia's core content policies. For this reason, the use of LLMs to generate or rewrite article content is prohibited...

+

The use of LLMs to generate or rewrite article content is prohibited...

It doesn't really have any benefit, because it doesn't modify the actual prohibition in any way, nor clarify its scope, it merely justifies it, which isn't really necessary to do on the guideline page itself. Conversely, it has historically resulted in confusion (see WP:AINOVIOLATION which I wrote specifically in response to such confusion)

I'm not going to do this BOLDly because it's kind of a big change, so consider this an informal quasi-RfC on the matter. Athanelar (talk) 12:29, 7 September 2026 (UTC)

I like the clarity. Maybe move the phrase to its own sentence in a second paragraph. Dw31415 (talk) 13:39, 7 September 2026 (UTC)

I don't think the text should be removed from the page, rules that people understand the purpose of are more likely to be complied with. That said, I don't know that it needs to be part of the rule itself, so making it a second paragraph as DW suggests is not a bad idea. Thryduulf (talk) 15:06, 7 September 2026 (UTC)

Yes, drop the preamble. That's been the direction of travel on WP:MOS and its subpages for years. Too many editors argue that a prohibition based on a preambling justification doesn't apply to them because their edits are different. Meanwhile, ideal editors, the ones that read guidelines before editing, shouldn't be made to slog through a preamble and get distracted thinking how that relates to them before they get to the message that foo's prohibited anyway. NebY (talk) 20:13, 7 September 2026 (UTC)

I also wonder to what extent the community actually even agrees that this is the reason LLMs are banned these days. I think the much more common understanding currently is that we ban LLMs due to the difference between the effort required in creating massive amounts of LLM-generated text and the effort required in verifying and proofreading all of that text. It's not necessarily its tendency to break core content policies so much as the risk that any amount of such mistakes slipping through the cracks would be compounded exponentially by the sheer volume of LLM-generated text that would be flooding Wikipedia if we didn't actively work to prevent it. Athanelar (talk) 22:51, 7 September 2026 (UTC)

Given the agreement so far I've gone ahead and moved the information that was in this preamble to the second paragraph, which now opens LLM-generated text often violates several of Wikipedia's core content policies, and LLMs allow such text to be generated at a scale that makes human verification of factual claims, source-to-text discrepancies, style issues, etc., unrealistic. Persistent use of large language models outside of the above exceptions is considered disruptive and will result in an editor being blocked from editing. Hope everybody finds this agreeable. Athanelar (talk) 22:59, 7 September 2026 (UTC)

I see you also boldly add this text.

Any use of LLMs which is argued to be within the exceptions below is still subject to sensible judgement by other editors. LLM use which consensus finds to be disruptive can still be prohibited under this guideline even if it can be argued to fit within the exceptions; they are not "get out of jail free" cards.

I reverted those two sentences until there is some discussion on them. Mostly the sentences have a very different tone than other guidelines I'm familiar with. The edits being "subject to sensible judgement" isn't very clear as an additional issue. Maybe we can hash out a better version here. I'll revist in the next couple of days myself, but hoping others some word-smithing suggestions. Dw31415 (talk) 02:23, 8 September 2026 (UTC)

I do think it's an important thing to clarify to stop a lot of the "but I only used it to copyedit!" arguments. Editors should know right away that LLM use can still be disruptive even if it ostensibly falls within one of the exceptions. Athanelar (talk) 09:02, 8 September 2026 (UTC)

I don't think this is a necessary change tbh, it doesn't really cause any problems. It was worded like this because the idea was to base NOLLM on AI-generated content violating PAGs, and have another discussion or RfC down the line about whether to ban it regardless of PAG vios, but practically no-one who !voted in the RfC viewed it like that so idm changing it if people want to.

The addition of LLM-generated text often violates several of Wikipedia's core content policies, and LLMs allow such text to be generated at a scale that makes human verification of factual claims, source-to-text discrepancies, style issues, etc., unrealistic. doesn't make sense, "factual claims" and "source-to-text discrepancies" are effectively the same thing, and "human verification of style issues" doesn't work. This is more the reasoning for WP:LLMPRV, which I suppose we should add a sentence about. I'll have a go at rewording Kowal2701 (talk, contribs) 10:17, 8 September 2026 (UTC)

How does that look? (notified WP:VPP) Kowal2701 (talk, contribs) 10:33, 8 September 2026 (UTC)

Nice. Well done. Dw31415 (talk) 23:18, 9 September 2026 (UTC)

Weighing to say that slight ambiguity is good actually. While Athanelar has an essay saying otherwise, we have many users (HTGS, Rusalkii, Myceteae, etc.) in the ongoing VPP discussion on AI and citations arguing (in SarekOfVulcan's words): "If you check it thoroughly enough so that you're sure there's no error, you're not going to get flagged, and if you don't, you should be." In short, the message we should be sending is "don't mess up the encyclopedia with your unverified content" rather than "we will hunt out your LLM usage to the maximum extent." The prior wording was more consistent with this.--Carwil (talk) 15:14, 9 September 2026 (UTC)

That message would fundamentally change the guideline, so start an RfC for it if you think the guideline should be changed. Gnomingstuff (talk) 16:53, 9 September 2026 (UTC)