Wikipedia talk:Writing articles with large language models

In what ways are LLMs useful?

This policy says that they can be useful tools but i cant seem to think of a reason they would be useful for Wikipedia it should either list useful reasons for use on Wikipedia or the claim be removed entirely of course they are not useful for creating articles and that part should remain in the policy Isla🏳️‍⚧ 23:25, 17 February 2026 (UTC)

One application I've found to be useful is to drop the PDF of a source I just read into the LLM with the prompt "please generate a Wikipedia-format citation for this document. If any bibliographic information is missing, search the web to find it, and provide me with links to where you found the missing information so I can verify it is correct". -- LWG ^{talk (VOPOV)} 23:48, 17 February 2026 (UTC)

Just to confirm, this usage is now banned under the new guideline, right? I also regularly use LLMs to format bibliographic metadata into {{cite}} templates; Claude Sonnet has substantially more coverage and is more reliable than our existing Zotero/Citoid integrations. I guess maybe formatting citations is considered "basic copyediting", since it's about readability and layout? Suriname0 (talk) 23:35, 26 March 2026 (UTC)

Another use case is what I described below - using an LLM to convert all the dates in a table to ISO format to make them sort correctly (and of course manually reviewing the result before saving). If I had had to do that by hand, I would surely have introduced errors (or more realistically, might not have bothered since it's so tedious).

I think routine technical uses like format conversions or wikitext help, like your citation formatting example, should be covered under the copyediting exception and should be explicitly added to the guideline, with the caveat that a human needs to review the output and is responsible for the edit. I also think the wording of the copyediting exception should probably be changed to remove "to their own writing" - it doesn't matter whether someone uses an LLM to format a {{cite}} template for a new citation they are adding or for a bare URL someone else added to the article. Helpful Cat🐈(talk) 02:15, 27 March 2026 (UTC)

I would absolutely put that under copyediting, as you're not adding new content but just formatting it properly. Chaotic Enby (talk · contribs) 21:33, 27 March 2026 (UTC)

You could also give article content to an LLM and have it check for errors. Pmsyyz (talk) 02:37, 23 March 2026 (UTC)

Wikipedia:Responsibly using large language models Moxy🍁 23:42, 26 March 2026 (UTC)

I've used LLMs to ask quick grammar questions, ask research questions (e.g. potential sources or which library to use), get suggestions for articles or uncovered content (ideas, not the content itself), perform tedious technical changes that don't add or subtract content (e.g. fixing the ordering of a big table), and in working on technical things, generate example code (e.g. SQL, JavaScript, Lua) for particular uses I can't get my head around. Of course, I carefully review/massage everything that is generated before any use here. the Stefen 𝕋ower 09:09, 8 April 2026 (UTC)

Reverted removal of nutshell

@EEng, I just saw that you reverted my removal of the "in a nutshell" box. Your edit summary of Just relax does not really help me understand your reasoning. Could you explain? Rootvegetablecurator (talk) 05:45, 27 February 2026 (UTC)

It means your edit summary betrayed excessive worry about form or balance or something, over punch. EEng 06:22, 27 February 2026 (UTC)

I think a guideline/policy page is one of the worst spots to introduce unnecessary clutter, especially if the goal is to actually have editors read it. I don't think "punch" should be a big concern in a two-sentence long guideline - it's not an opinion piece, it's not trying to convince anyone of anything, it's just telling you what you have to do. Rootvegetablecurator (talk) 16:26, 27 February 2026 (UTC)

I like that the warning is reinstated. It helps tech bros who are not that good at reading (that's why the want to use AI in the first place) obey.

--~2026-28930-14 (talk) 18:55, 13 May 2026 (UTC)

Discussion at Wikipedia:Writing articles with large language models/RfC

 You are invited to join the discussion at Wikipedia:Writing articles with large language models/RfC. Chaotic Enby (talk · contribs) 14:27, 15 March 2026 (UTC)

Footnote

@NicheSports: I included the footnote because most users of LLMs do not know what LLMs are. The same footnote is used at WP:G15. Do you have any suggestions for alternative wordings? SuperPianoMan9167 (talk) 14:30, 21 March 2026 (UTC)

My fault @SuperPianoMan9167. I misread the diff and should have first checked how it looked in the page. I think it is fine to keep, thank you for adding. NicheSports (talk) 14:34, 21 March 2026 (UTC)

Lower case redirect

Why doesn't WP:newllm redirect to this page? It seems unique amongst the policy/essay pages like WP:cite. I'm not used to typing in all caps, so this surprised me quite a bit when I landed on the empty page. LkL-70547 (talk) 16:04, 21 March 2026 (UTC)

I've now created it. Kovcszaln6 (talk) 16:25, 21 March 2026 (UTC)

Should WP:LLMWRITE redirect here instead? Kowal2701 (talk, contribs) 13:55, 22 March 2026 (UTC)

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)

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)

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)

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)

Extended-confirmed-protected edit request on 24 March 2026

~2026-18396-60 (talk) 19:38, 24 March 2026 (UTC)

 Not done: it's not clear what changes you want made. Please detail the specific changes in a "change X to Y" format and provide a reliable source if appropriate. This edit request is empty. SuperPianoMan9167 (talk) 19:46, 24 March 2026 (UTC)

Does this fall under an exception?

I have never used LLMs to generate article content or rewrite prose, nor do I personally use them to copyedit in the sense of grammar, spelling, etc. I did use them twice to convert all the dates in a table to ISO 8601 format, because the dates were not sorting correctly otherwise, as someone pointed out on talk. Per my edit summaries, I disclosed the use of an LLM, and manually reviewed every line of the diff to ensure the conversions were correct and nothing else was changed; nobody has objected to the changes, and subsequent editors have added new entries in ISO format.

(ISO dates are allowed in tables per MOS:DATE, and suggested as one of the solutions to the date sorting problem in Help:Sortable tables#Year, month, day. Using numbers. ISO date YYYY-MM-DD. I was not aware of Template:dts at the time, but if I had been aware, I could equally have asked AI to add {{Date table sorting}} to all the dates and then reviewed the wikitext myself.)

Under the new guideline, would this kind of purely technical edit fall under the copyediting exception? If not, that's a pity because it's unproblematic in my opinion as long as a human reviews the changes, and would have been extremely tedious to do by hand. I would probably have introduced more errors doing it myself than by having AI do it and then checking the AI's output. I suggest this type of technical edit should be explicitly allowed with a caveat that a human must review all changes and is responsible for them. Helpful Cat🐈(talk) 01:36, 26 March 2026 (UTC)

In my opinion this kind of mechanistic format change falls into the realm of copyediting, and is fine (though personally I probably would have used a more deterministic script to make the changes, just to ease proofreading). -- LWG ^{talk (VOPOV)} 02:03, 26 March 2026 (UTC)

Thanks! I don't know how to set up scripts, so AI was the easiest way. Currently, the copyediting exception says Editors are permitted to use LLMs to suggest basic copyedits to their own writing (emphasis mine), but the dates I changed were not my own writing as they had been added by previous editors. Should we change the wording of the guideline, so that it no longer limits copyediting to one's own writing? Helpful Cat🐈(talk) 02:14, 26 March 2026 (UTC)

That's a good point, "to their own writing" seems arbitrary and should probably be removed. It shouldn't matter whether the copyeditor is the original writer, since no one owns their writing on Wikipedia. Justin Kunimune (talk) 11:51, 26 March 2026 (UTC)

That addition makes sense as we don't want people to run ChatGPT/Grok/etc. to "rewrite" sentences to "improve flow", which I've already seen happen multiple times, never with an actual improvement. I interpreted "their own edits" as meaning "the edits they're currently working on", but I understand how that could be confusing. Similarly, I didn't see technical edits as falling under content generation, but it could be helpful to clarify that they're allowed if someone wants to propose such a wording clarification. Chaotic Enby (talk · contribs) 21:38, 27 March 2026 (UTC)

Thanks! I proposed some wording below. I think a clarification is useful to prevent a situation where some editors reject such edits as "rewriting article content" rather than copyediting. Helpful Cat🐈(talk) 10:03, 28 March 2026 (UTC)

IAR exists for stuff like this. Re "to their own writing", it’s to stop cases like this Kowal2701 (talk, contribs) 06:11, 27 March 2026 (UTC)

Going full IAR is not necessary. If the llm use is indistinguishable from a regex (which a human would also have to review, we get some bad find and replaces), then the use falls under WP:SEMIAUTOMATED. This does not affect the prose, to the point it is more basic than the basic copyediting exception already listed. Under SEMIAUTOMATED, there is only a huge issue if there is no disclosure or it is being done rapidly at scale. Fixing up one table in one article does not fall into this (even though it is more "rapid" than doing it manually). CMD (talk) 06:26, 27 March 2026 (UTC)

Yep, anything that doesn't affect the prose but just does regex-like formatting shouldn't count as "generating content" more than copyediting. Chaotic Enby (talk · contribs) 21:35, 27 March 2026 (UTC)

Proposal: exception for technical edits

There seems to be consensus that this type of technical edit is fine. What do people think about this wording for a third exception:

3. Editors are permitted to use LLMs for purely technical assistance that does not change the meaning of article content, including but not limited to help with wikitext. Editors are expected to review such edits before publishing, and remain responsible for any changes made.

I think this would cover my date format example, as well as the example other users have mentioned of formatting citation templates. Helpful Cat🐈(talk) 09:58, 28 March 2026 (UTC)

I would support that. Justin Kunimune (talk) 11:52, 28 March 2026 (UTC)

I'd say that's already covered under basic copyediting, as – for the purposes of provided the LLM does not introduce content of its own – I wouldn't consider this "content", but I'd support mentioning this as an example or clarifying the word "content". Kovcszaln6 (talk) 12:04, 28 March 2026 (UTC)

I would also fold this into exception 1, perhaps with a link to WP:SEMIAUTOMATED. CMD (talk) 12:17, 28 March 2026 (UTC)

Yep, I would also prefer this. Chaotic Enby (talk · contribs) 12:33, 28 March 2026 (UTC)

If people prefer to fold this into exception 1 instead, how about this rewording of exception 1?

1. Editors are permitted to use LLMs to suggest basic copyedits to their own prose, and to incorporate some of them after human review, provided the LLM does not introduce content of its own. Caution is required, because LLMs can go beyond what is asked of them and can change the meaning of the text such that it is not supported by the sources cited. Using LLMs for purely technical assistance that does not change the meaning of article content, including but not limited to help with wikitext, is also considered copyediting and is allowed. Editors are expected to review such edits before publishing, and remain responsible for any changes made. If editors use LLMs to assist with repetitive technical tasks, the guidelines for semi-automated editing apply.

In the first sentence, I changed "their own writing" to "their own prose". This is because people have pointed out above that we don't want editors trying to "improve" existing article text with LLMs. However, non-prose technical changes should be fine regardless of who originally added the content (e.g. converting dates added by other users, or formatting citation templates for bare URLs someone else added). Helpful Cat🐈(talk) 12:52, 28 March 2026 (UTC)

That definitely sounds like an improvement! My only worry is that some folks may use the "purely technical assistance" as a loophole, as it is less clearly defined than "basic copyediting", but we can clarify it later down the line if there is clear abuse. Chaotic Enby (talk · contribs) 12:58, 28 March 2026 (UTC)

My thoughts were along the line of changing "suggest basic copyedits to their own writing" to "suggest basic copyedits to their own writing or aid with repetitive tasks". This copies over the wording from SEMIAUTOMATED, and the link provides examples of what that wording means. CMD (talk) 13:00, 28 March 2026 (UTC)

Much more concise! Both are fine with me, although concision can often win. Chaotic Enby (talk · contribs) 13:05, 28 March 2026 (UTC)

Is that a bit too narrow? The use case others mentioned above of using LLMs to format citation templates isn't a repetitive task, for example. Helpful Cat🐈(talk) 13:06, 28 March 2026 (UTC)

Reformatting a large number of citation templates can feel pretty repetitive. There are situations where citation editing can be a bit trickier for a few reasons, but those are also situations where the LLM is going to be less helpful. I suspect trying to try and find where different configurations of citation formatting lay on that spectrum may be a fools errand. CMD (talk) 13:20, 28 March 2026 (UTC)

True. I was thinking that if someone is using an LLM to format and insert one citation at a time, or to help with wikitext for a single, non-repetitive edit, those are not repetitive tasks but should still be allowed. How about this concise but broader version: Editors are permitted to use LLMs to suggest basic copyedits to their own writing; to aid with repetitive tasks; and for purely technical assistance, such as help with wikitext, provided the LLM does not introduce content of its own. Editors should only publish such edits after human review. Caution is required, because LLMs can go beyond what is asked of them and can change the meaning of the text such that it is not supported by the sources cited. Helpful Cat🐈(talk) 13:26, 28 March 2026 (UTC)

I like this proposal. Concise, and "repetitive tasks" captures most of the relevant use cases. As additional uses become apparent, we can revisit the specific scope of "basic copyediting" and "repetitive tasks". Suriname0 (talk) 17:54, 28 March 2026 (UTC)

I don't think this should be an exception. LLMs suck at wikitext and templates' syntaxes, and clueless "technical editing" can easily break a few MOS guidelines (most commonly Wikipedia:Manual of Style/Accessibility). sapphaline (talk) 13:25, 28 March 2026 (UTC)

Why is translation allowed to English but not from English?

The article says:

> Editors are permitted to use LLMs to translate articles from another language's Wikipedia into the English Wikipedia,

But this is going to raise in many readers minds the question of why only that direction. This should be addressed in the page.

In non-culturally-inflected subjects such as math/physics/computing etc I think using machine translation to achieve higher quality corpora in non-English languages would be highly beneficial to non-English speakers. Myrmornis (talk) 16:04, 27 March 2026 (UTC)

From English is irrelevant for this guideline, because it implicitly means the new content is being published on other wikis, which cannot be bound by enwiki guidelines. Other-language Wikipedias likely have their own guidelines on translating text into their own languages with LLMs. Cremastra (talk · contribs) 16:07, 27 March 2026 (UTC)

Wikipedia:Large language model/Noticeboard

We may need to create a noticeboard: Wikipedia:Large language model/Noticeboard (WP:LLMN) where editors can deal with issues, get advice, and show their LLM work and ask if what they are doing is allowed.

We have Wikipedia:WikiProject AI Cleanup/Noticeboard, but I'm not sure that's the same thing.

We are well past the stage where denying or forbidding all LLM use is feasible or realistic. It's now all about figuring out constructive ways to use it as a "new technology" that is not going away, and when it should or should not be allowed, all while assuming good faith. Editors need a safe place to discuss this stuff. -- Valjean (talk) (PING me) 22:14, 27 March 2026 (UTC)

If we're going down a preventative path, then that sort of forum might help hammer out specific and detailed guides on responsible uses of LLMs, in line with policy-- this sort of institutional knowledge could then be linked to in future requests for advice at the noticeboard. Joko2468 (talk) 01:35, 28 March 2026 (UTC)

There are already ample forums for relevant discussions: the teahouse, help desk, WP:VPP/R/I, WP:AINB (aka WP:LLMN), WT:AIC, WT:AITOOLS, etc. I don't see any need to create another venue.

We are well past the stage where denying or forbidding all LLM use is feasible or realistic. It's now all about figuring out constructive ways to use it... – This policy doesn't forbid all use, and if it did, it wouldn't then follow that unless we can prevent all LLM use we must instead integrate it. "Can't beat 100% of it, join it" is not a logical basis for anything. fifteen thousand two hundred twenty four (talk) 03:02, 28 March 2026 (UTC)

@Fifteen thousand two hundred twenty four: I think you have misunderstood me. I am not saying that we are forbidding all use (we have two exceptions which I agree with), nor am I proposing "Can't beat 100% of it, join it" thinking. That's not what I'm saying. I am saying that LLMs are here (in the world) to stay, so we need to channel and control how they are used here at Wikipedia, not surrender. That is what we are doing, and we have two good exceptions. In that regard, I am not saying anything new.

My proposal of a noticeboard does not change any of the good proposals being made. I am just proposing a central place for discussion, problem solving, and dealing with disputes, just like other noticeboards. This just happens to be an important enough topic to deserve its own noticeboard. -- Valjean (talk) (PING me) 15:45, 28 March 2026 (UTC)

I agree that we needn't (and perhaps shouldn't) be recommending uses of LLMs-- we could explicitly state as such and that the focus of any such forum is policy compliance. The only real benefit to such a forum that I can see is to potentially develop guides that could give users the best chance of not violating policy. The more we understand about how editors are using LLMs the better (and this could potentially facilitate inhouse research that could support more assertive and informed advice on the prevalence of risk, since typical researchers are unlikely to treat LLM synthesis as an inherent problem). Ofcourse this requires community engagement, if there's little appetite for it then it's probably not a great idea. Joko2468 (talk) 08:46, 28 March 2026 (UTC)

If it isn't clear, I see LLM misuse as analogous to the war on drugs-- we should avoid a negligent or unenforceable approach. Joko2468 (talk) 12:22, 28 March 2026 (UTC)

Definitely, and essays such as Wikipedia:Responsibly using large language models play a major role in this, and should be prominently linked. Chaotic Enby (talk · contribs) 12:43, 28 March 2026 (UTC)

I understand where you're coming from (and giving editors advice regarding constructive use is helpful), but, like @Fifteen thousand two hundred twenty four points out, we already have places to ask about this. While a more specialized one could be helpful, a "noticeboard" might not be the appropriate way to go at it, as these usually carry the connotation of bringing up problems that have to be solved, rather than asking for advice. Maybe we can, for now, redirect them to the teahouse or help desk, and maybe down the line split it into a new page if either the volume of requests or the need for more specialized editor advice makes it necessary? Chaotic Enby (talk · contribs) 12:32, 28 March 2026 (UTC)

The teahouse should be a safe space, especially for newer editors (ie. those most likely to be unfamiliar with how LLMs are handled here). Given LLM use often overlaps with issues such as reliable sourcing, tone, plagiarism, notability, and other challenges often faced by newer editors, I wouldn't want to pull newer editors away to the teahouse to a specialised board with perhaps less friendly expectations. CMD (talk) 12:58, 28 March 2026 (UTC)

Yep, I'm thinking we could redirect them to the Teahouse, and maybe have a "welcome" warning template (maybe a softening of {{uw-ai1}} that genuinely assumes that the editor doesn't not know about the issue) that points them towards the Teahouse and useful essays/advice pages? Chaotic Enby (talk · contribs) 13:03, 28 March 2026 (UTC)

Perhaps there is a case to add a Teahouse link to all level-1 warnings, I would raise that somewhere. CMD (talk) 13:22, 28 March 2026 (UTC)

Done! Wikipedia talk:Template index/User talk namespace § Add teahouse links to level 1 templates? Chaotic Enby (talk · contribs) 13:42, 28 March 2026 (UTC)

(edited my comment, meant to write "does not" or "doesn't" and ended up with the cursed "doesn't not") Chaotic Enby (talk · contribs) 13:26, 28 March 2026 (UTC)

I think this could be a subpage of ANI, considering how many reports about AI misuse there have recently been (something like Wikipedia:Administrators' noticeboard/Incidents/LLM abuse). sapphaline (talk) 13:10, 28 March 2026 (UTC)

I remember floating a similar idea once (although it was about making WP:LLMN into this kind of page, rather than creating a standalone one). With how much the problem has grown since, it definitely sounds like it is needed too. We might want to be careful not to mix up a potential place to report AI misuse (last resort, involuntary, threat of sanctions) with a place for new editors to ask for feedback on their AI use (first resort, voluntary, threat of... being helped), as the atmospheres of both will be very different. Chaotic Enby (talk · contribs) 13:24, 28 March 2026 (UTC)

"We might want to be careful not to mix up a potential place to report AI misuse (last resort, involuntary, threat of sanctions) with a place for new editors to ask for feedback on their AI use" - that's why I called the subpage "LLM abuse". sapphaline (talk) 13:26, 28 March 2026 (UTC)

Of course, I was just clarifying how it was meaningfully different from what was suggested above! Both would be helpful in my opinion. Chaotic Enby (talk · contribs) 13:27, 28 March 2026 (UTC)

The FAQ

I feel the phrasing here needs work:

Current:

Q2: This guideline doesn't explain why it's necessary to be this strict. A2: Guidelines aren't information pages. We have plenty of information already about why using LLMs for content is a bad idea at WP:LLM.

First off, this is phrased like a statement, not a question. Is this S2 of a FAS or is it Q2 of a FAQ?

Secondly, this comes across as defensive. We don't have to be defensive about this stuff.

Also, this begs the question "why is the guideline based on non-binding essay-level information?"

WP:LLM is only an essay. Essays can argue for or against things without that thing necessarily becoming policy. At the very least, the guideline should endorse that essay (over others, theoretically there could be another essay "why we should use LLMs to write Wikipedia) in the guideline itself. A FAQ on a talk page isn't part of the guideline.

Also, please add a Q3 along the lines of

Q3: Why doesn't this guideline explore the downsides of being this strict?

I am chiefly thinking about arguments (that I'm sure have been aired) such as "why be this strict when many editors will use LLms anyway", "why be this strict when plenty of LLM generated additions will slip through the cracks", and/or "why be this strict when LLms will get gud soon enough".

When a guideline enforces a clearly controversial decision like here it should ideally make it clear which disadvantages have been considered and still found not persuasive enough to overturn the guideline.

Finally, let me clear say I approve of the policy and I do not wish to change it. This is about the logic gap of telling users guidelines don't have to justify themselves, but then justifying them anyway - but through thoroughly inofficial content like random essays. CapnZapp (talk) 13:40, 28 March 2026 (UTC)

Shortcuts

Should WP:LLM and/or WP:LLMPOLICY be retargeted here? The first one is more natural (as P&Gs are more relevant than essays), although I already added a hatnote for disambiguation and it might be more broadly linked. The second one is an early draft proposal for an LLM policy, and, given how editors (especially newcomers) might use "policy" and "guideline" interchangeably, it could be helpful to retarget it either here or towards the same target as WP:AIPOLICY. Chaotic Enby (talk · contribs) 19:35, 28 March 2026 (UTC)

Yes retargeting WP:LLM here would be helpful and reduce confusion. I've already (since WP:NEWLLM was updated) encountered an editor invoking WP:LLM to argue that using AI to condense article text is an accepted use. There is a separate question about what to do with the current WP:LLM page - I think most (or even all) of it should be deleted as it overlaps with several PAGs - not just NEWLLM but also see WP:LLMCOMM and WP:LLMCHAT which could either be deleted or retargeted to WP:AITALK (I also wonder if WP:CHATGPT and WP:HATGPT should be deleted as LLM provider specific). If someone wants to advance this overall effort that would be very helpful. NicheSports (talk) 19:50, 28 March 2026 (UTC)

A quick search reveals 2200+ uses of WP:LLM in the wild. I don't think it's a good idea to retroactively change the meaning of so many messages at once, WP:LLM contains different content than WP:NEWLLM and I've personally linked other users to it for that differing content specifically.

There's also the issue of accuracy, WP:LLM is a very general shortcut better suited for broad coverage, such as what the essay attempts to do. More specific targets should have more specific shortcuts like with WP:AITALK, WP:LLMTRANSLATE or WP:NEWLLM.

This would be better handled with a hatnote (exists, but could be more descriptive) and integrating a guideline and policy overview into the essay (doesn't exist). fifteen thousand two hundred twenty four (talk) 21:01, 28 March 2026 (UTC)

The best option is probably to move this whole page to Wikipedia:Large language models (making WP:LLM the target). Not really convinced that the existing links are a problem -- there are probably thousands of references to policy pages that changed since the time of the comment. Gnomingstuff (talk) 17:59, 31 March 2026 (UTC)

Then where will the essay be? SuperPianoMan9167 (talk) 18:09, 31 March 2026 (UTC)

maybe tack a (historical) onto it, make it a subpage, idk, lots of options Gnomingstuff (talk) 21:35, 31 March 2026 (UTC)

Merging in parts of Wikipedia:Large language models?

Since WP:LLM no longer leads to the page explaining why LLMs are bad for writing content, we should probably merge in sections from said info page to justify the guideline. A guideline is not just a rule to follow without any explanation as to why it should be followed. SuperPianoMan9167 (talk) 20:19, 28 March 2026 (UTC)

Guideline worded as a policy

Hello

I come from frwiki and your recent update of Wikipedia:Writing articles with large language models led us to probably strengthen our rules.

I noticed that you 'prohibit' the use of LLMs, but this is only a content guideline, not a content policy, so even if you use the word 'prohibited', ' exceptions may apply' (sic, in {{guideline}}).

Shouldn't it rather be a 'policy'?

I know it's being nitpicky, but a large part of problems is due to that, and to rules that aren't specific enough, leaving room for interpretation, which is never a good thing when it comes to applying them.

Kind regards,  Şÿℵדαχ₮ɘɼɾ๏ʁ 19:02, 31 March 2026 (UTC)

Technically, even policies have exceptions. SuperPianoMan9167 (talk) 19:11, 31 March 2026 (UTC)

The short version is that reaching a consensus on AI-generated content was somewhat contentious and required multiple RfCs to arrive at the current guideline:

• The November 2025 RfC that resulted in a much narrower guideline than this
• Attempted Dec 2025 RfC to make that guideline stricter
• Subsequent RfC in 2026, resulting in this guideline

Gnomingstuff (talk) 21:39, 31 March 2026 (UTC)

You may also be interested in de:Wikipedia:Künstliche Intelligenz, which came before en.wiki's guideline and differs slightly in its detail. CMD (talk) 03:14, 1 April 2026 (UTC)

Agreed. The implementation of the RfC is pretty sloppy imo and leaves a lot of stuff vague, the guideline just isn't detailed enough to be ready for implementation as an actual guideline. The guideline gives a bare minimum that it "often violates policies" and doesn't really explain why this is or why the issues can't be mitigated by responsible usage (e.g. AutoWikiBrowser?), which makes it a weak guideline on a site where there are no real rules and action is primarily decided by actual discussions. The guideline also overstates I think, a guideline nor even a policy can't entirely prohibit something in most cases and it will probably be quite interesting how the guideline is applied in practice. Respectfully dissenting, Coleisforeditor (talk) 01:57, 3 April 2026 (UTC)

@Coleisforeditor : AWB doesn't use AI and cannot use AI, so it is not part of the problem.

The only thing I can see is help with creating regexes, but that has nothing to do with creating content.  Şÿℵדαχ₮ɘɼɾ๏ʁ 10:32, 3 April 2026 (UTC)

Of course, but it sets the precedent that semi-automated editing is okay given responsible usage and - in the case of AWB - after permission is granted. Respectfully dissenting, Coleisforeditor (talk) 13:36, 3 April 2026 (UTC)

You are comparing things that are not comparable, and that are not related to my initial question anyway.  Şÿℵדαχ₮ɘɼɾ๏ʁ 13:49, 3 April 2026 (UTC)

Your question was in regards to that the text of the guideline expresses that it is prohibited to use LLMs but it is a guideline which contradicts the language (per WP:GUIDES guidelines are best-practices that should be generally followed with common sense and occasional exceptions whereas the text asserts that it is an absolute prohibitory rule). Your question also expresses doubt that the guideline is specific enough and leaves too much to interpretation (a large part of problems is due to that, and to rules that aren't specific enough, leaving room for interpretation). I merely intended to add to these concerns, I apologise if it came off as changing the subject. Respectfully dissenting, Coleisforeditor (talk) 13:57, 3 April 2026 (UTC)

There are no exceptions for certain parts of the external links guideline. SuperPianoMan9167 (talk) 15:27, 9 April 2026 (UTC)

The policy based restriction is just an uncontroversial application of WP:COPYVIO, a legal policy, and is also a restatement of WP:LINKVIO which is a policy. I would be fine with the guideline if it were more specific and explained how it often violates content policies and how that cannot be mitigated by responsible usage and disclosure. The guideline doesn't really say what "generate or rewrite article content" means, it states that using it to review and do basic copyediting on your own work provided you then review its edits yourself is fine but doesn't then define what actually falls under this restriction. Would it be a violation of this guideline to use AI to generate a draft of an edit and then heavily improve that draft (not that I do that, I find AI to be very difficult to work with, but different people have different workflows) which would then make it compliant with the other policies and guidelines? Respectfully dissenting, Coleisforeditor (talk) 11:59, 10 April 2026 (UTC)

Would it be a violation of this guideline to use AI to generate a draft of an edit and then heavily improve that draft (not that I do that, I find AI to be very difficult to work with, but different people have different workflows) which would then make it compliant with the other policies and guidelines? Yes. The only allowed uses are copyediting your own writing (not generating new content) and translation. Any other use of LLMs in articles is not allowed. Unless, of course, you can successfully convince the community that an exception should be made. SuperPianoMan9167 (talk) 12:10, 10 April 2026 (UTC)

Except the guideline doesn't actually say that. That's your personal interpretation of it but the guideline just reads the use of LLMs to generate or rewrite article content. This is bound to create grey areas in application, and, while policies and guidelines just document the existing consensus and do not usually themselves create binding rules, the guideline would be much more useful if thought was given to why AI-generated content often violates content policies. This would give editors forming a consensus in these grey areas more to go off for handling them than their own subjective opinions on edit quality, human authorship, etc. The guideline as it reads now is more like something you'd see in a rulebook than a Wikipedia guideline or policy as it just creates an expectation and states two exceptions without explaining it. Respectfully dissenting, Coleisforeditor (talk) 13:15, 10 April 2026 (UTC)

The guideline as it reads now is more like something you'd see in a rulebook than a Wikipedia guideline or policy as it just creates an expectation and states two exceptions without explaining it. Exactly. I totally agree with you that this guideline really should explain itself. Guidelines are supposed to guide.

On your first point, the use of LLMs to generate or rewrite article content is (IMO) broad enough to include drafts, which are explicitly intended to become article content. Generating a draft with LLMs is not much different from generating an article in mainspace with LLMs. Generating a draft LLM edit would be okay as long as the edit is limited to basic copyediting. SuperPianoMan9167 (talk) 13:23, 10 April 2026 (UTC)

Apologies, I don't mean a draft as in articles in draftspace, I mean editors generating a first draft (as in the noun) and then heavily editing that so that it is almost entirely human authored/verified before publishing to Wikipedia Respectfully dissenting, Coleisforeditor (talk) 13:26, 10 April 2026 (UTC)

Yeah I realized that after I wrote my comment. That sort of LLM editing is indeed a bit of a gray area. I would say that using an LLM to make a "first draft" of an edit is generally not allowed unless the edit is limited to copyediting. SuperPianoMan9167 (talk) 13:28, 10 April 2026 (UTC)

I agree. This is not a policy (yet), but it seems to be masquerading as one. Is there precedent for guidelines prohibit[ing] editors from, or permitt[ing] them to do, certain things? I think this guideline ought to be reworded to say, for example, [...] For this reason, editors should not use LLMs to generate or rewrite article content, except in the following cases: [...]. Pink Bee (talk) 15:09, 9 April 2026 (UTC)

Yes, there is precedent. Wikipedia:Spam prohibits self-promotion: Adding external links to an article or user page for the purpose of promoting a website or a product is not allowed It's not a policy. It's a guideline. Wikipedia:Notability is a guideline. It says Topics that do not meet this criterion are not retained as separate articles. Wikipedia:Do not create hoaxes is a guideline that tells editors to not do things. Wikipedia:External links is a guideline. There are no exceptions for the prohibition against linking to copyright violations.

The main argument made in this section is based on the common misconception that only policy pages have unbreakable rules. SuperPianoMan9167 (talk) 15:26, 9 April 2026 (UTC)

Thanks for finding those examples. I should have made my meaning clearer: I think the exact choice of wording is wrong – "prohibit" in particular, and also "permit", though to a lesser degree, since most things are permitted by default.

I am certainly not saying that a guideline can't tell people what to do, but I think it is enough simply to say that editors "should", "should not", "may", "must", etc., as is done on the pages you linked. (The only one of those that uses the word "prohibit" is WP:EL.) Alternatively, "generally prohibited" and "generally permitted" are popular choices.

I agree that guidelines can have rules that are effectively unbreakable, but is this really such a guideline? There are plenty of edge cases (see discussions further up this page) that make language as confidently authoritative as "prohibited" (without qualification) a poor choice. Pink Bee (talk) 16:21, 9 April 2026 (UTC)

Pink Bee, please see WP:BUREAUCRACY Kowal2701 (talk, contribs) 16:19, 9 April 2026 (UTC)

Which part? I agree with what it says, but I don't see how it applies to the wording of policies and guidelines – it seems to be about the application of them. Pink Bee (talk) 16:25, 9 April 2026 (UTC)

In a practical sense, there's no difference between policies and guidelines, they are all 'rules'. The authority sits with what the community actually does for which there is consensus for, not what's written on a page in project space. We could delete all 'rules' pages and live by convention, and the project would still function okay Kowal2701 (talk, contribs) 16:29, 9 April 2026 (UTC)

Indeed, though I do not think that is a reason not to word guidelines and policies as best we can. Pink Bee (talk) 16:54, 9 April 2026 (UTC)

Imo, the best wording is "prohibited" because it stops more disruption than other suggestions. Wording is based on applying them to get the best result. Also see WP:PWIH, Good policy writers remember that the real policy is what good editors really do, and that the words on a page with a "policy" tag at the top are only pale shadows of the true policy – the operational, day-to-day consensus of how Wikipedia is managed. Kowal2701 (talk, contribs) 16:59, 9 April 2026 (UTC)

On WP:PWIH, I also see the following, both directly relevant to the discussion at hand:

• Have you been careful to neither understate nor overstate a rule, in a way that a determined POV pusher might later dismiss as merely optional or irrelevant advice, or conversely as a hard requirement for circumstances you didn't intend? (I don't mean that as a flippant rhetorical question; I only mean to reference the point that PWIH makes with that question.)
• Try to signal the range of editorial judgment that is usually appropriate [...] This can be done partly by using words like should, usually, optionally, and must. [...] Use words like must, required, always, or never when there are no acceptable options (italics original).

That said (or quoted, I suppose), I'm not sure I'm going to get anywhere with this, so I'll stop. Pink Bee (talk) 17:18, 9 April 2026 (UTC)

I noticed that you 'prohibit' the use of LLMs, but this is only a content guideline, not a content policy, so even if you use the word 'prohibited', ' exceptions may apply'

This is a common misconception. SuperPianoMan9167 (talk) 15:31, 9 April 2026 (UTC)

The only reason it's a guideline and not a policy is because the people that proposed it didn't discuss whether it ought to be guideline or policy. As one of those people, my thinking was that we only have 7 content policies, all of which are very high level, fleshed out, and tremendously important. Given the near unanimous consensus at the RfC, this could have been a policy, but it doesn't matter anyway, the difference between guidelines and policies are negligible, and anyone wikilawyering over that will most likely end up blocked quicker Kowal2701 (talk, contribs) 16:16, 9 April 2026 (UTC)

Is Google Translate an LLM?

I have the question. What about the human editors that use Google Translate for translating text to other languages? Is Google Translate one of the Large Language Models? RW Gamer (a typical editor and a gamer) (talk) 14:09, 6 April 2026 (UTC)

My understanding is that it depends on what languages you are translating between. To be on the safe side, it is best to follow the guidance at WP:LLMTRANSLATE when translating articles using Google Translate (or even better, just translate the article text manually from languages you know, and leave languages you don't know to be translated by people fluent in those languages). -- LWG ^{talk (VOPOV)} 14:26, 6 April 2026 (UTC)

Google Translate switched fairly recently, December 2025 iirc. It's a little unclear how broadly it applies -- the blog post implies that it's everything, but other sources I've seen say it's gated behind an "Advanced" feature that only shows up sometimes.

but it's definitely in there in some capacity, given that someone managed to pull off prompt injection on it Gnomingstuff (talk) 20:28, 7 April 2026 (UTC)

There are two modes you can switch between: "Classic" and "Advanced".

I'm guessing "Advanced" makes an API call to Gemini with a prompt like "Translate the following text in [language 1] to [language 2]". "Classic" seems to use the older transformer models specialized for translation. SuperPianoMan9167 (talk) 20:32, 7 April 2026 (UTC)

Ah, I see that now. On the web app it says Advanced (Improved accuracy, built with Gemini) with a note that says Advanced is supported for text translation on in select languages. Trying it out translating some of my own academic writing from English to Spanish, the "Advanced" option seems to produce more natural Spanish flow but does slightly worse at capturing the exact factual content of what I wrote and stumbles on some jargon. I wouldn't say one is necessarily better than the other. Neither would be good enough for inserting unedited into Wikipedia. -- LWG ^{talk (VOPOV)} 21:05, 7 April 2026 (UTC)

Unfortunately, prompt injection seems to not work anymore. SuperPianoMan9167 (talk) 21:09, 7 April 2026 (UTC)

I believe Google Translate does use an LLM nowadays, yes, as does most major translation software. Regardless, the guidance at WP:LLMTRANSLATE is good advice. Toadspike [Talk] 14:40, 6 April 2026 (UTC)

Yes, "Advanced" translation is just Gemini under the hood. Google Translate has been using transformer models for some time now. SuperPianoMan9167 (talk) 16:51, 6 April 2026 (UTC)

Proposal to add note on citation scraping

I often use LLMs to scrape the internet for citations on an obscure topic. For example, if at FARC an article has many {{cn}} tags it may be useful to use a LLM to help find these. This is because they often have a better ability to scrape text than google. I find Gemini is terrible at this but ChatGPT is quite good. Obviously, I then go and check what the sources say and use the sources as usual, because ChatGPT has, anecdotally, around a 10% false positive rate of saying a source says something it doesn't. However, I think we should be careful to not discourage people from doing this as getting it to find URLs that you then manually verify is a good use of AI. I was thinking something like:

LLMs can be used to find sources on a topic or to find a reference for an uncited statement. However, these must be manually verified by a human editor before being added due to the risk of hallucination.

Any thoughts? JacobTheRox^{(talk | contributions)} 20:36, 10 April 2026 (UTC)

We are against the wholesale AI creation of content, but using AI for research, like to find sourcing, should be all right. I think of this use of AI as augmented search. AI may become very useful in assisting us to reduce a lot of our missing citations. Of course, what I state here is my opinion and doesn't necessarily fit the current guideline, but if the guideline needs clarifying in this area, I'm for it. the Stefen 𝕋ower 21:02, 10 April 2026 (UTC)

I think LLMs can be useful for a great number of things. As long as they are not being used to generate or rewrite article content, this guideline does not apply, and clarification isn’t needed. I₂Overcome ^talk 21:17, 10 April 2026 (UTC)

I saw someone describe this guideline as being akin to WP:BROTHER, which seemed clear and useful to me. I.e. I'd describe an alternative/informal Nutshell as being along the lines of: "Nobody is allowed to blame an LLM for any mistakes they make; the human has full responsibility for all edits or errors made from their account". ~2026-22190-34 (talk) 22:26, 10 April 2026 (UTC)

In my opinion, that isn't "generating content" any more than googling for sources is "using Google to generate content". While an LLM isn't a search engine, the function it's serving here is the same - it's finding links that already exist on the web, not generating new content. Helpful Cat🐈(talk) 02:58, 11 April 2026 (UTC)

This guideline is specifically prohibiting using LLMs to write the text that you put into Wikipedia. Feel free to use whatever tools you like to find resources or learn information, so long as the actual text you put on Wikipedia is written by you, not an LLM, and verifiable to sources you have actually read, not just had an AI summarize for you. -- LWG ^{talk (VOPOV)} 03:42, 11 April 2026 (UTC)

Discussion at Wikipedia:Village pump (idea lab) § Presumptive deletion for AI-blocked users?

 You are invited to join the discussion at Wikipedia:Village pump (idea lab) § Presumptive deletion for AI-blocked users?. Chaotic Enby (talk · contribs) 06:55, 11 April 2026 (UTC)

My thoughts

The following discussion is closed. Please do not modify it. Subsequent comments should be made on the appropriate discussion page. No further edits should be made to this discussion.

---

I completely understand this ban, having witnessed a number of major issues with LLMs, particularly with claims and problems with verification in the sources given. ChatGPT still makes basic errors over simple things, though is gradually improving. At present level, they aren't trustworthy enough to write content. However, I think it's only a matter of time before AI is advanced enough to be able to consistently draft articles better than humans without all the sourcing issues, or at least save a great deal of time in writing. I think at some point Wikipedia runs the risk of becoming obsolete in the future, a dinosaur of the early web, if it doesn't accept AI editing, and will be dwarfed by the tech giants in both size and quality eventually and we will gradually lose traffic (and subsequently not raise enough in donations to last, continue to grow and compete with giant AI encyclopedias). The world is quickly becoming an AI world whether we like it or not. I don't think this ban can be a permanent thing, it should only be until LLMs are consistently highly accurate and accomplished at writing encyclopedia content. I think the ban should be reconsidered at some point when the LLMs become much more competent, which I think is inevitable. Ideally our own AI technology which can assist with research and drafting and making editing more efficient would be ideal. ♦ Dr. Blofeld 16:21, 13 April 2026 (UTC)

Wikipedia is written by humans for humans. Our readers come here because they want verifiable information written by humans. If anyone wants to know what AI's got to say, then go and ask AI, but that is not what Wikipedia is for. Kovcszaln6 (talk) 16:47, 13 April 2026 (UTC)

I agree, my point is that I believe we are heading into a world in which AI does just about everything and in the future the amount of people wanting an encyclopedia written by humans will be a small, niche thing, given that at some point I believe it will become vastly superior to human editing. ♦ Dr. Blofeld 16:54, 13 April 2026 (UTC)

LLMs rely on Wikipedia for training data. (Human-written) Wikipedia is actually necessary for AI improvement, unless you want model collapse from training LLMs on LLM output. SuperPianoMan9167 (talk) 17:22, 13 April 2026 (UTC)

2022: "It's only a matter of time before AI makes Wikipedia obsolete."

2023: "It's only a matter of time before AI makes Wikipedia obsolete."

2024: "It's only a matter of time before AI makes Wikipedia obsolete."

2025: "It's only a matter of time before AI makes Wikipedia obsolete."

2026: "It's only a matter of time before AI makes Wikipedia obsolete."

*cough cough* So why hasn't it happened yet? SuperPianoMan9167 (talk) 17:20, 13 April 2026 (UTC)

@User:SuperPianoMan9167 ChatGPT launched in Q4 of 2023 so I do not know what you think you are talking about by reputedly quoting claims that date as earlier as 2022 claiming that "AI will replace Wikipedia." No one even today seems to be claiming that; what we are claiming, those of us that oppose this anti-AI policy, is that generative AI (LLMs) should not be uniformly banned given its present research capabilities and human fact-checking safeguards. It is best to keep up with the changing world rather than instituting retrogressive, extreme, outright ban with no expiration in sight. The policy should at least contain a clause setting forth clearly what threshold an LLM must meet to supersede the policy. Simply stating that OpenAI is gonna go under and Google and Meta will dissolve their AI programs is not remotely supported by their massive building plans nor by the reality of the AI arms race forcing US companies to try to keep up and get ahead of rival Chinese companies that are also pursuing heavily this research. Ballardy Talk Page 17:14, 24 April 2026 (UTC)

ChatGPT came out November 2022, not 2023. (I'm pretty sure fiscal years are one year ahead of the current year You have the date wrong. Please ignore the crossed-out text, it's not relevant.). My point is that repeated predictions of AI integration into Wikipedia (i.e. statements saying, in essence, "adapt to this new AI-centric world or fall into obscurity") are hard to take seriously when they've been made nearly constantly for the past three years.

SuperPianoMan9167 (talk) 17:24, 24 April 2026 (UTC)

Even so, where are the claims that AI is going to supplant Wikipedia? And how does that relate to this policy? All I am saying is give it a chance. Ballardy Talk Page 18:35, 24 April 2026 (UTC)

Why? AI is still in its infancy. We're not going to be replaced in the near future. But you think decades into the future humans will still be writing encyclopedia articles without AI at a snail's pace? I would love for you to be proven right. In 2100 do you think people will still want information written exclusively by humans? ♦ Dr. Blofeld 17:27, 13 April 2026 (UTC)

If we survive the existential problems we are currently facing, and survive for 100 years, I doubt there will be notable topics we haven't covered left. Even if we put that aside, AI isn't profitable. Either there will be bubble burst or a gradual decline in investment, and AI companies will need to find a way to make a profit. My best bet is AI companies will integrate ads into the text, thus creating a cycle of enshitification, or charge people to use the AI. We need to outlast the cycle. We need to be a quality source of information for all, existing long after the AI's wild west era is over. We need to be immortal. We survived the end of the internet's wild west era, and we will survive AI's wild west era. Mikeycdiamond (talk) 18:21, 13 April 2026 (UTC)

Yes, one of our strengths has always been no adverts. But I believe there will always be topics to write about, I remember comments in the 2000s that by the 2020s we'd be almost complete! The thought of Wikipedia turning into a vast, automated, soulless encylopedia is a horrid thought. I just think our world will change markedly over the next 20 years and gradually AI start to do things we used to do and become very normal. At some point I believe the technology will become so advanced that we would be greatly hampering our development and potential by remaining the same way. I think at the very least WM should develop some AI research tools to vastly improve how we research and make editing much more efficient. I don't think we have to worry much at the moment, and, yes, it's true that LLMs currently need human content to train on.

I am just thinking well into the future and I think it is inevitable we will have to adapt to the times at some point. I can also see the possibility at some point that text based encyclopedia content starts to evolve into something more video/VR visual based. Imagine reading an article on a castle in Scotland and being immersed in that environment as you learn about it and feeling like you are at that place as you learn. Obviously I'm talking later 21st century, if our silly leaders haven't destroyed us all by then! I just want our beloved pedia to stand the test of time and not become a relic of the old web, supported by a niche community that's all. I wouldn't underestimate how powerful AI could become, even with it being currently lousy at writing and sourcing great Wikipedia content.

♦ Dr. Blofeld 19:32, 13 April 2026 (UTC)

AI is still in its infancy

Read this. SuperPianoMan9167 (talk) 18:29, 13 April 2026 (UTC)

I think it's unlikely AI will surpass humans in this regard any time soon, but if that happens then we can certainly reassess this guideline at that time. There's no need to worry about it preemptively. Justin Kunimune (talk) 19:25, 13 April 2026 (UTC)

Agreed. ♦ Dr. Blofeld 19:36, 13 April 2026 (UTC)

Agreed with this as well. SuperPianoMan9167 (talk) 19:37, 13 April 2026 (UTC)

Disagreed. Why not get ahead of the curve? It is inevitable that this policy will grow outdated and be set aside.Ballardy Talk Page 14:21, 24 April 2026 (UTC)

It is not our job to make such predictions. SuperPianoMan9167 (talk) 14:28, 24 April 2026 (UTC)

This is not a prediction. It is guaranteed; just a question of when. Ballardy Talk Page 14:38, 24 April 2026 (UTC)

I think it is more likely that OpenAI and Anthropic will go bankrupt and run out of cash before this policy has to be revised. SuperPianoMan9167 (talk) 14:52, 24 April 2026 (UTC)

Let me guess--you suspect Google and Meta will as well? Ballardy Talk Page 14:59, 24 April 2026 (UTC)

No. Google and Meta, as big tech companies, are able to make profit even without AI. The vast majority of their revenues comes from selling ads, not AI subscriptions.

Anthropic and OpenAI are currently not profitable; OpenAI says it won't be profitable until at least 2030. SuperPianoMan9167 (talk) 15:06, 24 April 2026 (UTC)

If that is your view on the future of AI, you are entitled to it. But please preserve this chat. In ten years one year, we'll be able to check back and see who is right. Ballardy Talk Page 15:28, 24 April 2026 (UTC)

Not a problem, as every discussion here gets archived. In fact, a prediction I made in November of last year ended up coming to pass when the stricter version of this guideline passed last month. SuperPianoMan9167 (talk) 15:46, 24 April 2026 (UTC)

More recent versions of AIs hallucinate more. As AIs get more unreliable, it will be less likely that they can contribute on Wikipedia without making errors/hallucinations. Mikeycdiamond (talk) 15:36, 24 April 2026 (UTC)

Moreover, the policy to which you've referred concerns articles about unknown future events- it does not concern policy questions. Ballardy Talk Page 14:39, 24 April 2026 (UTC)

The idea behind predicating this guideline on LLM-generated edits violating PAGs was that if/when LLM-generated edits generally comply with PAGs, this would need rewriting and the community would have to decide on an ideological ban or to allow their use. This guideline's only meant to serve our needs in the short term. I’m in favour of an ideological ban, for reasons which will hopefully be outlined in a future essay, but we’d need either a very good system to detect such edits, or AI companies would need to be regulated to make LLM output easy to detect. Given the effect on education systems, hopefully it’s the latter, but if not, imo this site is fucked Kowal2701 (talk, contribs) 22:17, 15 April 2026 (UTC)

Wikipedia will doubtlessly someday be forced to repeal this policy once AI reaches an advanced-enough stage

When? 2026? 2027? 2030? 2050? 2100? SuperPianoMan9167 (talk) 14:26, 24 April 2026 (UTC)

Maybe 2026. Maybe 2100. More likely closer to the former. I would argue that its deep-research abilities extant already are sufficiently capable for being permitted. Either way, please, if you oppose the idea, do vote below. Ballardy Talk Page 14:30, 24 April 2026 (UTC)

Well, when and if that happens, it will happen; WP:CCC exists for a reason. But it certainly hasn't happened now. And of course part of your argument (I assume!) is that the technology will improve - but it hasn't improved yet, so that's hardly an argument for allowing it now. --Aquillion (talk) 15:02, 24 April 2026 (UTC)

Also, Wikipedia typically avoids the use of polls to gauge support for a policy. SuperPianoMan9167 (talk) 14:31, 24 April 2026 (UTC)

Not trying to use this as a substitute for discussion; I encourage editors to include their reasons with their vote below. But discussion is easier to follow when differing views are sorted into their respective category. Ballardy Talk Page 14:32, 24 April 2026 (UTC)

It is very much possible that, one day, LLMs will regularly produce human-quality writing with minimal hallucinations and text-source integrity errors. When that is the case, we can absolutely revisit the discussion. However, we shouldn't get ahead of ourselves, as the community generally agrees that LLMs in their current state are much more often cause for disruption. This guideline is barely a month old, and, while consensus can change, it is doubtful that such a massive shift happened in just a month. Chaotic Enby (talk · contribs) 14:38, 24 April 2026 (UTC)

Tell me the metric by which this is being judged. Because it is apparent that already, Gemini and GPT have research modes that are highly superior in efficiency and no inferior in quality to human work. Ballardy Talk Page 14:41, 24 April 2026 (UTC)

This policy assumes that users would copy-and-paste without fact-checking instant outputs. It should at least allow for certain research deep-thinking mode work to be usable if confirmed as valid. Ballardy Talk Page 14:44, 24 April 2026 (UTC)

The problem is that it takes a lot of effort to confirm output as valid, so much so that in many cases it would be easier to just write the article yourself. SuperPianoMan9167 (talk) 14:48, 24 April 2026 (UTC)

That may be, but it does not support the conclusion that barring editors from doing so has a rational basis. Ballardy Talk Page 14:51, 24 April 2026 (UTC)

Additionally, we know that, usually, people don't do this effort, even though our policies already require editors to be responsible for the content of their edits and to not add things like fake citations. Chaotic Enby (talk · contribs) 14:50, 24 April 2026 (UTC)

So your defense of the policy is an assumption of bad-faith? That is contrary to hornbook policy, see Wikipedia:Assume good faith Ballardy Talk Page 14:52, 24 April 2026 (UTC)

After three years of cleaning up poorly written LLM-generated content, it is understandable that many Wikipedians have less patience for LLM use, especially because it takes a lot more effort to clean up than it does to produce. SuperPianoMan9167 (talk) 14:55, 24 April 2026 (UTC)

The governing rule here remains that we assume good-faith. Ballardy Talk Page 14:58, 24 April 2026 (UTC)

We gave AI four years. People submitted unchecked and hallucinated work. People used AI to generate content that deliberately violated NPOV. People wasted editor time generating talk page topics. AI wasted hundreds of hours of editor time fixing content generated by it. Whether people intentionally used AI to damage Wikipedia or not, AI has caused immeasurable amount of damage that we will be fixing for years to come. The fundamental programing of AI makes it unreliable; AI was meant to predict the next word in a sentence, not write factual and verifiable information. We are done with spending hours to fix damage that takes people seconds to make. Allowing any exceptions would be reopening the floodgates of slop that plagued us for so long. Mikeycdiamond (talk) 20:10, 24 April 2026 (UTC)

That may well be your position now, but there will likely come a time, likely sooner than you think, when those past problems will no longer be sufficient to justify a continuing uniform ban on the technology. If you are of the view that Wikipedia must forever be a human-edited encyclopedia, I would suggest that such a view is too extreme and you should consider being a bit more open-minded as AI research proceeds and LLMs continue to evolve. Ballardy Talk Page 21:46, 24 April 2026 (UTC)

Defining the Scope

This policy has been described, albeit in dicta (in this talk page only), as a short-term solution and its formality as a binding policy is opaque at this point to me based on the statements throughout this talk page. Is this formal policy, a temporary restraining order, an advisory rule, or what? Want to understand exactly what it is. If it is not yet a formal policy, then I wish to know this so that I can be an active voice against codifying it in the appropriate forum when it does come up for formal ratification. Ballardy Talk Page 17:27, 24 April 2026 (UTC)

This is a formally defined guideline, which was enacted by the community a month ago in a widely publicized discussion. Wikipedia is not a court of law, and all our policies and guidelines (including this one!) are subject to future review and changes, although bringing it up for discussion only one month after it was already discussed may be seen as disruptive as there haven't been any major changes in the meantime. Chaotic Enby (talk · contribs) 17:47, 24 April 2026 (UTC)

GPT 5.5 was released yesterday, so to say no major changes have occurred without probing the new model at all is both incorrect and unfair. How do you know that it still remains incapable of outputting reliable information without auditing the new model in any form? Ballardy Talk Page 22:12, 24 April 2026 (UTC)

It is a guideline. While there are technical differences between policies and guidelines, they are treated the same in practice. Mikeycdiamond (talk) 18:09, 24 April 2026 (UTC)

When do you anticipate it will next be revisited formally? Want to know so I can participate fully. Ballardy Talk Page 22:14, 24 April 2026 (UTC)

References

The discussion above is closed. Please do not modify it. Subsequent comments should be made on the appropriate discussion page. No further edits should be made to this discussion.

I should add my two cents here, considering that this is uh where the "revisit later" camp resides. I hate the existing LLM edits characterized by "hot air", unnecessary emboldening, and worst of them all mangling of existing content (especially refs with names). At the same time I do believe that a competent Wikipedian and AI wrangler can coax AI into producing correct, MOS-compliant, and not "LLM flavored" wikitext with a combination of: a good model (perhaps one existing right now), an "agent" framework that actually reads the sources and interacts with talk pages etc., and a good amount of examples and coaching (maybe one of the newfangled SKILL.md things). But I'm not going to test that. There's no fun in risking my whole editing career for a more boring way to write.

As for whether an LLM-specific guideline is required: while in principle all the LLM badness is covered by existing policies, it does help to tell inexperienced writer-prompters to "not even try". --Artoria2e5 🌉 14:40, 8 May 2026 (UTC)

Grok

Why is Grok included in the LLM guideline? Absolutiva 23:27, 13 April 2026 (UTC)

It is one of the most well-known LLMs, so it seems natural to cite it as an example alongside others. Chaotic Enby (talk · contribs) 02:30, 14 April 2026 (UTC)

Specifically, it's (likely) the third most popular LLM after ChatGPT and Gemini, at least in the US. Gnomingstuff (talk) 06:22, 14 April 2026 (UTC)

Motion to Repeal

We had a clear RfC result almost exactly one month ago. While consensus can change, there is no reason why it would change so rapidly and drastically as to reverse a one-old-month RfC. As such, this "motion" is disruptive relitigation and wasting community time. Fram (talk) 14:59, 24 April 2026 (UTC) The following discussion has been closed. Please do not modify it.

Wikipedia will doubtlessly someday be forced to repeal this policy once AI reaches an advanced-enough stage and is integrated deeply enough into academic work. We need to keep up with the times and be open to emerging technologies as opposed to antagonistic to it. I would just suggest that AI-generated content fact-checked fully by human editors be made permissible sooner rather than later, because it is likely inevitable either way that this will happen. Motion to repeal this policy; indicate your position below.Ballardy Talk Page 14:08, 24 April 2026 (UTC)

Wikipedia will doubtlessly someday be forced to repeal this policy once AI reaches an advanced-enough stage

When? 2026? 2027? 2030? 2050? 2100? SuperPianoMan9167 (talk) 14:26, 24 April 2026 (UTC)

Maybe 2026. Maybe 2100. More likely closer to the former. I would argue that its deep-research abilities extant already are sufficiently capable for being permitted. Either way, please, if you oppose the idea, do vote below. Ballardy Talk Page 14:30, 24 April 2026 (UTC)

Well, when and if that happens, it will happen; WP:CCC exists for a reason. But it certainly hasn't happened now. And of course part of your argument (I assume!) is that the technology will improve - but it hasn't improved yet, so that's hardly an argument for allowing it now. --Aquillion (talk) 15:02, 24 April 2026 (UTC)

Also, Wikipedia typically avoids the use of polls to gauge support for a policy. SuperPianoMan9167 (talk) 14:31, 24 April 2026 (UTC)

Not trying to use this as a substitute for discussion; I encourage editors to include their reasons with their vote below. But discussion is easier to follow when differing views are sorted into their respective category. Ballardy Talk Page 14:32, 24 April 2026 (UTC)

It is very much possible that, one day, LLMs will regularly produce human-quality writing with minimal hallucinations and text-source integrity errors. When that is the case, we can absolutely revisit the discussion. However, we shouldn't get ahead of ourselves, as the community generally agrees that LLMs in their current state are much more often cause for disruption. This guideline is barely a month old, and, while consensus can change, it is doubtful that such a massive shift happened in just a month. Chaotic Enby (talk · contribs) 14:38, 24 April 2026 (UTC)

Tell me the metric by which this is being judged. Because it is apparent that already, Gemini and GPT have research modes that are highly superior in efficiency and no inferior in quality to human work. Ballardy Talk Page 14:41, 24 April 2026 (UTC)

This policy assumes that users would copy-and-paste without fact-checking instant outputs. It should at least allow for certain research deep-thinking mode work to be usable if confirmed as valid. Ballardy Talk Page 14:44, 24 April 2026 (UTC)

The problem is that it takes a lot of effort to confirm output as valid, so much so that in many cases it would be easier to just write the article yourself. SuperPianoMan9167 (talk) 14:48, 24 April 2026 (UTC)

That may be, but it does not support the conclusion that barring editors from doing so has a rational basis. Ballardy Talk Page 14:51, 24 April 2026 (UTC)

Additionally, we know that, usually, people don't do this effort, even though our policies already require editors to be responsible for the content of their edits and to not add things like fake citations. Chaotic Enby (talk · contribs) 14:50, 24 April 2026 (UTC)

So your defense of the policy is an assumption of bad-faith? That is contrary to hornbook policy, see Wikipedia:Assume good faith Ballardy Talk Page 14:52, 24 April 2026 (UTC)

After three years of cleaning up poorly written LLM-generated content, it is understandable that many Wikipedians have less patience for LLM use, especially because it takes a lot more effort to clean up than it does to produce. SuperPianoMan9167 (talk) 14:55, 24 April 2026 (UTC)

The governing rule here remains that we assume good-faith. Ballardy Talk Page 14:58, 24 April 2026 (UTC)

On the motion to repeal this policy, please leave your signature below under the position you support:

Those in favor:

● Ballardy Talk Page 14:49, 24 April 2026 (UTC)

Those in opposition:

Motion to Reopen Debate

I have consolidated the debate-only content of the closed discussion to the previous already existing ongoing discussion and we will continue it there. Matter resolved. Motion withdrawn. Ballardy Talk Page 16:05, 24 April 2026 (UTC)

@Fram I am respectfully moving for the reopening of the above discussion. Cloture is not warranted simply on the ground that you feel this issue has been resolved and are unpersuaded by new arguments presently. You are free to reject my arguments and explain your reasons in the discussion. But let me have a voice. Ballardy Talk Page 15:08, 24 April 2026 (UTC)

Yeah, @Fram, this looks like a WP:SUPERHAT. Specifically a WP:BADHAT. People are commenting in the discussion even after you hatted it. SuperPianoMan9167 (talk) 15:10, 24 April 2026 (UTC)

One person commented to basically say the same thing I said in my close. Fram (talk) 15:17, 24 April 2026 (UTC)

You don't want to "have a voice", you want a new RfC (never mind that you call it a "motion") to overturn the previous one. You have presented no arguments why the closure of that one was procedurally incorrect, you just don't like the result. It's very normal to have discussions (AfDs, RfCs, whatever) reach a consensus you don't agree with, but in such a case, you just have to live with it and move on to other things on Wikipedia. Simply restarting the debate like you did here is wasting community time. Fram (talk) 15:13, 24 April 2026 (UTC)

While that may be true, I don't think it was a good idea to collapse the entire discussion while people were still discussing. SuperPianoMan9167 (talk) 15:15, 24 April 2026 (UTC)

"People" weren't still discussing, and in any case the hat was closed "to stop timewasting discussions on a just-decided thing". Fram (talk) 15:18, 24 April 2026 (UTC)

I agree with Fram. The "motion" above was an improper way was to attempt to change policy. If you want to debate, sure, go do it, but don't frame the debate as having the same power as a RFC. @Ballardy open motions with signatures aren't how we do things here. If I were you, I would wait a few months to see if consensus changes and then open a RFC. In the meantime, you can open another topic to discuss this policy. Mikeycdiamond (talk) 15:28, 24 April 2026 (UTC)

It was not an attempt to change the policy. If I wanted to do that I would have gone directly to the policy and changed it. I styled it as a "motion" because that is how debates usually are started. Ballardy Talk Page 15:34, 24 April 2026 (UTC)

I am happy to get rid of the voting portion of the discussion, but the rest of the debate was lively and intelligent and good-faith. Ballardy Talk Page 15:39, 24 April 2026 (UTC)

Maybe that is true, but if I cannot have that then I do still want to be able to discuss the issue. Depriving me of that too is not justifiable. Ballardy Talk Page 15:40, 24 April 2026 (UTC)

Nobodoy is depriving you of anything, you were already discussing it above. Kowal2701 (talk, contribs) 15:42, 24 April 2026 (UTC)

@Ballardy, hi, gonna save you some time: Wikipedia is not a court of law (nor a debate club) the process for these sorts of things is at WP:PROPOSAL, ie. an RfC. The RfC to adopt this guideline was closed a few weeks ago with near-unanimous consensus in favour, and it’s usually inappropriate to have another RfC on the same topic very soon after Kowal2701 (talk, contribs) 15:27, 24 April 2026 (UTC)

When I used the word "motion" it was not meant in a parliamentary sense. It was intended to get a discussion going, which in fact it did, until it was improperly teriminated. Ballardy Talk Page 15:31, 24 April 2026 (UTC)

Wikipedia is WP:NOTADEMOCRACY though. --Gurkubondinn (talk) 15:38, 24 April 2026 (UTC)

My Discussion was not intended to be any formal "RfC" and I am happy to restyle it as something other than a "motion". I am just asking that I be allowed to voice my outlook. Ballardy Talk Page 15:32, 24 April 2026 (UTC)

You can do that in a normal discussion, like the one above (beware of WP:BLUDGEON). But please don't expect people to treat your opinion as any more important than anyone else's, most people's views here are informed by editing experience, ie. having to clean up/ deal with the endless spam of slop we get, and even uses of LLMs by highly competent people to write content have been proven problematic Kowal2701 (talk, contribs) 15:37, 24 April 2026 (UTC)

That is not acceptable when a debate was already ongoing. Reinstate the discussion and we'll relabel it to ensure no one is confused regarding the nature of it. Ballardy Talk Page 15:41, 24 April 2026 (UTC)

Is there no procedural difference between a formal RfC and informal ongoing discussion? Ballardy Talk Page 15:55, 24 April 2026 (UTC)

I support Fram's hatting. There's no reason to think that consensus has changed in the last month. Please respect our RFC process and please respect experienced editor time. Re-litigating things too early is not efficient. –Novem Linguae (talk) 15:33, 24 April 2026 (UTC)

Your reasoning does not sustain hatting at all; only that the policy not be amended. But that is not the issue; consensus may well have not changed in a month; I am sure it has not. That has nothing to do with permitting continuing discussion. It should be permitted and deeming my post as something else because I labeled it a "motion" is censorship. I said we can relaabel it to merely "discussion". If that is not good enough, then this is improper cloture of a valid discussion. Ballardy Talk Page 15:34, 24 April 2026 (UTC)

You can still discuss, but RFCs--what you were trying to do with your motion--take time. On many RFCs tens of editors spend a lot of time formulating logical arguments to support their point of view. To create an RFC so soon after the most recent one, without any major developments, is a waste of time that could be used for other purposes. Consensus has not changed. Mikeycdiamond (talk) 15:39, 24 April 2026 (UTC)

The discussion you now claim you want to have is the one directly above, the section "My thoughts". You can continue participating there and post your thoughts there. One other thing, above you said "It was not an attempt to change the policy. If I wanted to do that I would have gone directly to the policy and changed it. " That would end badly, you should never do that. Fram (talk) 15:45, 24 April 2026 (UTC)

I am well aware of that. Once again, this is all unnecessary time wasting. Open up my discussion on my word that I will swiftly remove the references to voting or motions. The rest of the debate was quality and I want to continue it. Ballardy Talk Page 15:47, 24 April 2026 (UTC)

You don't seem to be reading the room very well here, nor being respectful of other's time by making 25 comments so quickly. Please see WP:BLUDGEON. –Novem Linguae (talk) 15:48, 24 April 2026 (UTC)

I do not take kindly to outright censorship. This would all be over if the above discussion were reopened. If you insist on being a dog in a manger and refusing to reopen it, then yes, I am going to respectfully protest it and ask that it be undone. Ballardy Talk Page 15:51, 24 April 2026 (UTC)

I support it too, but now there's this second thread which is also becoming a WP:TIMESINK. --Gurkubondinn (talk) 15:39, 24 April 2026 (UTC)

Just open it up again and I will remove all references to motions or voting to ensure no confusion. We'll keep it discussion only. That is it! Ballardy Talk Page 15:43, 24 April 2026 (UTC)

The next step is an administrative appeal, and I am asking you to please not force me to do that. I will remove references to voting or motions to prevent confusion upon reopening of the above discussion. Please do not make this more time consuming than it ought to be. Ballardy Talk Page 15:49, 24 April 2026 (UTC)

You are free to make an administrative appeal of course, but frankly, there is only one person here making "this more time consuming than it ought to be". Nothing is gained by reopening and refactoring a discussion when a similar discussion (where you already participate as well) is right above it. Fram (talk) 15:56, 24 April 2026 (UTC)

I implore you to be merciful. An RfC is not merely any attempt to discuss a topic. It is more formal and specific than that. I concededly couched the discussion in terms too formal and that created the confusion; I said I am happy to amend that. Would you please just end this? Ballardy Talk Page 15:59, 24 April 2026 (UTC)

"Nothing is gained by reopening and refactoring a discussion when a similar discussion (where you already participate as well) is right above it" is (A) your opinion and (B) not a basis for cloture. Ballardy Talk Page 16:01, 24 April 2026 (UTC)

WP:MULTI Kowal2701 (talk, contribs) 16:02, 24 April 2026 (UTC)

Edit request

Could you please replace "save for" with "except for"? The latter is more accessible language. TheAuroraBorealis (u/t/c) 19:57, 24 April 2026 (UTC)

I think except for these two exceptions is awkward wording, though I agree "save for" is less accessible Kowal2701 (talk, contribs) 20:07, 24 April 2026 (UTC)

except these two exceptions? TheAuroraBorealis (u/t/c) 20:13, 24 April 2026 (UTC)

 Done Mikeycdiamond (talk) 20:21, 24 April 2026 (UTC)

How about "besides these two exceptions?" Repeating "except" twice feels redundant. I₂Overcome ^talk 20:27, 24 April 2026 (UTC)

@I2Overcome +1 that may be a better solution TheAuroraBorealis (u/t/c) 20:29, 24 April 2026 (UTC)

 Done Mikeycdiamond (talk) 20:31, 24 April 2026 (UTC)

...excluding these two exceptions? Kowal2701 (talk, contribs) 20:30, 24 April 2026 (UTC)

…except in these two specific cases, maybe? I₂Overcome ^talk 20:36, 24 April 2026 (UTC)

Like Kowal2701 (talk, contribs) 20:39, 24 April 2026 (UTC)

 Done I₂Overcome ^talk 20:47, 24 April 2026 (UTC)

This whole fiasco...

regarding the attempted relitigation above is, IMO, all the more reason that our prohibition on AI needs to be on ideological grounds, not argumentative ones. So long as we forbid AI on the grounds that it doesn't generate good Wikipedia content, there will be a Ballardy somewhere who will argue "guys, Schlop 3.7 hallucinates 73% less than its predecessor, can we unban AI usage now?" In that sense, a valid point is made that we are just kicking the can down the road; but not for the reason that AI will eventually be suitable for Wikiprdia, rather that there will always be someone interested in forcing us to relitigate the issue; and while that might not be an issue now while we can write it off as too soon since the previous RfC; do we want to be having the same conversation about whether AI still generates problematic content every 6 months or every year? If our position is that AI is unsuitable for Wikipedia and we don't want to welcome it here; then let's just say that outright.

We should take an anthropocentric position. Wikipedia; by humans, for humans, full stop. AI is actively malignant, the CEOs of AI companies want to integrate their product into as many facets of life as possible in order to create dependency which they can then monetise. Let's not even entertain the possibility. Athanelar (talk) 12:46, 25 April 2026 (UTC)

I think there might be a bit of an issue with that since the WMF is actively looking into using AI & seem to be rather positive about it (as with most large organisations).

It's been mentioned on the recent AI/TomWikiAssist Village Pump thread, it's freaking massive so forgive me for not providing a diff right now.

If it's integrated into the software, what do we do?

We also use very basic AI in the form of bots, so the difficulty would be in where to draw the line.

Finally, there's the perennial issue of getting more than three editors to agree that the sky is blue.

Honestly I'd love for this to be a set Thing with it's own policy page, but I'm not sure I see it happening. Blue Sonnet (talk) 14:21, 25 April 2026 (UTC)

Bots do not introduce prose into articles as far as I'm aware, and I doubt any AI functionality the WMF introduces would do so either. Our stance can still be "every word in every article should come directly from a human brain" without falling afoul of those things.

Will someone try to call this out as 'inconsistent'? Sure, I don't doubt it. But we should reject those criticisms for what they are; the continuing attempts by AI-poisoned technophiles to convince us that surrendering humanity's greatest knowledge project to the whims of Sam Altman's cadre of ghouls is a wise or inevitable future path. Athanelar (talk) 14:49, 25 April 2026 (UTC)

Bots do not introduce prose into articles as far as I'm aware – I would like to introduce you to Category:Articles created by bots. These are stubs, but we have had bots (not LLMs, but still bots) write articles before. Chess enjoyer (talk) 15:12, 25 April 2026 (UTC)

Even with those, the bots aren't introducing prose to articles, since all the prose that ends up in the articles was written by the human who wrote the bot's source code. All the bots are doing is importing (human-written) content from bot-readable databases and slotting that content into predefined holes in (human-written) prose templates. -- LWG ^{talk (VOPOV)} 16:02, 25 April 2026 (UTC)

Agree very much and will support this when the time comes. One thing: "by humans, for humans" - LLMs will read wikipedia and I have no problem with that... it is really "by humans, for all" or something. The critical part is that it is written by humans of course. If you can come up with something pithy then most appreciated NicheSports (talk) 15:47, 25 April 2026 (UTC)

LLMs scraping Wikipedia is fine, and I don't think it really undermines the idea of it being 'by and for humans'. It's fine if automated processes are 'reading' Wikipedia text as a byproduct, but I think our aim should still ultimately be to create a human-written encyclopedia for human readers. Athanelar (talk) 18:26, 25 April 2026 (UTC)

I'm still hoping that LLMs and their sloppy problems will go the way of blockchain technology in a couple years as everyone realizes they don't have the transformative effect people were hoping for. Failing that I would support an ideological commitment to human-written content. It's basically a reverse Pascal's wager: if the current problems with LLMs are somehow solved and they become capable of consistently producing verifiable, neutral, and faithfully attributed information on any subject, Wikipedia will become totally unnecessary, since our goal of "a world in which every single person on the planet is given free access to the sum of all human knowledge" will be achieved for everyone who has ActuallyGoodChatBot.AI on their phone. If that doesn't happen, then it's really important that we preserve human-written resources for use after the AI bubble pops. Either way, the correct move for us as a community is to continue to curate quality human-written content. -- LWG ^{talk (VOPOV)} 16:25, 25 April 2026 (UTC)

Well, the good ting is that it's going to become impossible for LLMs to truely replace Wikipedia in that case - iff for no other reason that, as much as we're biased towards content availible over computers and the internet, they're even more so. If I'm reading old newspapers or books that haven't been digitized, then, well, the LLM just can't compete. Or even if the majority of the information isn't easily scrapable from the web - I've been making fun of the Google AI overview a lot in the past few days, as I'm trying to write about the histories of some of our totem poles. What with the inconsistent transliterations, constant name changes, the the fact that GNG coverage for these mostly exists in print. (Mind you, I'm not that worried in general. I think machine learning is really cool, and really useful to Wikipedia in some formats, Clubebot being the obvious one, and just the fact that chatbots like ChatGPT are able to create such human sounding text is, genuinely, super cool. But, well, while the chatbot-powered spam isn't going away any time soon, I'm with you that they don't quite have the transformative effect people want.)GreenLipstickLesbian💌🧸 18:15, 25 April 2026 (UTC)

@Athanelar, I’ve been meaning to write an essay at Wikipedia:Written for humans, by humans that covers the arguments in favour of an ideological ban if you fancy it? I’m just wary of it becoming an ideological rant that repels people rather than convinces them Kowal2701 (talk, contribs) 17:07, 25 April 2026 (UTC)

I'd be happy to contribute. (If it 'repels' the pro-AI crowd I'd consider that a plus...) Athanelar (talk) 18:27, 25 April 2026 (UTC)

We should take an anthropocentric position. Wikipedia; by humans, for humans, full stop.

We could have done that in 2023, but there is so much of a backlog of AI text already on Wikipedia -- and no doubt a lot of AI text still undetected -- that this would be falsely advertising the project to readers. I know this is meant to be a "should" statement but readers are going to interpret it as an "is" statement. Gnomingstuff (talk) 19:48, 25 April 2026 (UTC)

Sure, but if we make clear in no uncertain terms that we're fighting against that AI content rather than embracing it, I still think it's desirable to express our position that way. Athanelar (talk) 20:20, 25 April 2026 (UTC)

Same could be said for Neutrality or Verifiability or No Undisclosed Paid editing or even Free (since we can never be sure all copyvio material has been purged). All Wikipedia ideals are targets we move towards iteratively. For high-importance articles that are actively patrolled, I think we can fairly confidently put eyes on the diff between 2022 and now and clear them of AI content if necessary. I know I've started checking any high-byte-count edit on my watchlist feed for AI signs and reverting if I see them. And as new page patrollers get savvier about spotting AI signs I think we can catch a lot of low-visibility articles on arrival. -- LWG ^{talk (VOPOV)} 20:26, 25 April 2026 (UTC)

I've started a first draft of an essay regarding human-written Wikipedia, and I put it for a discussion at Wikipedia talk:WikiProject AI Cleanup#First draft of an essay on human-written Wikipedia. I'd like to hear this conversations' thoughts regarding this draft. guninvalid (talk) 20:15, 25 April 2026 (UTC)

Non-articles

I created Uhr as a disambiguation page using Gemini. The only changes I made were swapping "Uhr" and "UHR" in the first line; removing a space which was incorrectly introduced; and adding disambiguation page categories.

According to WP:ARTICLE, disambiguation pages are not counted as articles. Therefore, my writing of a disambiguation page with an LLM does not violate the WP:LLM guideline. Is my interpretation correct? feminist🩸 (talk) 14:34, 13 May 2026 (UTC)

I sent it to AfD for participants there to decide. SuperPianoMan9167 (talk) 14:45, 13 May 2026 (UTC)

Noting that WP:ARTICLE is an informational page; however the MOS is a guideline, and MOS:DAB states in the second paragraph, right near the top in prominent placement: Note that even though most disambiguation pages are kept in the Article namespace (mainspace), they are not articles; rather, they are aids in searching for articles. ⇒SWATJester ^{Shoot Blues, Tell VileRat!} 01:32, 23 May 2026 (UTC)

Use case for AI with respect to error-filled content that will not otherwise be fixed

St. James Church, Puthenkadai has been tagged as containing confusing writing since 2022. In particular, the "History" section of that article currently reads:

AD 16th century in the initial period of Christian regional work for the disciples in coastal areas and work bulk of the Christian religion, By the embracing of the Name "Protector Of Fisher Men", They considered the St. James as a Main god to worship.,The people Called "CHAVAKIZHAR" On set of terminating the daily works Build up The Small church which is made up of sand. As the days pass over the number of Christians increasing in the area a big society of Christianity made over in this society of groupism Village. As By the info of this village initially after the standardity it has separated from the head Church of "Muzhagumoodu".

All of this is sourced to a permanent dead link, http://www.jamesputhenkadai.com/.

As an experiment, I put this into one of the standard AIs with an instruction to clean up the text; since the AI appeared to infer some facts not clearly stated in the paragraph (although it is a stretch to say that anything is clearly stated in the paragraph), I re-prompted it to clean up the text without assuming any any facts beyond what is presented in the section. It provided the following:

In the 16th century AD, during the early period of Christian missionary activity in the coastal regions, many people embraced Christianity under the name "Protector of Fishermen" and regarded Saint James as their principal figure of worship. The people known as the "Chavakizhar", after completing their daily work, built a small church made of sand. As the Christian population in the area increased over time, a larger Christian community developed in the village. According to local information, the village church was initially affiliated with the head church at Muzhagumoodu before later becoming separated from it.

Now, I do see some problems with this, since the intent is almost certainly that they considered Saint James (specifically James the Great) to be the protector of fishermen, and not Christianity as a whole, and I would bet that the original church was made of sandstone, not actual sand. However, I think the prompted text is an order of magnitude better than the original, and there are plenty more examples of this sort of writing in Wikipedia. In fact, I think it would be a boon to have a built-in functionality that would allow any editor to highlight poorly written text and get suggestions for improved versions of that section. BD2412 T 00:48, 30 May 2026 (UTC)

Your text is not "an order of magnitude better than the original"; it's just written in prettier English. The actual issues with the text remain woefully unaddressed and there are severe issues with source-text integrity. For the obvious starters, it's a large section of text sourced to a website belonging to the actual church (independent sources, anybody?) w/ historical claims about the development of religious group in Southeast Asia. In the 1500s. This alone should be raising alarm bells. And the source text integrity... well, I don't speak Tamil, but looking at an archived copy of the website and putting it through Google Translate, it's not looking promising.

If an AI tool just encourages editors to mask problematic articles like this, then no thank you. GreenLipstickLesbian💌🧸 01:21, 30 May 2026 (UTC)

I don't disagree with your source assessment, but that is orthogonal to the issue. It is trivially easy to find items of text with no sourcing concerns, but with comparable writing problems (random capitalization, punctuation and spacing errors, typos, and grammatical issues, all folded into one). Just go through recently added foreign films and look at their plot summaries. BD2412 T 01:24, 30 May 2026 (UTC)

For typos, capitalization, spacing errors and the like, you can use a spellcheck to your hearts content. These are installed in pretty much every computer device at this point, I don't entirely care if they use an LLM or not. Grammatical errors need somebody who can actually read and analyze the sources, to make sure that their grammar fixes are an accurate reflection of the sources. If you, as an individual, want to use an LLM to flag those for you, then I'm not stopping you, but too many times I've had to deal with "fixes" people made to ungrammatical sentences which completely changed their meaning in a way that was not reflected in the sources; if you are not willing or capable to do that (as an example, I not capable of fact checking the nuances of a text against a source in Tamil), then you can leave it to somebody else who is. A tool like this would just set people up for failure. GreenLipstickLesbian💌🧸 01:30, 30 May 2026 (UTC)

I have yet to see this extreme of writing problems without corresponding sourcing issues Drew Stanley (talk) 02:05, 1 June 2026 (UTC)

Exactly. I don't think I have ever seen prose this bad in an article that doesn't also have sourcing, NPOV, or some kind of other content issue. It's usually a WP:CIR problem at root, which the AI is only going to put a pretty bandage on. Cremastra (talk · contribs) 14:58, 5 June 2026 (UTC)

Frankly, I don't see any other editors carrying out the equivalent of my project to rigorously find and fix all of the hundreds of thousands of minor punctuation spacing errors in the encyclopedia, and I therefore doubt that anyone else is paying that much attention to these things. Had I not flagged this passage in this discussion, I doubt the page would ever have been PRODded; rather, it would have just sat there more or less forever looking just as it does, as it has for the past seventeen fifteen years. BD2412 T 17:13, 5 June 2026 (UTC)

Lots of people do care; I personally spend a lot of my time cleaning up old copyrighr issues, NPOV issues, BLP issues, and so on.

However, when I'm at my wit's end, I'm dealing with somebody who has been banned for repeated LLM use, plagiarism, synth, OR, ecetera, and I get an admin showing up at AN to say that enforcing WP:BANREVERT through G5 is abusive of deletion privileges, then, um.... Okay, I'm grateful for your AWB edits. Genuinely. But bringing up your AWB project to fix "thousands of minor punctuation spacing errors" as some form of evidence that you're the only one who cares, while actively accusing people who try to clean up more severe issues before they've hung around for 17 years of abuse, and advocating that people use AI to prettify articles, not check for source-text integrity, is insulting.

Readers can live with typos and "minor punctuation spacing errors". Readers shouldn't have to live with blatant misinformation and NPOV-riddled articles. GreenLipstickLesbian💌🧸 18:29, 5 June 2026 (UTC)

It is, however, possible for an article to have typos that need fixing and still be an otherwise accurate and informative article. BD2412 T 18:42, 5 June 2026 (UTC)

I know; Again, I'm genuinely grateful for you for fixing all those issues, and I'm happy that we don't have to live in a Wikipedia where we chose between fixing prose issues and fixing content issues. But please don't make it harder for people to fix those content issues -- part of the reason that articles like that stick around for so long is not because people don't care, but because we have large backlogs and because, when you act like you did at that AN/I, you really discourage people from trying to fix them. GreenLipstickLesbian💌🧸 19:01, 5 June 2026 (UTC)

My point in the ANI was about reflexive and unthinking mass deletion of content based solely on a ban, some of which is good content which leaves gaps in projects when deleted without review. BD2412 T 20:39, 5 June 2026 (UTC)

"Unthinking". Well, that's a step up from "abusive", I think.

Either way, your statement here is incompatible with your above I therefore doubt that anyone else is paying that much attention to these things; people are paying attention to these sorts of articles, you know they are, you just refer to the community's attempts to clean up these articles as "unthinking" and "abusive". GreenLipstickLesbian💌🧸 21:01, 5 June 2026 (UTC)

The ANI discussion was solely about rote deletion of article by blocked/banned users, which does not encompass this article, or most articles that I have seen that have these kinds of problems. The things that get users banned are often oddly orthogonal to the things that make users bad writers on a technical level. Had this article been written by a banned user, it would probably have been deleted a long time ago, but it wasn't, so here it sits in this incredibly bad shape for fifteen years. BD2412 T 21:12, 5 June 2026 (UTC)

We should not have a built in functionality to fuzzily guess at the meaning of text we don't understand. It's said that it's easier to write a FA from scratch than to try and adapt existing text. That's not because the existing text has grammatical or spelling issues, but due to issues such as sourcing and due weight. CMD (talk) 03:26, 30 May 2026 (UTC)

This requires far more than a minor grammar fix. No LLM, or human editor for that matter, should ever try to ‘improve’ that sort of thing by guessing at unclear meaning. That simply adds a layer of unsourced speculation, masking the unsourceable rubbish. Regardless of the reliability or otherwise of the original source, it’s clear that the editor was not competent to summarise it. The section contains nothing salvageable, actively harms the encyclopedia, and should be deleted. MichaelMaggs (talk) 04:54, 30 May 2026 (UTC)

I agree with the "masking" term that has come up several times above. The original has the advantage of being recognizably flawed, and the revision could be mistaken for more reliable material. Ability to compose coherent text at WP:CIR level and ability to cite material at WP:CIR level are correlated. Dekimasuよ! 05:34, 1 June 2026 (UTC)

I mean, no:

• "Protector of Fisher Men" is almost certainly referring to St. James, given that St. James was a fisherman. The LLM twisted it into referring to Christianity as a whole, which is nonsensical.
• "A main god to worship," with the indefinite article, leaves open the possibility that St. James is one of many important gods they worshiped. "Their principal figure of worship" is not only wordier but implies that there is only one.
• Admittedly I don't know what the second sentence is supposed to mean, but I highly doubt that The people known as the "Chavakizhar", after completing their daily work, built a small church made of sand. is it.

Gnomingstuff (talk) 06:57, 5 June 2026 (UTC)

I put this into one of the standard AIs Which one? ChatGPT? Gemini? Claude? Grok? DeepSeek? Something else?

Also, please say "LLMs". SuperPianoMan9167 (talk) 18:40, 5 June 2026 (UTC)

The test for purposes of the discussion was done with ChatGPT. Obviously, I did not add the content to the article. I noted in my post that there are things the LLM came up with that are likely wrong, including the way St. James is referenced and the church more likely being made of sandstone, but those are easily fixable. Look, I'm a law professor, and I am teaching a generation of students who will, within the next year or two, be going to work for law firms that now require new associates to use AI (I don't know that it will just be LLMs) to expedite their work. I have an obligation to understand how my students will be using these platforms during the semester and in practice. If I have a student who turns in work filled with typos and grammar/punctuation/spelling errors, that's a good sign that they have not used such a platform, but that quality of work will also receive a lower grade. BD2412 T 21:08, 5 June 2026 (UTC)

I don't know that it will just be LLMs Because of marketing, when the average person says "AI" nowadays, they usually mean "LLM-powered chatbot". SuperPianoMan9167 (talk) 21:16, 5 June 2026 (UTC)

I'm aware of that, but large language models are not the only models currently in development, and I'm looking a year or two down the road for students entering the workforce. BD2412 T 21:18, 5 June 2026 (UTC)

large language models are not the only models currently in development Yes, because there's also image generation, video generation, music generation, ...

Unfortunately, the generative AI hype machine has pretty much squandered any other kind of machine learning research for the past four years. SuperPianoMan9167 (talk) 21:25, 5 June 2026 (UTC)

Also, which model did you use in ChatGPT? I would assume GPT-5.5. SuperPianoMan9167 (talk) 21:19, 5 June 2026 (UTC)

That is correct. When I understood that I would need to evaluate the capacity of the platform on a continuing basis, I subscribed. I have experimented with all the platforms I have heard of, to some extent. BD2412 T 21:29, 5 June 2026 (UTC)

By the way, just for fun, I asked it whether this article should be deleted from Wikipedia, and it said "probably, yes" and gave the lack of sourcing, notability, and poor translation issues as reasons. BD2412 T 21:31, 5 June 2026 (UTC)

Looking at Women Looking at War

Can this be redirected as an WP:AtD? Bearian (talk) 04:41, 3 June 2026 (UTC)

I think that's a regular PROD, not an LLM PROD, so the answer should be yes. Even if it were an LLM PROD, I think redirecting would not violate the spirit of the rules. Toadspike [Talk] 09:53, 3 June 2026 (UTC)