Requests for comment/Artificial intelligence policy

This is a subpage; for more information, see the Requests for comments page.

---

In the last few years, many Wikimedia projects have developed individual policies and guidelines regarding the use of artificial intelligence, either to generate new content or in discussions between users (see Artificial intelligence/Policies by project). While varied and adapted to the specific needs of each project, these policies share commonalities revealing baseline expectations: that artificial intelligence may be helpful for auxiliary tasks, but must at least be used with caution, and ideally transparency, regarding matters such as content generation. Formalizing these as shared goals will minimally impact existing policies, while providing us with a common ground to rely on in the future.

While many large Wikimedia projects have had the opportunity to develop these policies, most smaller ones haven't yet. This leaves them especially vulnerable to AI disruption, which compounds with the existing lack of manpower on many projects. An opt-out policy would provide these projects with the policy tools they need, while still offering the opportunity to be refined for their specific use cases. To respect the choices of individual projects, it will not apply to those with an existing AI policy or guideline, or that previously voted against such a policy.

The baseline policy and the opt-out policy are independent of each other, and may be voted on separately. The specifics of each, including the definitions used, can be found at Artificial intelligence/Draft policy. Chaotic Enby (talk) 18:13, 23 April 2026 (UTC)

Baseline policy

Should the following be adopted as a global policy?

Content

The use of AI models to generate or rewrite content must be clearly disclosed, except for translation and basic copyediting. Any AI-generated content must be subject to human review before publication. Editors using AI for this purpose take full responsibility for their edits.

Discussions

The use of AI models to generate comments in discussions must be clearly disclosed, except for translation, grammar correction, and language refinement. Any AI-generated comment must be subject to human review before publication. Editors using AI for this purpose take full responsibility for their comments.

Discussion

•  Support as proposer. Chaotic Enby (talk) 18:13, 23 April 2026 (UTC)
•  Support per above. AI bubbles are about to be popped out soon. So a universal policy could help small wikis to be active and not passive. Ahri Boy (talk) 05:42, 24 April 2026 (UTC)
•  Comment is it possible to publicise this such that we get input from small wikis? A lot of the larger wikis have created strong restrictions on LLM use over the past year or so, and I think there's an assumption smaller wikis just haven't had the resources/impetus to follow suit, but there might be positive reasons like a more permissible attitude towards LLMs due to the need to scale Kowal2701 (talk) 15:29, 25 April 2026 (UTC)

  Publicizing it would be great! I'm not especially familiar with publicizing global RfCs on individual wikis, but please feel free to do so! Chaotic Enby (talk) 15:31, 25 April 2026 (UTC)

  It's probably easiest to do a MassMessage, there's Distribution list/Global message delivery. Even if this is unsuccessful, it'd be very valuable for gathering global perspectives on this for a future proposal Kowal2701 (talk) 15:48, 25 April 2026 (UTC)

  Filed a request at Meta:Requests for help from a sysop or bureaucrat#Mass message request for Requests for comment/Artificial intelligence policy! Chaotic Enby (talk) 16:10, 25 April 2026 (UTC)

•  Oppose on the basis that this would be an imposition onto local communities who may wish to establish different policies. We should respect the full autonomy of local communities, and if they wish to allow undisclosed LLM usage, they should be allowed to, even if that's a policy I would personally disagree with. — chrs [talk] 16:29, 25 April 2026 (UTC)
•  Support – as a draft contributor. – Aca (talk) 01:05, 26 April 2026 (UTC)
•  Oppose as an unreasonable intrusion on the autonomy of individual projects. Extraordinary Writ (talk) 01:08, 26 April 2026 (UTC)
•  Oppose as worded – in certain projects even translation and basic copy editing need to be disclosed. In addition, in editing circles it’s well known that current AI models cannot be trusted to perform copy editing.Alıƨsi (talk)
•  Support -- Infrogmation (talk) 01:14, 26 April 2026 (UTC)
•  Support It makes sense that AI usage should be disclosed. EdoAug (talk) 01:20, 26 April 2026 (UTC)
•  Support -- Jeff G. ツ (please ping or talk to me) 01:21, 26 April 2026 (UTC)
•  Support as a default policy for all communities, especially for smaller ones, but we should allow individual communities to override should they accept a looser or stricter policy. 1F616EMO (on zhwiki) 01:48, 26 April 2026 (UTC)
• Strongly opposed as written, even though the intention behind is good. As it is written now, it means that machine translation doesn't even have to be disclosed. It must be the other way around: All LLM-generated content must be subjected to human review, including translation. --Amir E. Aharoni (talk) 02:01, 26 April 2026 (UTC)

  This is a baseline policy, i.e. intended to be the minimum everyone can agree with, which is why it is deliberately very lightweight. Many projects already have much stricter policies, and it shouldn't be seen as a cap on what is allowed or expected. Chaotic Enby (talk) 02:09, 26 April 2026 (UTC)

  Well, I cannot agree with it. As written, it explicitly allows machine translation. Vandals will use it to publish machine translations in languages with few or no people to enforce it, and even the global sysops may be unable to enforce it, because the vandals will say that they don't even have to disclose machine translation. Amir E. Aharoni (talk) 02:22, 26 April 2026 (UTC)

•  Comment -- there are use cases inside wikimedia projects where requirement of disclosing use of AI is not useful, because use of AI is a very common tool used when users are doing these and disclosing requirement would just add deadweight. Say editing CSS, editing templates or Lua-modules, SPARQL-queries, basic coding tasks etc. --Zache (talk) 02:05, 26 April 2026 (UTC)

  This policy is explicitly for content and discussions (as defined here), and deliberately avoids topics such as coding where the dynamics of AI use can be very different. Chaotic Enby (talk) 02:07, 26 April 2026 (UTC)

  @Chaotic Enby: What about on Wikifunctions, where the reader-facing material is code? JJP_Master (she/they) 02:09, 26 April 2026 (UTC)

  Wikifunctions doesn't have reader-facing material per se, instead providing tooling for other projects (such as Abstract Wikipedia, in which these functions can then be assembled into content), so I don't envision it being affected at all. Chaotic Enby (talk) 02:12, 26 April 2026 (UTC)

  Wikidata and Structured data on Commons would be clearer examples than Abstract Wikipedia. --Zache (talk) 04:25, 26 April 2026 (UTC)

  LLM use on Wikidata absolutely must be disclosed, though I don’t know how. Same with Commons if you’re talking about translation. Alıƨsi (talk) 08:04, 1 May 2026 (UTC)

  I think something simple like P887 is sufficient for that. Well very well (talk) 08:14, 2 May 2026 (UTC)

•  Oppose. as user Al12si said, translation should also be disclosed. The case of some users translating full pages and doing lazy reviewing, if any, will be very real. Genoskill (talk) 02:16, 26 April 2026 (UTC)
•  Support because knowledge is human. Best, HouseBlaster (talk • he/they) 02:50, 26 April 2026 (UTC)
•  Support. Yamato Shiya (talk) 03:09, 26 April 2026 (UTC)
•  Oppose That's far too weak. As in German Wikipedia, AI-generated content shouldn't be allowed at all. Chaddy (talk) 03:28, 26 April 2026 (UTC)

  Nothing stops German Wikipedia from banning it altogether? Trade (talk) 11:10, 26 April 2026 (UTC)

  That's not the point. I want such a ban globally. Chaddy (talk) 16:07, 26 April 2026 (UTC)

  I would agree with you. Globetrotter30 (talk) 16:11, 26 April 2026 (UTC)

•  Support On the one hand, I understand for those people who don't have English as their first language LLM may seem like an easy way for them to get an article in good form for Wiki, it really isn't, because English isn't their first language, and they may not know that the article contains things that are outright wrong. Therefore I support this policy as it's pretty much the same thing as the bots policy. The user takes responsibility for the action taken on wiki. W.K.W.W.K ...Straight outta Underverse 03:55, 26 April 2026 (UTC)
•  Comment several remarks:
  • This should certainly say explicitly that it is a minimum, and that individual projects may make stronger rules.
  • Also: this is very confusing as to the status of AI translation in discussions. AI translation is called out explicitly for content, but there is no mention of it with reference to discussions.
  • The rule about what is required by way of disclosing AI translation for content is unclear. For example, when writing Spanish (in which I am just short of fluent) I will often check to see if Google Translate might hit on a more felicitous word than I would. Am I expected to disclose explicitly each time I do that? I suspect many contributors have a similar practice, and few (if any) are currently explicit about it.
  • Also: nothing here seems to address AI generation of images or multimedia. - Jmabel (talk) 04:07, 26 April 2026 (UTC)

    That's an oversight on my part, sorry. I meant to include the exemption on translation for discussions, just like for content editing. Chaotic Enby (talk) 04:16, 26 April 2026 (UTC)

•  Oppose for now per Amir Aharoni. If this becomes stronger, I could support. Abzeronow (talk) 04:39, 26 April 2026 (UTC)
•  Support --ELexikon (talk) 05:04, 26 April 2026 (UTC)
•  Support Grueslayer (talk) 06:25, 26 April 2026 (UTC) Mandatory disclosure still leaves it to local communities how to deal with the topic in a way they prefer. It's just a basis, a minimal consensus.
•  Support --𝓰𝓲𝓷𝓪𝓪𝓷기나ㅏㄴ(T/C) 06:43, 26 April 2026 (UTC)
•  Support It's not the best. Should be stronger and AI should be absolutely prohibited. But at least there is something. --Chelin (talk) 06:50, 26 April 2026 (UTC)
•  Oppose I agree with the general terms discussed here, but I don't think we need to impose that opinion on communities that may be thinking otherwise. There are many other uses of AI which are part of our workflows, for example OCR in Wikisource, noise reduction in Commons images or reconciliation in Wikidata. Humans should take care of the results, but if we impose limits somewhere, this can create problems in places we wouldn't think about. -Theklan (talk) 07:14, 26 April 2026 (UTC)

  Exactly, this is a very good point. That's why I don't recommend using the term "AI" at all. Words that are used in policies, rules, laws, etc., are supposed to be as unambiguous as possible, and the word "AI" is inherently ambiguous. If this policy is supposed to be for text generation or translation, then it should probably say "LLM", which is much more focused and well-defined, and "AI" can be mentioned in parentheses occasionally. Technologies for graphics, OCR, video that are often dubbed as "AI" are different, and there should be different policies for them, maybe allowing them and maybe disallowing them. And it is pretty much guaranteed that in a few years, some completely new thing will be labelled "AI" that we don't even imagine yet. The most important point in the whole discussion is "Editors [doing anything at all with or without AI] take full responsibility for their edits", Amir E. Aharoni (talk) 10:25, 26 April 2026 (UTC)

•  Support -- Sobaka (talk) 07:40, 26 April 2026 (UTC)
• I Strongly oppose the section allowing machine translation without disclosure, as per Amir E. Aharoni above. Support the rest of the proposal.--Naḥum (talk) 07:52, 26 April 2026 (UTC)
• Partly  Oppose, per Amir E. Aharoni and others: full disclosure if AI has been used at all. Jaap-073 (talk) 08:06, 26 April 2026 (UTC)
•  Strong oppose Per Amir E. Aharoni. Huñvreüs (talk) 08:28, 26 April 2026 (UTC)
•  Comment I'm inclined to support this as better than nothing, but I believe that all uses of AI must be disclosed, and that disclosure of the use of AI for "basic copy editing" is extremely important, because that goes to writing style. Ikan Kekek (talk) 08:48, 26 April 2026 (UTC)

  Exactly. Even the use or place of a comma can change the meaning of a sentence. Style and meaning are overlapping concepts. Jaap-073 (talk) 09:00, 26 April 2026 (UTC)

•  Strong oppose for the current version, as I think AI-generated translations should 100% be disclosed. It is very possible for LLMs to just hallucinate translations and terms, or write something else entirely. This is especially true for languages which are underrepresented in the training data, which are also the languages this proposal seeks to protect. Otherwise  Strong support. --QuickQuokka ^{[⁠talk • contribs]} 08:59, 26 April 2026 (UTC)
•  Oppose – AI translation is unacceptable in certain projects. tsca (talk) 09:15, 26 April 2026 (UTC)
•  Oppose Not really up to meta to make up content rules on the wikis.--Snævar (talk) 09:17, 26 April 2026 (UTC)
•  Oppose. I think many people would use this as an excuse to post low-quality content that someone else would then have to review. Only genuine content written by humans should be allowed. And this applies especially to comments: if don't want to think your own comments, you shouldn’t contribute to Wikipedia. At the very least, individual Wikipedia editions should decide for themselves. --PNNu (talk) 09:25, 26 April 2026 (UTC)
•  Oppose – The proposal implicitly assumes that editors and administrators are able to reliably identify AI-generated content. From my experience on the Dutch Wikipedia, this assumption does not hold. AI-generated text is often either obviously low-quality, in which case existing quality control mechanisms should suffice, or near-indistinguishable from human writing, while false positives based on misunderstandings about how LLM's work are common. This risks inconsistent enforcement (deletion? blocks?) based on subjective judgments, and shifts focus from content quality to difficult or impossible to verify claims about how the content was produced. Moreover, a disclosure requirement may create a false sense of security, where content without such a disclosure is assumed to not be AI-generated and therefore less scrutinized. Jeroen N (talk) 09:44, 26 April 2026 (UTC)

  Just a comment - LLM output isn't "near-indistinguishable" from human writing if you know what to look for. At least in English, LLM output is not only fairly easy to identify, but also fairly easy to date within ~6-12 months based solely on phrasing in the text (because GPT-4o sounds different from GPT-5, which sounds different from Claude, etc.)

  I don't know about Dutch specifically, but there's some research suggesting that LLM output in other languages "assumes" English under the hood, and anecdotally speaking I've seen the same phenomenon happen in French and German, where AI-generated content does the same things English does, enough that you can sometimes identify it even without speaking the language. (edit: and Esperanto apparently per a commenter below) Gnomingstuff (talk) 17:51, 27 April 2026 (UTC)

•  Support and translation should be included. Yann (talk) 09:50, 26 April 2026 (UTC)
•  Oppose This is IMO very similar to the case of UPE. Mandatory disclosure of all AI-writing is unenforceable just like paid-editing usually cannot be proved by volunteers. If an AI-generated text is flawed, then any sanctions should be taken because of these flaws (incidental or repeated, depending on the case). If AI-generated article is flawless, I don't see much of a difference from an ordinary human-written text. IMO, the rule about AI on our projects should be something like You're responsible for edits that you save and for ensuring they meet the project's principles. Being assisted by tools is not an excuse (no matter how advanced the tools are). (so, basically the two last sentences). Whatever above that is IMO either not enforeable or not future-proof. Msz2001 (talk) 10:26, 26 April 2026 (UTC)
•  Oppose Per Amir E. Aharoni. AI translations and AI copy editing are problematic. Projects that want such edits can write their own more permissive policy as easily as one can write a more strict policy. Machine translations have been a problem on some projects with few native speakers. It is also unclear how this policy explicitly on AI would relate to possibly more restrictive local policies not explicitly talking about use of AI. –LPfi (talk) 10:44, 26 April 2026 (UTC)

  I think I have an issue with the concept of a mandatory baseline based on what restrictions we might agree on being reasonable. If individual projects can override this, then the opt-out policy should suffice. If not, then this should be about what we as a movement cannot accept. I don't think the proposal is written with enough consideration of all possible corner cases for the latter use. Let's have the opt-out policy first and return to possible minimum standards when we know what actual problems cannot be handled by other means. –LPfi (talk) 11:15, 26 April 2026 (UTC)

• Simplify
  1. Use same wording for content and discussion
  2. Acknowledge AI translation
  3. Combine with opt-out policy
• For now, create a guideline instead of policy to stimulate discussion but allow project communities to more time discuss their local approach
  -- Jtneill - Talk 10:47, 26 April 2026 (UTC)

  I oppose posting AI-generated comments. Translation and copy editing your comments is mostly OK (while large-scale copy edits of articles is not), but I don't see why you would have an AI tell you what to tell. The examples I have seen (read: recognised) have been disruptive. An AI disclaimer would not help. –LPfi (talk) 10:52, 26 April 2026 (UTC)

  The one case I used AI for a comment was when I wrote out the comment, realized that it would be unhelpful, and had ChatGPT rewrite it to avoid driving away a new user. I lightly edited it before posting, and stand by it. I don't have AI write things for me because I have a clear voice, but sometimes I need someone to help me moderate.--Prosfilaes (talk) 03:21, 28 April 2026 (UTC)

  This is an excellent example of how AI may be of use to improve discussions. -- Jtneill - Talk 12:48, 1 May 2026 (UTC)

  I think that would fall under copyediting. From what I've seen, the only reason to allow AI generated comments is to allow for AI based bots. Like if en:User:Fermiboson/AIlog was automated. MetalBreaksAndBends (talk) 16:25, 1 May 2026 (UTC)

•  Oppose I have had issues with auto-translated content, where there was hallucinated information coming out of nowhere. Translations should also be covered by the policy. — IмSтevan ^talk 11:02, 26 April 2026 (UTC)
•  Support --Saroj (talk) 12:22, 26 April 2026 (UTC)
•  Oppose per Amire80. Naut-rena (talk) 13:04, 26 April 2026 (UTC)
•  Comment I agree with the above discussion regarding the listed exclusions - every use of generative AI should be documented as a baseline. However I think there is a more fundamental issue here: the definitions list on the linked page does not define "AI"! Are well-calibrated statistical models for example for OCR or machine translation, or automation tools for generating content that are not "LLM's" excluded? Some of those are surely fine, some also probably not. On Wikidata there has been for many years complaints about the 'cebwiki' (and some others) where, long before generative AI, somebody decided to run a bot that generated articles in that language for every entry in GeoNames (I think), resulting in massive numbers of duplicated items in Wikidata based on those sitelinks. What we are trying to tackle with policies like this is, I believe, avoiding creation of content unreviewed and beyond the capacity of human editors to monitor and manage. Generative AI tools make it far too easy to create vast amounts of content that swamp the capacity of us humans. I think these policies for disclosure and potential blocking should have that clear background justification, making clear that Wikimedia sites are fundamentally by and for humans. ArthurPSmith (talk) 13:24, 26 April 2026 (UTC)
•  Oppose AI translation is de bane of all small Wikipedias. It should be completely banned. Ieneach fan 'e Esk (talk) 13:26, 26 April 2026 (UTC)
•  Oppose. Long story short: Never disclose anything; forbid unsupervised AI. A human who publishes content always takes responsibility for it. Supervised AI can be used on very different degrees, which makes it pointless to create an abstract rule about disclosure (e.g. one can simply use AI to rewrite a list in alphabetical order: would that need to be disclosed?). On the other hand, we should forbid unsupervised AI (not simply disclose it). --Grufo (talk) 13:29, 26 April 2026 (UTC)
•  Oppose. As noted in the discussion, the problem isn't the tool's use, but the errors it creates. Mandatory labeling of text or comments in a discussion as AI-generated is meaningless; editors are already fully responsible for their actions. Some communities may adopt a more stringent rhetoric and force editors to make such labels. However, as a global policy, this is unacceptable. Лоття (talk) 13:34, 26 April 2026 (UTC)
•  Oppose per Amir E. Aharoni and Лоття. -- Futbollo (talk) 13:43, 26 April 2026 (UTC)
•  Strongly oppose per others. Redmin (talk) 13:55, 26 April 2026 (UTC)
•  Weak support I'm not sure the approach/concept of a global policy decided by people not part of these projects outside of these projects is legitimate/valid but if it is or to the extent it is, the proposed policy seems quite reasonable and useful – what constitutes as 'clearly disclosed' is up to the communities. The approach of global policies like this means the policies of communities and local/regional projects will be biased toward the demography of the global movement (their views). Ultimately this means that policies will be biased toward what happens to be hip in the US today. I'm also concerned about the deliberation process used that's not streamlined for rational nuanced outcomes&deliberation. Often people have strong opinions about things they don't know much about or that simply lack nuance etc. So basically, I oppose that global policies can be made by people not part of these projects, and with possibly quite different conclusions/needs/views, but at the same time this policy makes sense and ultimately I decided to weakly support this as these communities could still opt out of this rough baseline policy via a local policy. --Prototyperspective (talk) 14:00, 26 April 2026 (UTC)
•  Oppose Each individual project should be able to govern their own use of AI/LLM tools.
  Not listed on the "Policies by project" page is English Wikinews' proposed policy: n:WN:AI, which essentially mirrors this proposal in that humans are ultimately responsible—both contributors and reviewers—and no unsupervised generation, i.e., no fully-AI-generated articles were allowed. Michael.C.Wright (talk) 14:13, 26 April 2026 (UTC)
•  Oppose -- XXBlackburnXx (talk) 14:21, 26 April 2026 (UTC)
•  Support I believe that this is necessary. AI should not be used in a wiki, but may be helpful on encyclopedias for finding correct sources. --Globetrotter30 (talk) 14:40, 26 April 2026 (UTC)
• Support. Sapphaline (talk) 15:54, 26 April 2026 (UTC)
•  Comment I wonder, what might be the ways to disclose it. Weather you are fine disclossing it in edit summary, or it should be disclowsed on the page itself. What would be usefull to track ai generated content with a category or tag. The problem of AI generated content is not in quality but in velocity, were traditional mechanisms may fail to check such pages. So sorting them by tags or in cats would be usefull.--Juandev (talk) 17:10, 26 April 2026 (UTC)

  At en.WN, we used both categories and templates:

  • n:Category:AI usage declared
  • n:Template:AI usage (talk pages only, used both manually and automated, via SubmitWizard)
  • n:Template:AI disclosure (proposed, never put into production)
  • n:Wikinews:The use of AI in generating content

  Michael.C.Wright (talk) 20:21, 26 April 2026 (UTC)

•  Strongly oppose It is nonsense to think that if there is human control over the use of AI models, their use is okay. Moreover, there are children freely editing here. We literally have millions of entries, and millions more are being added. It is therefore impossible to assess the quality of AI-generated content. If you allow the use of AI models for writing content or translating, the reader might as well skip Wikipedia. Furthermore per user Jeroen N. Happytravels (talk) 17:28, 26 April 2026 (UTC)
•  Oppose – While I broadly agree with parts of the proposal, I don't support the underlying sentiment against AI assistance. Automation, including AI, is one of the few practical ways smaller wikis can keep pace with larger ones despite limited contributors. In workshops I’ve participated in, younger contributors consistently identify AI-assisted workflows as a top priority, and in my own admin work on my home wiki I rely on these tools for everything from interface improvements to bug fixes and content cleanup. Misuse, such as blind copy-pasting, should be addressed case by case, as with any low-quality edits. More importantly, this approach seems to run counter to Wikipedia’s original open model, which deliberately lowered barriers to participation compared to other more formal academical projects or even online ones like w:en:Nupedia. That openness enabled its growth, and discouraging AI risks moving in the opposite direction; the focus should instead be on better integrating these tools on its current workflows or on developing our own AI tools. - Klein Muçi (talk) 17:36, 26 April 2026 (UTC)
•  Oppose - this should be decided on a per-wiki basis. BrokenSegue 18:02, 26 April 2026 (UTC)
•  Oppose - the use of AI models should be disclosed without exceptions, including translation, copyediting, etc. Deltaspace42 (talk) 18:33, 26 April 2026 (UTC)
•  Support Nizzan Cohen (talk) 18:42, 26 April 2026 (UTC)
•  Support can be helpful for revisors on litte-documented language wikis to distinguish specialized words or text in a niche dialect from possible AI-hallucination. With raising AI quality, the need for translation transparency seems language-specific, so no global rule is ok. Krol111 (talk) 19:52, 26 April 2026 (UTC)
• I though hard about this... but I  Oppose. Bad actors won't care about this rule, and decent people would have to disclose using AI and face criticism or don't use AI at all for searching and improving edits. So this might make good people use worse tools. Furthermore most of the time, we won't be able to tell when AI was used, so this would be a dead letter law / unenforced law (I did use AI to check the idiom, how does that make you feel?). Some here even say any use at all would have to be disclosed. Really? Google Translate is AI, OCR is AI, heck, even your spam filters have been AI for the whole current century (and more). Dictionaries are slow to use, and hand-filtering your emails would be a nightmare. Almost nobody has been copying book text by hand for ages (you even have built in OCR tools in Wikisource). AI has become more advanced recently, but let's not freak out about it, please. On plwiki, we do have some abuse filters that hint when an LLM was used to generate parts of a page – we tag those edits. Edits we tag as AI should be double-checked for bad sources and nonsense information, but it might be fine and in many cases it is fine-ish. I mean if a novice user used their knowledge of facts and used AI to conform with our lengthy rules and guidelines... then I'm actually mostly OK with that. I must admit I'm annoyed by people adding AI nonsense or AI slop to wikis, but I'm not sure if the rule proposed here would help to minimise that. --Nux (talk) 20:07, 26 April 2026 (UTC)
•  Support (strong) Accipiter Gentilis Q. (talk) 20:54, 26 April 2026 (UTC)
•  Support baseline policy. It provides essential protection for small projects without overriding existing local policies. Mirzali (talk) 21:10, 26 April 2026 (UTC)
•  Oppose According to this proposal, anything AI goes, provided it is disclosed. With the loophole for 'basic editing' the proposal is even weaker than the otherwise similar policy for paid editing. And we all know how that pans out. If I wish to read AI text on a topic, I talk to ChatGPT. -<(kmk)>- (talk) 21:27, 26 April 2026 (UTC)
• Strongly opposed as written. Any use of AI should be disclosed no matter the reason. No content-related post written with AI should be published in the open without a thourough check by users (if possible, this could be done by users who have been attributed the role of "AI-content controllers" by their peers on each project following a vote). If some parts, but not all, of an article has been written with AI, the AI "watermark" should be clearly visible at the beginning of the page with the mention of the sections that are at risk of AI hallucinations. Treehill (talk) 21:28, 26 April 2026 (UTC)
• Strongly opposed. AI has no conscience and no responsibility. So I support keeping the project purely human. AI makes different mistakes compared to living people, and thus the errors can stay alive. EDIT: Adding my own (sorry): He must have the human touch in what we write. The simple calculations are just good to make with computer, but the AI will add his own opinions if taught. EDIT II As I say, please, translate from my user page from Finnish, or let the AI try to do that. "Tekoälyhän meillä kaikilla on (paitsi uskovaisilla)."--Höyhens (talk) 22:03, 26 April 2026 (UTC)
•  Comment I find the exemption on disclosing AI use for translations questionable in the context of translations made for Wikisource. ~2026-25529-47 (talk) 22:49, 26 April 2026 (UTC)
•  Comment All use of AI within Wikipedia must be banned for now. --Sinucep (talk) 23:11, 26 April 2026 (UTC)
•  Oppose --Qualitätssicherung (talk) 23:34, 26 April 2026 (UTC)
•  Support any method of clearing out ai. ltbdl (talk) 00:29, 27 April 2026 (UTC)
•  Support Andrei Stroe (talk) 08:04, 27 April 2026 (UTC)
•  Oppose The harm caused by the ContentTranslate tool, which the WMF actively imposed even on communities that had voted to ban it, is a very good reason to oppose the undisclosed use of AI for translation and copy-editing.--Leptictidium (talk) 09:33, 27 April 2026 (UTC)
•  Oppose because not strict enough. The use of IA in discussions is a no-go for me. I don't see the point communicating with a guessing words machine. Translation could be an exception. The use of GenIA in articles should be completely banned. We are creating human fact-based knowledge, not an approximative-kinda-looks-probably-true-if-the-machine-says-so knowledge. --M0tty (talk) 10:01, 27 April 2026 (UTC)
•  Support Esperanto Wikipedia is at risk of being overwhelmed by uncited AI-generated content, often containing dubious generalizations (e.g. about what the article topic "is best known for," "represents" etc.). Most of this is created by one well-meaning but misguided user with the complacency of admins who don't seem to see it as a serious problem. Eowiki has always been lax in practice about verifiability, but presumably most uncited content has actually been based on a source - AI drastically increases the probability that it isn't.
  A stricter policy might be desirable but I don't see how that is grounds for opposing - rejecting light protections doesn't get us heavier protections, it leaves us with no protections.
  Por: La Esperanta Vikipedio estas en danĝero de drono en senreferenca AI-kreita enhavo, ofte enhavanta dubindajn ĝeneralaĵojn (ekz. asertanta pro kio la artikoltemo estas "plej konata", kion ĝi "reprezentas" ktp.) La plejparto estas kreata de unu bonintenca sed missupoza uzanto kun la permesemo de administrantoj, kiuj ŝajne ne rigardas ĝin kiel seriozan problemon. Nia vikipedio ĉiam malstriktis en praktiko pri kontrolebleco, sed supozeble la plejparto de la senreferenca enhavo reale baziĝis sur fonto - AI draste pliigas la ŝancon ke ne.
  Pli strikta politiko eble estas dezirinda sed mi ne vidas kiel tio estas kialo por kontraŭo - rifuzo de malstriktaj protektoj ne havigas pli striktajn, ĝi lasas nin senprotektaj. --Arbarulo (talk) 13:51, 27 April 2026 (UTC)
•  Oppose AI usage has the risk of generating content which are full of hallucination and misleading, and AI usage for discussion will generate content that sounds too unsincere. Hakimi97 (talk) 14:09, 27 April 2026 (UTC)
•  Oppose on the ground that other usages such as translation should also be disclosed, as AI in languages other than English were often not as matured, and the global policy should take account of this. —— Eric Liu_（Talk） 15:49, 27 April 2026 (UTC)
•  Strong oppose Per Alıƨsi and Amir E. Aharoni. I work in localization, I know first-hand how rubbish -- and sneakily so -- AI translation can be. ANY use of AI to generate article / entry content MUST BE DISCLOSED, as far as I'm concerned. ‑‑ Eiríkr Útlendi │^{Tala við mig} 18:04, 27 April 2026 (UTC)

As rubbish as w:English as She Is Spoke? The perfect is the enemy of the good, and AI and machine translation are a good bit better than a lot of human translation.--Prosfilaes (talk) 03:32, 28 April 2026 (UTC)

My concern is not that editors will use AI -- use the tools available, that's always a good idea.

My concern is that editors will use AI and not recognize the pitfalls, not actually understand the source and/or target languages at all, and post the output without proper evaluation.

If we create a policy that requires that editors disclose their use of AI, the broader editor community can use that disclosure as a flag to double-check the content. Note: I'm specifically talking about public-facing article / entry pages, not Talk pages or other fora.

I'll also point out that, for those of us trained for and working professionally in localization, we see where AI and machine translation are noticeably not better than humans. ‑‑ Eiríkr Útlendi │^{Tala við mig} 09:22, 28 April 2026 (UTC)

No. Unedited machine translation is worse than no translation at all. I see more and more apps and websites in recent years that publish stuff with unedited machine translation, and it's horrible. I guess that they think that it's good for their business, and while I think that they are very wrong in a lot of ways, I have no energy to fight against them all. But Wikimedia sites must absolutely not do this. Our business is providing information that is fact-checked and well-written by humans, and unedited machine translation is completely opposed to this. Amir E. Aharoni (talk) 01:40, 5 May 2026 (UTC)

•  Oppose Consequences? Reliability for proof? --Julius Senegal (talk) 18:57, 27 April 2026 (UTC)
•  Weak oppose Like many others, I'd prefer if machine translation of content was forbidden, especially since it's so easy for users who aren't fluent enough to review their outputs to add machine-translated content to Wikimedia projects. If that part were changed I'd be willing to change to a "support" since using machine translation in discussions, particularly in multilingual spaces like Wikidata, makes sense to me. While I'd strongly prefer if it were disclosed in all cases, I think machine translation is probably one of our best options for experienced Wikidata users to get through to new users who don't share a language with them. Mcampany (talk) 19:08, 27 April 2026 (UTC)
•  Oppose Following Jeroen N and Msz2001 The matter of not being able to detect LLM content is disturbing, and this aspect should be explicitly mentioned in a global policy. I think humans should write the article text, not just the prompt. The paradox is, that LLM content cannot be malicious per se, because noone will be able to discern between manual and generated content; it is sort of not existing. I don't have any experience in this subject, inside or outside of Wikipedia, but I think that is the scenario to come, and is probably the working model for constructing LLMs. Sechinsic (talk) 19:46, 27 April 2026 (UTC)
•  Oppose for not going far enough. LLM-based translation and copy editing are every bit as problematic as LLMs creating content from scratch. --Carnildo (talk) 23:47, 27 April 2026 (UTC)
•  Oppose as too weak. I said this in relation to the proposed global ban below, but I think we need to be much stricter with AI than current. I would go as far as to say that use of AI should be as heavily discouraged as writing an autobiographical article, with exceptions for basic tasks like spellchecking, although these should still be discouraged. TheDowningStreetCat (talk) 00:43, 28 April 2026 (UTC)
• AI is not ready for translation, nowhere near. Think of concepts that don't exist in the target language. Or the industry-specific tone that may or may not have an equivalent model. Retired electrician (talk) 01:28, 28 April 2026 (UTC)
•  Support as a bare minimum which projects can improve on or strengthen. Secretlondon (talk) 08:58, 28 April 2026 (UTC)
•  Oppose - Every project should set their own rules, even if they are weaker than proposed here. There shouldn't be a global AI policy. Mbch331 (talk) 10:48, 28 April 2026 (UTC)
•  Support Well very well (talk) 12:28, 28 April 2026 (UTC)
•  Oppose, AI-generated translations should also be labeled. Ternera (talk) 13:16, 28 April 2026 (UTC)
•  Oppose Per many comments above. I believe this policy as worded is too weak. That being said, I am in favor of AI assisted editing in certain limited circumstances. The changes I would need to be able to support this proposal would be 1. Excluding the "translation" exception from the Content proposal's disclosure requirement; and 2. It should be clear that individual projects may adopt stricter requirements than listed here, and that those projects should develop guidelines for how to implement this policy and any complimentary local policies. I also would not be comfortable supporting any proposal unless and until sufficient input from Wikisource, Wikidata, and Commons editors has been solicited and incorporated, as well as from projects in minority languages. -- Mathmitch7 (talk) 14:41, 28 April 2026 (UTC)
•  Support - And support creating templates/buttons in editors to make it easy to add an AI notification, which ideally would be in the form of an easily recognizable icon. Whaledad (talk) 15:57, 28 April 2026 (UTC)
•  Oppose Meta typically doesn't set global policies that impact content. I'm not sure what benefit creating a global policy like this would have without a body of editors able to do enforcement across all projects - and similarly, I am not sure that such enforcement would be welcome or productive. – Ajraddatz (talk) 00:51, 29 April 2026 (UTC)
• Weak Oppose: I support the content section, but AI based bots should be permitted to engage in discussions with editors who are ok with receiving unreviewed AI messages. Perhaps a carveout for the user talk namespace? MetalBreaksAndBends (talk) 15:09, 29 April 2026 (UTC)
•  Comment I'm a little lost as what it means to adopt this as a baseline, so I want to ask a few clarifying questions. If this is adopted, do I understand correctly that:
  1. Projects are still free to further restrict the use of AI in article space, File space, etc., up to and including a total ban.
  2. Projects are still free to require disclosure of AI tools used in translations or copyediting.
  3. If a project allows someone use AI to mass-translate or mass-generate articles, they cannot absolve that individual of full responsibility for the results.

Jmabel (talk) 17:55, 29 April 2026 (UTC)

•  Support the content part,  Oppose the discussion part. For the latter, there is a terminological issue: the only use of AI in an actual discussion (i.e., arguing with real humans) should be for translation and formatting, no amount of disclosure can compensate for an AI's immensely faster generation of text. On the other hand, say, using AI to create a list of errors in our text for other editors to review is perfectly valid (subject to disclosure, naturally). --Викидим (talk) 20:43, 29 April 2026 (UTC)
•  Oppose - Every project should set its own rules. In fact, I believe that AI-generated content shouldn't be allowed at all.--Jalu (talk) 21:23, 29 April 2026 (UTC)
•  Oppose Every project should set it's own rules. There's nothing gained by having a global rule. Different projects experimenting with slightly different policies and learning from each other is the way content policy was always developed in Wikimedia and it should stay that way.

I would also note that Wikidata has different needs here. Letting an LLM write a SPARQL query that you use for adding data is a lot easier than writing that query by hand, but I don't consider it important to have that SPARQL query disclosed. Many edits on Wikidata also don't come with edit descriptions so disclosure can be more complex. ChristianKl ❪✉❫ 08:41, 30 April 2026 (UTC)

Re: "There's nothing gained by having a global rule.": Not true. An opt-out-able rule helps small wiki projects avoid a bureaucratic overhead involved in designing and approving a rule/policy, and still allows all rule-fähig/rule-capable wikis to implement an opt-out policy. --Dan Polansky (talk) 11:05, 30 April 2026 (UTC)

•  Oppose The use of AI models to generate or rewrite content must be strictly forbidden globally. At least English and German language Wikipedia, possibly other projects, have already rules in place that largely forbid their use, that's very good, but we need to introduce even stricter rules globally to ensure that Wikimedia projects remain made by humans, for humans. Gestumblindi (talk) 09:26, 30 April 2026 (UTC)
•  Support --Pafsanias (talk) 13:46, 30 April 2026 (UTC)
•  Weak oppose While I don't oppose AI in general (only irresponsible use) and I think a rule for disclosure could be a good compromise, there are also many stigmas for people who use AI, even if they use it only cautiously and wisely to improve their texts. This forced disclosure could encourage witch hunts and hounding by ideological opponents of AI, which should be prevented. Therefore I prefer the solution below that does not force disclosure, only encourages it. --Eddy 97 (talk) 16:08, 30 April 2026 (UTC)
•  Oppose The current proposal is under-defined for a global policy. Does content extend to media as well, while Commons is still discussing proper AI policy that's based on AI generation argument alone (and independent of other factors like the current project scope definition). The intent is definitely right, but should be as sound as possible due to its global scope.

Furthermore, most users still do not disclose their AI tools (be it LLM or others) usage so a policy without proper procedures to follow and actions to partake once abuse is detected is pointless. Scrolling through SRM and other related forums we see the global patrollers and certain local under-developed and understaffed communities already struggle with machine translation, which predates AI and LLM tools by 15 years where we have not yet argued an effective solution to crosswiki machine translation abusers. And now we're expected to detect AI abusers as well, which are usually somewhat harder, more obscure to discover from a crosswiki patroller experience? The whole global community is underfed to effectively apprehend the danger. A ban on AI tools might also interfere with rights of local projects too much.--A09|(pogovor) 20:50, 30 April 2026 (UTC)

•  Oppose The proposal as I understand it creates a conflict between itself and already existing local policies that are tailored to the specific needs of a project. A common ground is hard to find as there are very different needs in each of the projects. See for example Wikisource where it is not uncommon to use AI for transcribing texts and then (in a separate step) to proof it by a human reader. --AFBorchert (talk) 13:15, 1 May 2026 (UTC)

   Comment Transcribing sources with AI-based tools is, in my opinion, neither generating nor rewriting content with AI. As I understand this proposal, the meaning of AI for its purposes is basically generative AI. We have to distinguish between generative AI and other forms of AI. Gestumblindi (talk) 13:14, 2 May 2026 (UTC)

•  Oppose, baseline should simply be disclosure. The idea of "review" other than the wiki standard of edits + reviews + re-edits is ambiguous, slow, and not well suited to mediawiki. Also that sort of friction is particularly counterproductive on smaller wikis, which are also the most likely to not have the time / awareness to come up with their own bespoke policies. –SJ talk  20:55, 1 May 2026 (UTC)
•  Oppose It doesn't help much to say that Wikimedia projects must disclose the use of AI. This is because, for those who don't have a native language English, local language translation (LLM) based tools seem to be the easiest way to make a wiki article better. it is similar to the bot policy or Translations, as the user is responsible for the action taken on the wiki.Kasyap (talk) 11:43, 2 May 2026 (UTC)
•  Oppose I do not agree with the use of AI for translations in Wikimedia projects. Urci dream (talk) 18:30, 2 May 2026 (UTC)
•  Oppose I do not believe this is something that needs a global policy. If AI is used in a disruptive way, most projects have or can write local policies to handle disruptive editing in general. Also, the difference between "basic copyediting" and a "rewrite" may not be so clear. Many people use writing software, such as Grammarly, which can rephrase sentences to make them more readable. I don't think this is something that would need disclosure. TheJoyfulTentmaker (talk) 00:31, 3 May 2026 (UTC)
•  Support мы, как сообщество, не сможем остановить прогресс. Прямым следствием прогресса является широкое использование различных разновидностей ИИ (я так называю LLM) для целей википедии и братских проектов. Одним из проявлений является использование LLM для написания статей или значительных частей статей, перевода, редактирования, вылавливания багов и тп. В этот список стоит включить и обсуждения, форумы и другие подобные места. Вне зависимости от целей и методов использования LLM, ответственным за правки является человек разместивший правки. К подобным правкам следует применять все меры, которые применяются к любым другим правкам участников не использующих LLM (проверка на копипаст, проверка на корректность языка, проверка на соответствие источникам, проверка самих источников). Авторы, использующие LLM, могут открыто заявлять об этом в описании правки. Мы не можем их обязать так делать, но мы можем и должны поощрять участников указавших использование LLM. Если мы начнём приследовать участников за использование LLM, то они начнут скрывать эту информацию и снижать внимание к качеству написания статей. Тут стит поделиться своим опытом: я разместил у себя на странице участника шаблоны, раскрывающие моё использование LLM, в описаниях к правкам я всегда указываю использование ИИ, и, в случае переводов из других языковых разделов, я всегда указываю, что при переводе использовался ИИ. --VladimirPF (talk) 14:49, 3 May 2026 (UTC)
  • (As a community, we cannot halt progress. A direct consequence of progress is the widespread use of various types of AI (which I refer to as LLMs) for Wikipedia and its sister projects. One manifestation is the use of LLMs for writing articles or significant portions of articles, translation, editing, identifying errors, and so on. This list should also include discussions, forums, and other similar venues. Regardless of the purposes and methods of LLM use, the editor who submits the edits is responsible for them. Such edits should be subject to all the same measures applied to any other edits by editors who do not use LLMs (checks for plagiarism, language correctness, adherence to sources, and verification of the sources themselves). Editors who use LLMs may openly disclose this in the edit summary. We cannot require them to do so, but we can and should encourage editors who disclose their use of LLMs. If we begin to penalize editors for using LLMs, they will start to conceal this information and pay less attention to the quality of article writing. Here, it is worth sharing my experience: I have placed templates on my user page disclosing my use of LLMs; in edit summaries, I always indicate the use of LLMs; and in the case of translations from other language versions, I always note that LLMs were used in the translation.)--VladimirPF (talk) 14:49, 3 May 2026 (UTC)
•  Oppose, per above. IdanST (talk) 10:55, 5 May 2026 (UTC)
•  Strong oppose, let communities decide on their own without putting English Wikipedia policies on the spot as "the right ones" please. Sannita - not just another it.wiki sysop 11:22, 5 May 2026 (UTC)
•  Strong oppose: 1) AI is under strong and fast evolution, any rule could be obsoleted in short time, 2) Wikipedia is growing with freedom, without too many generalist top down policies witch are not considering local condition.--Bramfab (talk) 12:27, 5 May 2026 (UTC)
•  Oppose Let communities decide based on their local rules and sensitivity.--Superspritz^{tell me} 20:48, 5 May 2026 (UTC)
•  Oppose Global policies should not be imposed except in cases where it is actually necessary to overide the autonomy of individual projects. In this case, it does not appear necessary. I assume that individual projects will take steps to prevent inappropriate use of AI without being forced to do so. James500 (talk) 20:10, 6 May 2026 (UTC)

"Individual projects will take steps" I would have thought so too . . . But Esperanto Wikipedia is a counterexample. --Arbarulo (talk) 14:15, 8 May 2026 (UTC)

•  Support. But even for translation it should be disclosed if used significantly, and it must be clear that unpolished, AI-generated translations are not acceptable, either. --Paloi Sciurala (talk｜contribs) 09:26, 8 May 2026 (UTC)
•  Support, provided that this is clearly understood as a baseline policy: local communities must remain free to adopt stricter rules, including requiring disclosure for AI-assisted translation and basic copyediting, and even prohibiting some uses of AI-generated content if they consider this necessary for their project. --Gitz6666 (talk) 12:49, 14 May 2026 (UTC)

Opt-out policy

Should the following be adopted as a global opt-out policy?

Content

The use of AI models to generate or rewrite content is prohibited, except for translation, examples of AI-generated output, and basic copyediting. Pages created entirely by AI models and published without human review may be eligible for (speedy) deletion.

Discussions

The use of AI models to generate comments in discussions is prohibited, except for translation, grammar correction, language refinement, and examples of AI-generated output. Comments that are clearly AI-generated may be struck, collapsed or removed.

Discussion

•  Support as proposer. Chaotic Enby (talk) 18:13, 23 April 2026 (UTC)
•  Support Pretor90 (talk)
•  Support and I'll go further by blocking the use for "grammar correction and language refinement" as well since LLMs make such comments rather frustrating to read (+ they also tend to change the meaning entirely). Someone who cannot contribute proficiently in a language should not use it at all. This also applies to Meta, where users should be encouraged to use their own language instead of LLM-enhanced "English". Leaderboard (talk) 11:29, 24 April 2026 (UTC)

  @Leaderboard, see my plan in the discussion area on this RfC, but grammar correction and language refinement tools like Grammarly, are used by many people nowadays. Also, if there is a disabled editor like someone who is dyslexic, they may use these tools to help edit.

  Thanks, Globetrotter30 (talk) 09:23, 16 May 2026 (UTC)

  @Globetrotter30 I don't understand what you're trying to say - are you proposing that Grammarly be allowed or not? Leaderboard (talk) 12:00, 16 May 2026 (UTC)

•  Support as it would save smaller wikis time and resources. Not too keen on allowing copyediting as LLMs are terrible for most languages, but that can be adjusted locally if it's an issue Kowal2701 (talk) 15:25, 25 April 2026 (UTC)
•  Oppose, if this policy is adopted, we might as well ask the Board to make a decision to shut down Wikisource. Wikisource relies heavily on AI models for OCR, which helps with book transcription (the main thing Wikisource does), and enforcing a policy prohibiting the use of AI will make contributing to Wikisource extremely tedious and effectively kill the project. -- Sohom (talk) 19:10, 25 April 2026 (UTC)

  I don't think this would affect OCR? This policy is about content generation, something like transcription shouldn't be affected at all. Chaotic Enby (talk) 20:15, 25 April 2026 (UTC)

  I've added a definition of what falls under content generation, hopefully that should clarify things. Sorry for the misunderstanding! Chaotic Enby (talk) 21:40, 25 April 2026 (UTC)

•  Support – as a draft contributor. – Aca (talk) 01:05, 26 April 2026 (UTC)
•  Support AI-content is merely disruptive at the bare minimum, even more when it is being used in discussions (I'm here to have interactions with humans, not here to waste my time talking to machines). An opt-out policy at least allows individual wikis to take their loss if they wish to use AI for content matters without intruding on individual wiki autonomy. //shb (t • c) 01:15, 26 April 2026 (UTC)
•  Oppose as worded – in certain language pairs current AI models cannot be trusted to perform translation correctly. Alıƨsi (talk) 01:15, 26 April 2026 (UTC)
•  Support I believe this would be mostly beneficial to all projects. Machine translation is a bit icky in some situations, even in languages it handles somewhat well (e.g. bias on way of writing certain words and conjugation). EdoAug (talk) 01:20, 26 April 2026 (UTC)
•  Support -- Jeff G. ツ (please ping or talk to me) 01:22, 26 April 2026 (UTC)
•  Weak oppose as I don't like the "translation pass" for content. For discussion/mass messaging there's an argument for disclosed AI translation but ultimately it's a waste of time. The reader can use the same tools to translate the message if they want. Ignacio Rodríguez (talk) 01:55, 26 April 2026 (UTC)
• Chaotic Enby, how do you imagine this working for images? I don't know if you've read c:COM:AI and are trying to overturn Commons' guideline here, or if you didn't think about the non-Wikipedias. WhatamIdoing (talk) 02:10, 26 April 2026 (UTC)

  I did consider Wikimedia Commons, and clarified that this policy will not apply by default to any project that, in the past, formally discussed a policy or guideline regarding generative AI, whether successfully or unsuccessfully. In the case of Commons, as they already have their own set of rules regarding use of AI for image generation, they wouldn't fall under this global policy. The intent of it, as an opt-out policy, is specifically to not overturn any policies or guidelines, but to provide a "default" that can be adapted (or discarded) if nothing has been set up (or explicitly voted against) yet. Chaotic Enby (talk) 02:15, 26 April 2026 (UTC)

  What if they adopt such a rule "in the future", instead of "in the past"?

  Overall, I don't like this approach of asking people to vote on the 150 words on this page, while having another 300 words at Artificial intelligence/Draft policy that change the meaning. WhatamIdoing (talk) 03:39, 26 April 2026 (UTC)

• How will this work for projects that don't have a "speedy deletion" concept? d:Q4664109 lists 100 Wikipedias with such a policy, but only 10 (out of 174) Wiktionaries, 2 (out of 83) Wikisouces, etc. Some of them won't have a separate page, but many of them don't really have speedy deletion as a concept in their policies. WhatamIdoing (talk) 02:10, 26 April 2026 (UTC)

  Good point indeed. I'm amending it to "may be eligible for (speedy) deletion" to emphasize that they will always be eligible for the wiki's deletion process, with fast-tracked processes being privileged. Chaotic Enby (talk) 02:17, 26 April 2026 (UTC)

  Everything outside of the Special: namespace "will always be eligible for the wiki's deletion process". So this is either meaningless, or it's an attempt to impose a faster process than the local community might choose. For example, a low-traffic community might prefer that the local community first reach a consensus about whether a page is actually AI-generated instead of AI-translated. And that could require an ordinary deletion discussion, rather than saying that any passing admin can delete it on sight. WhatamIdoing (talk) 03:43, 26 April 2026 (UTC)

  I’d define “speedy deletion” as “deletion that is justified by obviously fulfilling the requirements instead of having discussion and reaching consensus every time.” It can as well be defined as “criterion that are designated by the community that deletion is automatically approved every time such a page is created.” Such an approach probably already exist on most of our wikis in dealing with spams, as you do not need to have a discussion to determine whether a given piece of text is spam or not. 1F616EMO (on zhwiki) 02:17, 26 April 2026 (UTC)

  Most of our wikis don't have a big problem with spam pages. That is primarily a problem at large, high-traffic Wikipedias. (Spam URLs are a problem everywhere, but those get blacklisted, not deleted.) WhatamIdoing (talk) 03:49, 26 April 2026 (UTC)

  Note: Edit conflict have eaten away my comment and the reply: [1]. I am on mobile, so please help me to restore the comments. 1F616EMO (on zhwiki) 02:23, 26 April 2026 (UTC)

  Fixed, @1F616EMO. Thanks for posting a note about the problem. WhatamIdoing (talk) 03:37, 26 April 2026 (UTC)

•  Oppose as we should not add constraints on the tools. If AI outputs are problematic, we can handle them through existing protocols as with all other problematic contributions; if AI outputs are refined enough that they meet the standard of what normal human contributors would have done, that's fine. That said, we can assume AI is problematic most of the time if the user cannot master the skills on prompting and refining, and thus prohibit any contents that clearly did not go through enough human review and refinement; for the same reason, we should require mandatory disclosure on the use of AI tools, as suggested in the first proposal. Even if this proposal is going to pass, we should allow individual communities to override it and/or expand it on their own. 1F616EMO (on zhwiki) 02:06, 26 April 2026 (UTC)

  Regarding your last point: yes, as this is an opt-out proposal, individual communities would be able to freely opt out of it, or refine/expand it as they will. Chaotic Enby (talk) 02:08, 26 April 2026 (UTC)

  Thought I oppose to a full ban on generating AI contents, I still support the prohibition of AI in discussions. While contents can be AI-assisted, discussions are a human thing; AI might be used to summarize discussions, but never to give opinions on behalf of a human. Obvious use of AI in discussions implies a lack of thinking. We are here to build the sum of human knowledge as humans. 1F616EMO (on zhwiki) 05:44, 26 April 2026 (UTC)

•  Support. Genoskill (talk) 02:16, 26 April 2026 (UTC)
• Opposed. I support the intention behind it, but as it is written, it is confusing, self-contradictory, and potentially harmful:
  1. General comment: Instead of AI, it's better to write language model (also known as "AI"). The term "AI" is common, but it means a lot of things, and gets redefined every few years, whereas the term "language model" is fairly stable.
  2. There must not be an exemption for translation. Machine translation without human review must be prohibited everywhere, whether the technology used to generate is "AI" or not "AI" (see above: the word "AI" is not well-defined anyway). And contrariwise: Machine translation that went through human review and conforms to other policies about citations, verifiability, style, etc., is OK. So there's no need to mention translation separately at all.
  3. The first sentence says that "The use of AI models to generate or rewrite content is prohibited", but the second sentence talks about "Pages created entirely by AI models and published without human review". So... if they did go through human review, are they not prohibited? As written now, it is potentially contradictory. --Amir E. Aharoni (talk) 02:18, 26 April 2026 (UTC)

  So... if they did go through human review, are they not prohibited? The proposal says nothing of the sort; it only implies that LLM-generated pages which have gone through human review are not necessarily eligible for speedy deletion. The proposal seems to be modeled after English Wikipedia policy; compare w:en:Wikipedia:Writing articles with large language models, which says the use of LLMs to generate or rewrite article content is prohibited (emphasis in original) and the narrower G15 speedy deletion criterion, which says This applies to any page that exhibits one or more of the following signs [...] generated by large language models (LLM) and would have been removed by any reasonable human review (emphasis in original). OutsideNormality (talk) 03:43, 27 April 2026 (UTC)

  Maybe it doesn't intend to say it, but it can be interpreted as if it does. Amir E. Aharoni (talk) 12:51, 27 April 2026 (UTC)

•  Support knowledge is human. Best, HouseBlaster (talk • he/they) 02:50, 26 April 2026 (UTC)
•  Support with addition proposal: Comments/arguments/replies that are suspected or identified being generated fully using AI will likely to be ignored/not counted as good faith comments/arguments/replies. Reasons: In Wikidata request for undeletion, and administrators noticeboard, I often finds new users fighting for unnotable items using AI to generate arguments that even misinterpret or fail to recall Wikidata Notability. Each time human users replies, each time they used long winding arguments that they even not check beforehand. Some using incorrect template/templates that arent used in Wikidata. Yamato Shiya (talk) 03:16, 26 April 2026 (UTC)
•  Support Far better than the above proposal. Our projects should be free of AI slop. Chaddy (talk) 03:32, 26 April 2026 (UTC)
•  Oppose I don't think using LLM should be an automatic deletion. If the article can be cleaned up, by all means do so and keep it, I prefer the policy above better, as it advises the user that they're responsible for the posting that LLM created article. W.K.W.W.K ...Straight outta Underverse 03:58, 26 April 2026 (UTC)
•  Comment If I read this correctly, this would completely ban the use of AI-based translation tools in discussions, even as an aid if you do not know how to translate a particular phrase. Is that correct? If that is the intent, it almost certainly should not apply to multilingual projects. I cannot imagine how (for example) we could operate the Commons help desk without some use of such tools. Similarly: would this mean that if (for example) a native Bengali-speaker needs clarification of something in the English-language Wikipedia, they are not allowed to use an AI-driven translation tool to help them frame a question in English? - Jmabel (talk) 04:14, 26 April 2026 (UTC)

  That's an oversight on my part, sorry. I meant to include the exemption on translation for discussions, just like for content editing. Chaotic Enby (talk) 04:15, 26 April 2026 (UTC)

•  Support, maybe even a strong one. Most large WP editions (for certain: en:WP:NEWLLM, de:WP:KI, fr:WP:IA and I'm aware of a discussion on RU-WP where the participants more or less uniformly stood against AI in general. Further references may be found among the interwikis in d:Q119702818 and d:Q137217991) already forbid LLM output, at least for articles, often also in discussions. It's a good time to make that universal for the Wikimedia world. Grand-Duc (talk) 04:17, 26 April 2026 (UTC)
•  Oppose per Amir Aharoni. We must not exempt translation. Abzeronow (talk) 04:42, 26 April 2026 (UTC)
•  Support.--ELexikon (talk) 05:02, 26 April 2026 (UTC)
•  Support--Andromeda2064 (talk) 06:22, 26 April 2026 (UTC)
•  Oppose Grueslayer (talk) 06:27, 26 April 2026 (UTC) Undeclared use of AI should lead to speedy deletion.
•  Support --Metsavend (talk) 06:32, 26 April 2026 (UTC)
•  Support --Chelin (talk) 06:51, 26 April 2026 (UTC)
•  Oppose per Amir Aharoni, point 2 and 3. — The preceding unsigned comment was added by NisJørgensen (talk) 06:57, 26 April 2026 (UTC)
•  Oppose This proposal might be interesting for large languages with huge communities, which already have policies. Imposing that to smaller and underserved communities is something we shouldn't be doing, because local dynamics, needs and uses are usually different. There's no need for a global policy dicussed in English. -Theklan (talk) 07:22, 26 April 2026 (UTC)
•  Oppose First, on my home wiki (kowiki), most articles are translated from enwiki. Before AI models became common, most users relied on machine translation tools, and I don't see a clear difference between using those and using AI now. And TBF AI may even produce better results than traditional translators. Kowiki also has a gadget that uses Gemini for translation. My point is that in a wiki where the majority of content is translated, banning AI tools would not change much, since users would simply continue using translation tools anyway, and it is not really possible to distinguish between content produced by traditional machine translation and that generated by LLMs. Therefore, I do not agree with the view that translation should also be included in such restrictions. btw I oppose applying CSD to a wiki where discussions about LLM generated content have not yet begun, regardless of the local community's consensus. In principle, CSD should follow local consensus.--𝓰𝓲𝓷𝓪𝓪𝓷기나ㅏㄴ(T/C) 07:30, 26 April 2026 (UTC)
•  Support. ImenaOphelia (talk) 07:34, 26 April 2026 (UTC)
•  Oppose as per Amir E. Aharoni above.--Naḥum (talk) 07:57, 26 April 2026 (UTC)
• Neutral. I'm sympathetic to the basic idea, enabling speedy deletion for AI-generated or -rewriten articles; but this isn't always clear. Jaap-073 (talk) 08:17, 26 April 2026 (UTC)
•  Oppose How do you prove that a page was created entirely by AI models and published without human review? Is merely reading the text by its author a sufficient review? Or fixing one comma? Msz2001 (talk) 08:24, 26 April 2026 (UTC)
•  Oppose as per Amir E. Aharoni Empat Tilda (talk) 08:42, 26 April 2026 (UTC)
•  Oppose as worded – for many smaller languages even translation and basic copy editing need to be disclosed. AI is a serious threat to small languages such as Cornish, of whose wikipedia I am an administrator. We face the publication of many books which have trawled the various spelling systems used and jumbled them up, invented totally new words and misused the basics of grammar. If we allow AI to write wikipedia articles, very soon AI will be using AI invented mistakes, presenting it as legitimate content, and learners will not be able to tell the difference because of the flood of new content. Brwynog (talk) 08:51, 26 April 2026 (UTC)

  Same concern as an active contributor of the Low Saxon wiki. The generated texts are mixed dialects, switch to German mid sentence etc. Similar to many other smaller wiki the Low Saxon wiki is by far largest Low Saxon text corpus. Allowing any AI use for content outside templates etc. will amplify the existing mistakes the AI makes since there is no other sizable text corpus to work as a corrective force. We already have bad AI translations regularly, that we can't double check and improve since we lack sufficient contributor to clean up the work of others Flaverius (talk) 09:27, 26 April 2026 (UTC)

•  Oppose AI is a moving target so this would need to be reconsidered regularly. For example, AI originally did not pass a 12th grade science test (https://www.vox.com/future-perfect/2019/9/13/20863269/ai-aristo-science-test-allen-institute ), until 2019. Also, it is up to the deletion nominee to prove AI usage, as he/she is making the claim. Just writing 'AI usage' as a reason is not good enough. Will reconsider if both points are addressed, but not if only one is addressed.--Snævar (talk) 09:20, 26 April 2026 (UTC)
•  Oppose – Per my comment on the baseline policy. Moreover, the name 'opt-out policy' seems to me misleading. It is not a policy setting out the conditions for opting out (of the baseline policy?), but appears to be the proposed default policy, which you can opt out of. This is not made clear, however. Jeroen N (talk) 09:51, 26 April 2026 (UTC)

  Opt-out policies are an established pattern across Wikimedia - compare, for example, the global bot policy. Omphalographer (talk) 17:41, 2 May 2026 (UTC)

•  Support --Yann (talk) 09:52, 26 April 2026 (UTC)
• Simplify
  1. Use same wording for content and discussion
  2. Acknowledge AI translation
  3. Combine with baseline policy
  4. For now, create a guideline instead of policy to stimulate discussion but allow project communities to more time discuss their local approach
     -- Jtneill - Talk 10:48, 26 April 2026 (UTC)
•  Support --Saroj (talk) 12:21, 26 April 2026 (UTC)
•  Oppose in favor of the baseline policy. Таёжный лес (talk) 12:53, 26 April 2026 (UTC)
•  Oppose per Amire80. Naut-rena (talk) 13:09, 26 April 2026 (UTC)
•  Oppose AI translation is de bane of all small Wikipedias. It should be completely banned. Ieneach fan 'e Esk (talk) 13:26, 26 April 2026 (UTC)
•  Oppose. The same reasoning as in the previous point. These restrictions have no practical meaning. Лоття (talk) 13:36, 26 April 2026 (UTC)
•  Oppose. A complete ban (with loopholes that threaten small projects) is senseless and should not be adopted. -- Futbollo (talk) 13:46, 26 April 2026 (UTC)
•  Strong oppose The term 'AI' is way too broad, and the technology involved is rapidly evolving, and having a global policy (even if it is an opt-out policy) restricting its usage would hamstring community members who dare to innovate in small wikis. Also, how does one even prove AI usage? Stylometry could be a somewhat good method for that but it does not work for new contributors. Speedy deletion without sufficient proof seems like a horrible idea to me. Redmin (talk) 14:09, 26 April 2026 (UTC)
•  Oppose Banning AI/LLM at this point is like trying to put the toothpaste back in the tube. The tools are prolific and increasingly difficult to detect. It is better to hold contributors accountable for their content regardless of how they generated it.
  Not listed on the "Policies by project" page is English Wikinews' proposed policy: n:WN:AI, which essentially mirrors this proposal in that humans are ultimately responsible—both contributors and reviewers—and no unsupervised generation, i.e., no fully-AI-generated articles were allowed. Michael.C.Wright (talk) 14:19, 26 April 2026 (UTC)
•  Oppose -- XXBlackburnXx (talk) 14:20, 26 April 2026 (UTC)
•  Oppose As already mentioned, an editor always takes responsibility for what they publish, and AI can be used on different degrees, thus we should not impose any rule on how humans compose their texts. On the other hand, we should forbid unsupervised AI (yes to speedy deletion). --Grufo (talk) 14:50, 26 April 2026 (UTC)
• Support. Sapphaline (talk) 15:54, 26 April 2026 (UTC)
•  Weak oppose As this may feel like a step too far- you had might as well ban use of Artificial Intelligence entirely.--Globetrotter30 (talk) 15:55, 26 April 2026 (UTC)
•  Oppose – Repeating what I've commented above: While I broadly agree with parts of the proposal, I don't support the underlying sentiment against AI assistance. Automation, including AI, is one of the few practical ways smaller wikis can keep pace with larger ones despite limited contributors. In workshops I’ve participated in, younger contributors consistently identify AI-assisted workflows as a top priority, and in my own admin work on my home wiki I rely on these tools for everything from interface improvements to bug fixes and content cleanup. Misuse, such as blind copy-pasting, should be addressed case by case, as with any low-quality edits. More importantly, this approach seems to run counter to Wikipedia’s original open model, which deliberately lowered barriers to participation compared to other more formal academical projects or even online ones like w:en:Nupedia. That openness enabled its growth, and discouraging AI risks moving in the opposite direction; the focus should instead be on better integrating these tools on its current workflows or on developing our own AI tools. - Klein Muçi (talk) 17:39, 26 April 2026 (UTC)
•  Oppose Nizzan Cohen (talk) 18:42, 26 April 2026 (UTC)
•  Strongly oppose This proposal is partly the same as the other one. I read: "Pages created entirely by AI models and published without human review may be eligible for (speedy) deletion." This still gives the possibility to write content with AI, which we shouldn't do. You can't count on human editors. It's impossible! There are children editing, and other amateurs (mainly amateurs). It makes no sense at all to introduce AI for writing any content, or translate text. And again also per Jeroen N. Happytravels (talk) 19:27, 26 April 2026 (UTC)
• *  Support I like the quality > scale sentiment, especially on small wikis Krol111 (talk) 20:00, 26 April 2026 (UTC)
•  Support (strong) Accipiter Gentilis Q. (talk) 20:55, 26 April 2026 (UTC)
• All use of AI within Wikipedia must be strictly and thoroughly banned for now, if possible. EDIT: We must have the human touch, as if always is inconmplete, it has the human capability. You may translate from my user page from Finnish: "Kaikillahan meillä tekoäly on (paitsi uskovaisilla)" --Höyhens (talk) 22:08, 26 April 2026 (UTC)
• All use of AI within Wikipedia must be banned for now. --Sinucep (talk) 23:09, 26 April 2026 (UTC)
•  Support, in total agreement with Sinucep --Qualitätssicherung (talk) 23:36, 26 April 2026 (UTC)
•  Support any method of clearing out ai. ltbdl (talk) 00:29, 27 April 2026 (UTC)
•  Oppose as per Brwynog, Flaverius, Theklan, and others. If I want machine-translated articles from the English Wikipedia I can make them myself, thank you very much. --Pseudo-philodoxia (talk) 00:42, 27 April 2026 (UTC)
•  Support the terminology used in the policy should use LLM rather than AI, as AI technically means a lot of other things. Other than that, having this as the default and allowing projects to opt out is a good defense.Andrei Stroe (talk) 08:16, 27 April 2026 (UTC)
•  Oppose No exceptions for AI translation.--Leptictidium (talk) 09:34, 27 April 2026 (UTC)
•  Support Esperanto Wikipedia is at risk of being overwhelmed by uncited AI-generated content, often containing dubious generalizations (e.g. about what the article topic "is best known for," "represents" etc.). Most of this is created by one well-meaning but misguided user with the complacency of admins who don't seem to see it as a serious problem. Eowiki has always been lax in practice about verifiability, but presumably most uncited content has actually been based on a source - AI drastically increases the probability that it isn't.
  A stricter policy might be desirable but I don't see how that is grounds for opposing - rejecting light protections doesn't get us heavier protections, it leaves us with no protections.
  Por: La Esperanta Vikipedio estas en danĝero de drono en senreferenca AI-kreita enhavo, ofte enhavanta dubindajn ĝeneralaĵojn (ekz. asertanta pro kio la artikoltemo estas "plej konata", kion ĝi "reprezentas" ktp.) La plejparto estas kreata de unu bonintenca sed missupoza uzanto kun la permesemo de administrantoj, kiuj ŝajne ne rigardas ĝin kiel seriozan problemon. Nia vikipedio ĉiam malstriktis en praktiko pri kontrolebleco, sed supozeble la plejparto de la senreferenca enhavo reale baziĝis sur fonto - AI draste pliigas la ŝancon ke ne.
  Pli strikta politiko eble estas dezirinda sed mi ne vidas kiel tio estas kialo por kontraŭo - rifuzo de malstriktaj protektoj ne havigas pli striktajn, ĝi lasas nin senprotektaj. --Arbarulo (talk) 13:51, 27 April 2026 (UTC)
•  Oppose Have you think of the minority languages with lack of online materials? The generated AI content will further contaminating the minority languages even for the purpose of correcting grammars and such. Hakimi97 (talk) 14:11, 27 April 2026 (UTC)
•  Oppose Linh Huynh (talk) 16:26, 27 April 2026 (UTC)
•  Strong oppose Per Alıƨsi and Amir E. Aharoni. I work in localization, I know first-hand how rubbish -- and sneakily so -- AI translation can be. ANY use of AI to generate article / entry content MUST BE DISCLOSED, as far as I'm concerned. I am also very leery about undisclosed use of AI to copyedit, for similar reasons. ‑‑ Eiríkr Útlendi │^{Tala við mig} 18:08, 27 April 2026 (UTC)
• Generally  Support. This is a good starting point, and seems to generally represent the consensus on large wikis. Projects which would prefer to establish stricter or looser policies on AI use are free to do so; small projects should not be exposed to the risk of unrestricted AI use. I would prefer that the disclosure of AI use be made mandatory, but not to the extent that I would oppose this. Omphalographer (talk) 22:51, 27 April 2026 (UTC)
• Conditional  Oppose. In isolation, I find this a suitable default policy, but the carve-out for translation and copyediting in the baseline policy means I can't support it. --Carnildo (talk) 23:55, 27 April 2026 (UTC)
•  Oppose, I don't see why a second policy is needed in addition to the first one. Well very well (talk) 12:29, 28 April 2026 (UTC)
•  Oppose including the "Translation" exception in the Content opt-out policy. If we only had one AI policy globally, and it was an opt-out policy like this, it should not allow for machine translation of content. -- Mathmitch7 (talk) 14:43, 28 April 2026 (UTC)
•  Support - And support creating templates/buttons in editors to make it easy to add an AI notification, which ideally would be in the form of an easily recognizable icon. Whaledad (talk) 15:58, 28 April 2026 (UTC)
•  Support, but I think the wording should mention LLMs as an example. --Minoa (talk) 20:01, 28 April 2026 (UTC)
•  Oppose. On the content side, the technique used by a human editor to create the text should be irrelevant to us as long as the result fits our requirements. For example, a typical newcomer editor tends to create bad texts at the beginning. However, we do not ban the newcomers altogether or set up exams to check their readiness for the mainspace. Instead, we have machinery (like en:WP:NPP, en:WP:Draft in English Wikipedia) to deal with sub-standard texts. On the discussion side, what is wrong with a list, compiled using AI, of errors found in the (human-created) text of an article and posted on the talk page? --Викидим (talk) 21:06, 29 April 2026 (UTC)
•  Oppose Every project should set it's own rules. There's nothing gained by having a global rule. Different projects experimenting with slightly different policies and learning from each other is the way content policy was always developed in Wikimedia and it should stay that way. ChristianKl ❪✉❫ 08:42, 30 April 2026 (UTC)
•  Support --Pafsanias (talk) 13:47, 30 April 2026 (UTC)
•  Oppose The usage of AI cannot be reliably detected. Such a rule cannot be enforced correctly and would scare away a lot of genuine authors whose work was falsely flagged as AI. Plus, not all forms of AI are bad. If authors use it wisely and cautiously to generate flawless texts, where is the problem? --Eddy 97 (talk) 15:54, 30 April 2026 (UTC)
•  Oppose Opt-out policy is even too lenient for certain projects which have fully banned AI and AI-related tools. Perhaps a good compromise is to add a clause stating projects can adopt a local policy by which they are not bounded to the global policy anymore.--A09|(pogovor) 20:52, 30 April 2026 (UTC)

  That's implicit in the opt-out policy - any project which wishes to set its own guidelines on AI/LLM use is free to do so, and its policy will override this one. Omphalographer (talk) 17:43, 2 May 2026 (UTC)

•  Oppose per Klein Muçi. Every project should set its own rules, there's nothing inherently wrong with AI assistance and in some cases it is essential. –SJ talk  20:55, 1 May 2026 (UTC)
•  Oppose, per above. IdanST (talk) 10:55, 5 May 2026 (UTC)
•  Strong oppose, per Klein Muçi. Sannita - not just another it.wiki sysop 11:23, 5 May 2026 (UTC)
•  Support Wikipedia is not Grokipedia, I don't see any argument for any Wikipedia edition to allow an usage of AI besides the usual exceptions (translations, copyriediting, etc).--Friniate (talk) 14:23, 5 May 2026 (UTC)
•  Support especially for discussions. Interactions on Wikimedia project must occur among human beings, not between humans and software. --Superspritz^{tell me} 18:01, 5 May 2026 (UTC)
•  Oppose (1) I suspect this proposed policy might result in editors casting aspersions on each other. w:en:Wikipedia:Casting aspersions#Selected examples 2008–2015 describes the problem of aspersions that resulted from rules about UPE, which have the same problem of requiring proof of off-wiki behaviour that is not recorded by diffs and is therefore difficult to prove. (2) This proposed policy also appears redundant. Projects generally already have policies that allow us to remove "bad" content and comments, and sanction editors who add "bad" content and comments. If our existing policies already allow us to deal with such edits and editors, what do we need this policy for?James500 (talk) 20:26, 6 May 2026 (UTC)
•  Support. But even for translation it should be disclosed if used significantly, and it must be clear that unpolished, AI-generated translations are not acceptable, either. --Paloi Sciurala (talk｜contribs) 09:26, 8 May 2026 (UTC)
•  Support AI generates very low quality content. AlbertRA (talk) 15:39, 17 May 2026 (UTC)

AI-assisted editing and human accountability

Should the following alternative policy be adopted instead (both in content and in discussions)?

AI-assisted contributions are permitted, provided they are thoroughly reviewed by a human editor who assumes full responsibility for accuracy, sourcing, and compliance with Wikipedia policies. Content created or published without adequate human review may be challenged and subject to deletion under existing processes. Disclosure of AI use is optional but encouraged, particularly where it may affect transparency or verifiability. Editors are fully accountable for any shortcomings in AI-assisted edits.

--Grufo (talk) 00:34, 27 April 2026 (UTC)

Discussion

•  Support Pretor90 (talk)
•  Support; I'm relatively new but this latter proposal appears a more enforceable policy that doesn't include questionable exemptions. On top of which, it appears to me (see the discussion page) that, should the alternative policies be adopted, there might be great difficulty in safeguarding local Wikipedias from unsupervised, low-quality machine translations. --Pseudo-philodoxia (talk) 01:16, 27 April 2026 (UTC)
•  Support (as the one who proposed it). I should add that on Latin Wikipedia we currently face the opposite concern: we struggle to limit the use of unsupervised AI in translations (other uses have never been a major issue). In my view, what matters most is ensuring proper human review of AI-generated content; disclosure of AI use is of limited relevance and may even be perceived as shifting or diluting user accountability. --Grufo (talk) 02:06, 27 April 2026 (UTC)
•  Strong support From what I've seen in the ongoing discussions, many editors are uncomfortable with machine-generated or anything automated. Whether this reflects a vocal minority or a majority, the situation is similar to academic integrity where accountability should lie with the editor, not the tool they use. AI is becoming mainstream and its use will increase regardless of how restrictive the policies are. This means the policy should focus on human responsability rather than temporary restrictions that will need to be rewritten once they fail to work as intended. --Profesorul Blazat (talk) 07:10, 27 April 2026 (UTC)
•  Oppose A proposal without any teeth in the face of the onslaught of AI slop. We don't need a new policy that specifies that the other policies will be enforced, nor one that says you may or may not disclose the use of AI (it's exactly the same without it).Andrei Stroe (talk) 08:14, 27 April 2026 (UTC)
•  Oppose, agree with Andrei. Due to words "adequate", "may", "optional" the statement leaves too much room for interpretation, basically means nothing more than obvious authors responsibility for his edits and therefore gives no protection against AI slop, especially on small wikis with no local policy. Krol111 (talk) 09:08, 27 April 2026 (UTC)
•  Oppose "Disclosure of AI use is optional" just scuppers this for me.--Leptictidium (talk) 09:35, 27 April 2026 (UTC)
•  Strong support While the argument that this policy may be unnecessary is generally valid, this version is the only viable option. We can't restrict a tool that is constantly improving with labels that malicious users would never apply and that would only hinder legitimate editors. Imposing such an idea on all projects without exception seems reckless. A complete ban on this tool would change nothing except the personal satisfaction of those who promote such ideas. Therefore, this is the only option left. Recognize that a tool is a tool, recognize that no matter how a person writes articles, we already have many established rules, and if an article complies with them, then great. This policy will serve as a good explanation for newcomers and a foundation for local policies in all projects, which, of course, could be much stricter if the community itself finds itself in need of something similar, as in the German Wikipedia, for example. Лоття (talk) 11:15, 27 April 2026 (UTC)
•  Oppose. This proposal is kind of a bit closer than the previous proposals to something sensible, practical, and mission-aligned, but it has several important shortcomings, both in style and in substance:
  • Don't write "AI". As I noted in other sections, the term "AI" is undefinable. It's much better to write language model (also known as "AI").
  • Don't write permitted. Say forbidden, unless. It means the same thing, but feels different. Many people—even many Wikimedians!—often don't finish reading whole sentences, and if those people see "AI-assisted contributions are permitted", they'll think that AI-assisted contributions are permitted.
  • The condition reviewed by a human editor is too short and open to interpretation. Suggested change: reviewed and corrected by a human editor who knows the language of the wiki well and coordinates their contributions with the wiki's community. Rationale: Theoretically, there may be some cases in which correction is not needed, but practically, it's needed pretty much always, and the style of the text must make this attitude explicit. And the requirement to know the language must be explicit, too, because there are a lot of people who add machine-translated or LLM-generated text to wikis in languages they don't know: sometimes with the intent to "help" those languages "develop", sometimes as spam, and sometimes as vandalism. The result is always bad, even if the intention is good.
  • "Content created or published without adequate human review may be challenged and subject to deletion under existing processes." - too weak. If it may (!) be "subject to deletion under existing process", then what is this policy even for? It should have more teeth and say "speedy deletion". --Amir E. Aharoni (talk) 12:48, 27 April 2026 (UTC)
    • Write language model (also known as "AI").: This can make sense.
    • It means the same thing, but feels different: Why should something permitted feel forbidden?
    • Theoretically, there may be some cases in which correction is not needed, but practically, it's needed pretty much always: Isn't that what “review” already means? If something does not need corrections you don't correct it, if it needs corrections you correct it. If errors remain we don't say “this has not been corrected”, we say “this has not been reviewed”… As you implied, if we say “reviewed and corrected” it might be interpreted as an obligation to correct every single word no matter what. --Grufo (talk) 19:54, 27 April 2026 (UTC)

      Why should something permitted feel forbidden? Because if there were no significant problems with using LLMs, and if it would be mostly OK to use them, we wouldn't be having this discussion and trying to write a policy at all. Rules, policies, laws, etc., are usually made to say what is not allowed, and not to say what is allowed. In political philosophy, there are negative and positive rights that (arguably) need to be protected by the government, such as freedom to life, freedom of religion, the right to not die of hunger, the right to self-expression in speech, art, clothing style, etc. I don't think that anyone is arguing for the right to use LLM-generated texts as free speech or self-expression. The problem is that way too often, LLM-generated texts are misused, and that's why they have to be regulated.

      Isn't that what “review” already means? - No, not necessarily. In software development, for example, "code review" usually means reading code and sending comments if something needs fixing, and most of the time, the reviewer is not expected to make the fixes. At least that's the custom in the Wikimedia code review process. That said, I understand that reviewing a text may also include fixing it, but that's just me: I already try to be responsible enough and to understand that publishing LLM-generated texts with made up facts is pointless and harmful, so I don't need a written policy to explain it to me. This policy is written for people who don't understand it and who may think that reviewing can be passive, and the fixing will be done by other people. So the requirement to correct has to be explicit. Amir E. Aharoni (talk) 13:56, 28 April 2026 (UTC)

      Because if there were no significant problems with using LLMs, and if it would be mostly OK to use them, we wouldn't be having this discussion: I think this might be the real point. I don't think (and with me the text I wrote) that using AI is the problem; I think that the problem is using unsupervised AI (i.e. “not corrected by a human”, or however we want to call it). As you said, “I already try to be responsible enough and to understand that publishing LLM-generated texts with made up facts is pointless and harmful”: not everyone knows that, some people even think that AI is the new Google, and so this policy fixes that, while trying not to impose anything on responsible editors. --Grufo (talk) 15:48, 28 April 2026 (UTC)

•  Oppose Naut-rena (talk) 14:11, 27 April 2026 (UTC)
•  Oppose per Amir Aharoni. Hakimi97 (talk) 14:13, 27 April 2026 (UTC)
• Oppose. Sapphaline (talk) 14:52, 27 April 2026 (UTC)
•  Oppose I cannot support this issue until clear adjustments are made. Linh Huynh (talk) 16:28, 27 April 2026 (UTC)
•  Strong oppose -- why is disclosure "optional"? AI is well known to add, subtract, and otherwise distort content when translating or copyediting or summarizing, etc. See also: w:Hallucination (artificial intelligence). Why on earth would we be so blasé about editors using known-faulty tools? What benefit is there from not insisting on disclosure? As written, this policy makes no sense to me. ‑‑ Eiríkr Útlendi │^{Tala við mig} 18:13, 27 April 2026 (UTC)

  Initially I was undecided between “optional but encouraged” and “optional but discouraged”. I chose “optional but encouraged” as a compromise to meet opinions like yours (and also to facilitate verifiability, in some cases), and I would naturally lean towards “optional but discouraged”. Why? Because, as I previously mentioned, disclosing AI usage may be perceived as shifting or diluting user accountability. My instinct would say “if you use AI, keep it for yourself and have the courage of being fully accountable for what you write no matter how you produced it; if you publish something, you believe in it.” --Grufo (talk) 19:44, 27 April 2026 (UTC)

•  Strong oppose Bad content is subject to be challenged and deleted regardless of the source. Fundamentally the point of the AI policy put forth on multiple wikis is to stem the flow of bad content that human editors have to review. The human who generates and is responsible for reviewing it for compliance will shrug, say they were just trying to help, and move on with their lives, leaving behind an absolute mess for other human editors to clean up--if someone even wants to clean it up as part of their volunteer work. This would do nothing to reduce the amount of bad AI content that humans would need to review, and will make it easier for users to add low-quality information to Wikimedia projects since they can claim they "reviewed" it. Mcampany (talk) 19:41, 27 April 2026 (UTC)
•  Oppose As I like the idea of human controlled AI, I think this is not usable: Who an how to check that someone has checked the AI generated imput. --GodeNehler (talk) 20:00, 27 April 2026 (UTC)
  • Who an how to check that someone has checked the AI generated imput: This is pretty straightforward. You check it by searching for errors or shortcomings: it has errors/shortcomings → not reviewed; it has no errors/shortcomings → reviewed. --Grufo (talk) 20:11, 27 April 2026 (UTC)
•  Oppose. This explicitly conflicts with policies which have been established by a number of major wikis, including enwiki (cf. en:WP:EPAI, en:WP:LLM, en:WP:AITALK). It is also essentially a non-policy - it amounts to a vacuous statement that editors should follow existing policies. Omphalographer (talk) 22:47, 27 April 2026 (UTC)
•  Oppose I agree with Amir Aharoni's points about language used as well. Abzeronow (talk) 04:00, 28 April 2026 (UTC)
•  Support. This text is in line with the existing policy ("guideline") ru:Википедия:Нейросеть in Russian Wikipedia. Nikolay Omonov (talk) 04:48, 28 April 2026 (UTC)
•  Support, per. earlier. --Werter1995 (talk) 07:33, 28 April 2026 (UTC)
•  Oppose. With "Disclosure of AI use is optional but encouraged" this policy is practically unenforceable. Well very well (talk) 12:31, 28 April 2026 (UTC)
•  Oppose - Too weak and unenforceable. Whaledad (talk) 16:00, 28 April 2026 (UTC)
•  Support – Per Лоття. Jeroen N (talk) 17:45, 28 April 2026 (UTC)
•  Oppose pretty much per Andrei Stroe: "A proposal without any teeth in the face of the onslaught of AI slop. [...]". The item "Disclosure of AI use is optional but encouraged" has no force of a rule; it is de facto a hint. --Dan Polansky (talk) 05:32, 29 April 2026 (UTC)
•  Oppose, lots of projects have local policies in place. There is no need to impose a top-down policy on all of them. It's presumptuous to think that we can know and anticipate the needs of all projects. Alaexis (talk) 06:15, 29 April 2026 (UTC)
• With no opt-out??  Strong oppose. Imagine what this would mean for Commons or Wikidata. - Jmabel (talk) 17:58, 29 April 2026 (UTC)

  What would it mean? --Grufo (talk) 19:14, 29 April 2026 (UTC)

•  Support. IMHO this is the only sane way to deal with a tool. --Викидим (talk) 21:12, 29 April 2026 (UTC)
•  Support This is well formulated, I support this. As I stated above, forced disclosure could encourage hounding by ideological opponents of AI, therefore I prefer a solution that only encourages, but not forces disclosure. --Eddy 97 (talk) 16:20, 30 April 2026 (UTC)
•  Support but strengthen disclosure. Not to dillute responsibility, but as an integral and necessary bit of context for future review, estimation of the independence of multiple edits/sources, &c. –SJ talk  20:55, 1 May 2026 (UTC)
•  Support, per above. IdanST (talk) 10:56, 5 May 2026 (UTC)
•  Oppose, per Amir Aharoni. Sannita - not just another it.wiki sysop 11:25, 5 May 2026 (UTC)
•  Strong oppose: 1) AI is under strong and fast evolution, any rule could be obsoleted in short time, 2) Wikipedia is growing with freedom, without too many generalist top down policies witch are not considering local condition.--Bramfab (talk) 12:28, 5 May 2026 (UTC)
•  Oppose as per Amir.--Superspritz^{tell me} 20:50, 5 May 2026 (UTC)
•  Oppose Global policies should not be imposed except in cases where it is actually necessary to overide the autonomy of individual projects. In this case, it does not appear necessary. I assume that individual projects will take steps to prevent inappropriate use of AI without being forced to do so. James500 (talk) 20:11, 6 May 2026 (UTC)

Full ban

I propose the following to be adopted instead of everything presented above:

Use of generative AI is fully banned on all Wikimedia projects' main and talk namespaces. Users posting generative AI output in these namespaces may receive a single warning; after that, they may be subject to any sanction an imposing admin thinks is reasonable. This doesn't apply to talk pages of multilingual projects.

Sapphaline (talk) 14:52, 27 April 2026 (UTC)

Discussion

• Support. Sapphaline (talk) 14:52, 27 April 2026 (UTC)
•  Oppose. Naut-rena (talk) 15:44, 27 April 2026 (UTC)
• Support Maarten1963 (talk) 15:47, 27 April 2026 (UTC)
•  Oppose. Full ban seems unreasonable. Deltaspace42 (talk) 16:18, 27 April 2026 (UTC)
•  Oppose I find this really silly because there are no clear, specific guidelines, and not everyone can identify what constitutes AI-generated content. Furthermore, if users violate the rules and are blocked harshly without warning, it can have the opposite effect. Linh Huynh (talk) 16:33, 27 April 2026 (UTC)

  It's also quite hard to identify what constitutes disruptive/tendentious editing. This doesn't mean Wikimedia projects shouldn't regulate them, and rules being vague is a feature, not a flaw. "if users violate the rules and are blocked harshly without warning" - there is a recommendation to give the user a single warning before ban in my proposed policy. Also, "any sanction an imposing admin thinks is reasonable" isn't reduced to a ban. Sapphaline (talk) 16:42, 27 April 2026 (UTC)

•  Strong oppose -- While I appreciate the idea, and am a proponent of human-generated content for humans, this policy as written would result in capricious and incorrect application. We have no clear means of identifying AI, as opposed to an editor simply being wrong, or trolling / vandalizing. AI tools used to identify AI-generated content are frequently wrong, producing false positives and false negatives. I've been accused of using AI-generated content, despite literally having never used it, presumably due to how I write (including things like—(gasp)—the em-dash, which I am able to input via my regular keyboard). Moreover, not all AI-generated content is necessarily bad. So long as such content is vetted by a human capable of understanding it and fixing any errors, and so long as such content is disclosed as using AI-generated content, I think a full ban may be a mistake. ‑‑ Eiríkr Útlendi │^{Tala við mig} 18:20, 27 April 2026 (UTC)
•  Strong oppose --GodeNehler (talk) 19:57, 27 April 2026 (UTC)
•  Strong oppose: I repeat this again and again: the only thing that matters is that you carefully review with your own eyes whatever you publish and you are held accountable for it. AI can be used in countless innocuous ways (for instance, you have a list of names and you ask the AI to convert all names to their initials). It does not matter to me if you generate your texts rolling dices and waiting that something meaningful comes out or if you are a monkey hitting keys at random on a typewriter, the only thing that matters is the result and that you are held accountable for it. --Grufo (talk) 20:04, 27 April 2026 (UTC)
•  Strong support I don't think generative LLM usage should be used in any project. EdoAug (talk) 22:23, 27 April 2026 (UTC)
•  Support Kill AI-generated content. AI-assisted is another matter.--Blockhaj (talk) 23:46, 27 April 2026 (UTC)
•  Oppose as going too far. As written, we'd need to remove the sample images from w:en:Stable Diffusion, the chat logs from w:en:ChatGPT, and possibly even the screenshots from w:en:ELIZA. At a bare minimum, this needs an exception that allows for examples of generative AI on relevant pages. --Carnildo (talk) 00:02, 28 April 2026 (UTC)
•  Weak support Phrased as it is currently, this is too broad and general to ever be enforceable or helpful. However, I would still support much stricter bans on AI than are currently in place, or that are being discussed above. A global ban is, in my opinion, the correct course of action, but it needs to be phrased better than this is currently. TheDowningStreetCat (talk) 00:17, 28 April 2026 (UTC)

  I should clarify: I would support a ban on generating content with AI globally, as well as very strict enforcement on anything AI-assisted (even apparently safe tasks like grammar-checking, etc. which are still often unreliable today). TheDowningStreetCat (talk) 00:38, 28 April 2026 (UTC)

•  Oppose People should use the best tools for them, and AI translations in multilingual discussion are useful, and they're useful for a starting point for translation.--Prosfilaes (talk) 03:27, 28 April 2026 (UTC)
•  Oppose. Since we are talking about policy making on Wikipedia, strict ban to all large language model (LLM) generated contents seems like going too far and contradicts with the policies made by various language editions of Wikimedia projects. AI-associated policy should guide users to think into a more narrow and concrete idea of what is allowed, what is conditionally allowed, what is highly discouraged, and what is forbidden when dealing with AI content. Hakimi97 (talk) 03:37, 28 April 2026 (UTC)

Support. AI is actively harming knowledge. TlinaR (talk) 07:44, 28 April 2026 (UTC)

•  Oppose This is extremely vague and overkill, creating limited room for individual project exceptions. //shb (t • c) 08:32, 28 April 2026 (UTC)
• Neutral But interesting that this might be better than the policies above. Well very well (talk) 12:32, 28 April 2026 (UTC)
•  Oppose - I'd be very happy to prohibit and prevent all the bad and unthoughtful uses of LLMs (and other "AI" technologies). However, you need to clearly define which uses are bad, even if most of them are. So a blanket ban just won't work. It's not perfectly enforceable because no LLM detection mechanism is bulletproof. For the same reason, it's easy to abuse: one user can accuse another of using an LLM just based on intuition. (There are also arguments in favor of banning LLMs completely for reasons other than bad output: for example, because they use too much energy or because they are bad for traditional media and knowledge development outside Wikimedia. I acknowledge that these arguments are quite reasonable, but that's probably not why we're discussing this topic here. Besides, banning the use of LLM output in Wikimedia projects won't fix those problems.) --Amir E. Aharoni (talk) 14:20, 28 April 2026 (UTC)
•  Strong oppose – Makes no sense. Jeroen N (talk) 17:47, 28 April 2026 (UTC)
•  Strong oppose --Wortulo (talk) 18:45, 28 April 2026 (UTC)
• Neutral Going in the right direction but perhaps too stringent. Using GenAI to brainstorm ideas, which is also use, would seem tolerable. What is not tolerable is placing unedited or edited-only-a-little GenAI output into a wiki, and similar. And what about posting an example GenAI output in a page that is about that GenAI? --Dan Polansky (talk) 05:34, 29 April 2026 (UTC)
•  Oppose - lots of projects have local policies in place. There is no need to impose a top-down policy on all of them. Alaexis (talk) 05:45, 29 April 2026 (UTC)
•  Strong oppose I think we should regulate not ban ! Alkmen-Alesia (talk) 07:13, 29 April 2026 (UTC)
•  Strong oppose Best approach for making sure Wikimedia stays behind and will be far less relevant. This would greatly harm our projects. Also no need or legitimacy for users participating in this metawiki RfC to impose a policy like this onto other projects which may have come or come to other conclusions and have other demographics than at best the overall demography of international Wikimedia toward which they would then be biased. --Prototyperspective (talk) 13:49, 29 April 2026 (UTC)
•  Strong oppose. Does this mean a Commons user couldn't use Google Lens to help identify a photo? That when I am trying to respond to a question in Romanian (which I speak decently but not fluently) I am not allowed to use Google Translate to help me in phrasing a sentence? etc. This is ridiculous. This is like saying you couldn't pose a question to a friend of colleague. In either case, you are still responsible for verifying, for finding sources as appropriate, and for what you write, but there is no reason to confine ourselves to methods available to a monk in a cell. - Jmabel (talk) 18:05, 29 April 2026 (UTC)
•  Strong oppose Banning a useful tool makes no sense. It will only give advantage to malicious fly-by-night players that by nature can disregard the rules. --Викидим (talk) 21:20, 29 April 2026 (UTC)
•  Strong oppose See my comments above. Cannot be detected reliably; not all ways of AI usage are bad. --Eddy 97 (talk) 16:10, 30 April 2026 (UTC)
•  Strong oppose Too restrictive, there are certain usages of AI tools that are beneficial in a usual, everyday workflow of a Wikipedian (and other projects as well, but not that well versed in those :P). Some projects did ban AI alltogether, but that's in their domain to decide, not ours. Afterall, AI usage is inherently tied to content, which is something Meta community should not intervene with, or should regulate very carefully.--A09|(pogovor) 20:55, 30 April 2026 (UTC)
•  Strong oppose Мы не можем остановить прогресс. Запрет приведёт к сокрытию таких правок и маскировке под правки живых людей. Мы загоним себя в тупик. --VladimirPF (talk) 14:54, 3 May 2026 (UTC)
  • (We cannot stop progress. A ban will lead to the concealment of such edits and disguising them as edits by living people. We will drive ourselves into a dead end.) --VladimirPF (talk) 14:54, 3 May 2026 (UTC)
•  Oppose, per above. IdanST (talk) 10:57, 5 May 2026 (UTC)
•  Strong oppose let communities decide on their own, do not force a decision without them being considered. Sannita - not just another it.wiki sysop 11:25, 5 May 2026 (UTC)
•  Oppose I don't even understand what sense this makes. We should promote technology, not ban it. Maybe restrict usage, but there should be a local community consensus on those restrictions, as others have pointed out.—super nabla(🪰 msg) 11:48, 5 May 2026 (UTC)
•  Oppose Exaggerated proposal, nonsense at this level ZandDev (talk) 13:52, 5 May 2026 (UTC)
•  Strong oppose this is a rule that cannot be implemented effectively in the reality.--Superspritz^{tell me} 20:53, 5 May 2026 (UTC)
•  Oppose Global policies should not be imposed except in cases where it is actually necessary to overide the autonomy of individual projects. In this case, it does not appear necessary. I assume that individual projects will take steps to prevent inappropriate use of AI without being forced to do so. James500 (talk) 20:12, 6 May 2026 (UTC)
•  Support It is a reasonable option. AlbertRA (talk) 16:01, 17 May 2026 (UTC)

Alternative policy

Note that this policy is meant to replace both baseline and opt-out policies above. It is meant to be a default policy, being possible to be override on any specific wiki.

Should the following be adopted as a global policy?

Content

The use of large language models to generate, translate, or edit content must be clearly disclosed. Any LLM-generated content must be subject to human review and correction before publication; unreviewed content is strictly prohibited. When there is strong evidence that the content was not sufficiently corrected, it may be freely removed; sanctions may be imposed on editors for repeated infractions. Editors using an LLM for these purposes take full responsibility for their edits.

Discussions

The use of large language models to generate comments in discussions must be clearly disclosed, except for basic copyediting. Any LLM-generated comment must be subject to human review and correction before publication; unreviewed comments are strictly prohibited. When there is strong evidence that the comment was not sufficiently corrected, it may be freely removed; sanctions may be imposed on editors for repeated infractions. Comment review is optional when LLM-translating to another language; disclosure is still mandatory. Editors using an LLM for these purposes take full responsibility for their comments.

Well very well (talk) 13:21, 28 April 2026 (UTC)

Discussion

•  Support as proposer. Well very well (talk) 13:21, 28 April 2026 (UTC)
•  Support — Ivan-r (talk) 13:49, 28 April 2026 (UTC)
•  Weak oppose - It's pretty reasonable, but:
  1. It must explicitly say "review and correct". See my comment in another section where I suggest "reviewed and corrected" and not just "reviewed".
  2. Even though it can be understood from the current text, it's still better to mention explicitly that posting unreviewed and uncorrected LLM content is prohibited, and not just that it will be deleted.
  3. The part about translation is redundant. This is already required by the general terms of use, section 7g (it says "re-use" and not "translation", but translation is a form of re-use). This detail should be removed. What should perhaps be added is a general requirement to comply with the Terms of Use and with the wiki's content policies (whatever they are). --Amir E. Aharoni (talk) 14:31, 28 April 2026 (UTC)

1. Yep, you are right. Fixed it.
2. IMO "must be subject" is already an explicit prohibition of the contrary. But I agree with you that it makes it more clearer.
3. It absolutely is redundant, but (judging at least by experience with machine translation) very many editors don't know that translation keeps original rights and don't do that. Maybe it's better to keep it as a reference instead of being in main text.

Well very well (talk) 14:48, 28 April 2026 (UTC)

@Ivan-r @Amire80 @Sapphaline FYI that changes were applied to the text. Well very well (talk) 14:53, 28 April 2026 (UTC)

• Support. Sapphaline (talk) 14:41, 28 April 2026 (UTC)
•  Oppose - I don't think there is a need for global policy, but this seems the most reasonable. Naut-rena (talk) 14:53, 28 April 2026 (UTC)
•  Strong oppose. I am against any mandatory disclosure of how editors compose their texts, which risks diluting their accountability. Editors are the only ones accountable for what they publish. However they compose their texts, Wikipedia's rules about sources, verifiability, etc. will apply. --Grufo (talk) 15:41, 28 April 2026 (UTC)

  This helps making editors accountable for that. If editor explicitly says that a text was LLM-generated, it will draw more attention from others and will make a more thorough review (including sources, verifiability, etc), and if editor doesn't say that, it's pretty disruptive and such a policy allows sanctioning them. Well very well (talk) 15:47, 28 April 2026 (UTC)

  Why should a text composed with AI's help should draw more attention than any other text? I expect that a bad editor who doesn't check what they publish will be an even worse editor if they don't use AI. It is the bad editor who must draw more attention, not the editor who uses AI. --Grufo (talk) 15:52, 28 April 2026 (UTC)

  Because a bad editor who uses AI can produce much more "good-looking" texts much faster than a bad editor who does not. And it takes more effort to distinguish bad AI text from a good (AI or not) text than just a general bad non-AI text. That's why they need more thorough review. Well very well (talk) 15:56, 28 April 2026 (UTC)

  I don't agree. More often than not, with very very bad editors I would suggest to them to use AI more often, at least it would filter some of their fantastic claims out. Wikipedia must be robust (as it already is) and not rely on editors' self reports on how they compose their texts to filter bad content out. I repeat it again and again: the problem is unsupervised AI, which, thanks also to how easy it can compose texts, risks overwhelming our forces. As for the rest, we can easily manage anything that implies some kind of work from a human. Moreover, when used responsibly, AI can be a great tool. --Grufo (talk) 16:03, 28 April 2026 (UTC)

  I don't see much difference between "reviewing" and "supervising" AI. If you mean just making a good prompt — no, that's not enough to ensure there would be no hallucinations, and it was demonstrated many times. A post-generation human review is required.
  As for the rest, we can easily manage anything that implies some kind of work from a human.
  That's not true. If a human just fixed two words but didn't check any of the sources that's not "easily manageable".
  Moreover, when used responsibly, AI can be a great tool.
  That's true, but mandatory diclosure wouldn't at all harm responsible AI use. Well very well (talk) 16:16, 28 April 2026 (UTC)

  I am not against calling a supervision a review or the other way around, I am against forcing any disclosure on responsible editors (who are the majority of Wikipedia editors). I am wary of unsupervised/unreviewed/unattended/however-you-want-to-call-it AI, because it can generate large quantities of text before anyone has the time to correct it. Changing a dot at the end of a sentence does not count as reviewing it. In any case, there is no point in imposing disclosure of unsupervised AI: unsupervised AI must simply be forbidden, period. There are no cases in which forcing AI usage disclosure indiscriminately can be useful: with responsible editors is useless and only a burden, with less responsible editors it risks relieving them of their accountability, and with totally unsupervised AI it is simply pointless, since disclosing it would mean asking for the removal of what it produced (we must assume that who uses unsupervised AI will want to be as stealth as possible about it). --Grufo (talk) 16:28, 28 April 2026 (UTC)

  who are the majority of Wikipedia editors
  Wikipedia editors consist not only from experienced editors, but also good-intentioned newbies and bad-intentioned vandals. Unsupervised AI in their hands can overwhelm experienced editors who must check all of its generations even if their number is smaller, because they produce content much faster.
  Changing a dot at the end of a sentence does not count as reviewing it. In any case, there is no point in imposing disclosure of unsupervised AI: unsupervised AI must simply be forbidden, period.
  I completely agree with that.
  There are no cases in which forcing AI usage disclosure indiscriminately can be useful: with responsible editors is useless and only a burden, with less responsible editors it risks relieving them of their accountability, and with totally unsupervised AI it is simply pointless, since disclosing it would mean asking for the removal of what it produced (we must assume that who uses unsupervised AI will want to be as stealth as possible about it).
  Ok, I think I understand your concerns now, but still don't fully agree.

• For experienced and responsible editors it's not much of a burden to add something like "ai-gen" into edit summary,
• For the second category it would not relieve editors of their accountability — because it's specifically written in the policy that they take full responsibility.
• And for the non-responsible users we either can immediately remove that AI generations, or, in the second case, when they get caught again after being warned they should mark it, they will just get immediately blocked.

Well very well (talk) 17:02, 28 April 2026 (UTC)

Case by case:

• Responsible editors: What counts as AI usage? Does re-ordering a list in alphabetical order count as AI usage? Does asking AI to simply point out grammatical errors count? How do you draw that line? And how should a disclosure count more than the trust we already have on responsible editors? If a responsible editor were in a (rare) situation in which they deemed opportune to disclose AI usage, they would do it anyway, because they are responsible.
• For less responsible editors: It does de-responsibilize them. On Latin Wikipedia we struggle to contain AI-generated translation from English. There you can see not-so-responsible editors publishing new pages that are clearly auto-translated, with only poor human checks. If they were to write in their edit summary “ai-gen with human review”, it would be worse than how it is now: right now we can simply mark the page as “auto-translated” (no matter what they declare) and delete it after a week, but how would we mark a page that says in its edit summary “with human review”? What makes it visibly auto-translated is its poor quality, not what the editor says.
• Irresponsible editors: Editors whose contributions are constantly removed don't end well already.

--Grufo (talk) 17:29, 28 April 2026 (UTC)

• When there is text (not a few words, at least a sentence) generated by AI. No, it should not count more — it counts more only in the cases below.
• You would mark it in exact the same way, if you see there clear bad translation artifacts or LLM indicators. That's specifically written in the policy: «When there is strong evidence that the content was not sufficiently corrected, it may be freely removed»
• Yes, but with what I described in the previous message they would be easier to control by responsible editors.

Well very well (talk) 17:57, 28 April 2026 (UTC)

P.S. On Latin Wikipedia even in front of pages that were auto-translated via completely-unsupervised-AI we give the author one week of time to fix the page before deleting it. Because, once again, what matters is the result. If they publish something AI-horrible and they fix it through hard human work, we have no problems with that. Of course, if someone does that repeatedly, we kindly invite them to use their userspace to do their AI experiments before publishing. --Grufo (talk) 17:49, 28 April 2026 (UTC)

If they publish something AI-horrible and they fix it through hard human work, we have no problems with that.
Absolutely — because that's human review. And if you judge by the result that it wasn't sufficiently reviewed, you can still delete it by the policy. Well very well (talk) 17:58, 28 April 2026 (UTC)

•  Oppose – Runs into the same objections as the initial proposals. Also per Grufo. Jeroen N (talk) 17:48, 28 April 2026 (UTC)
•  Oppose -- Why the cut-out exception for translation?

If an editor participating in a discussion doesn't know a language well enough to write in it natively, and is using an AI to communicate, and that AI gets things wrong, the editor likely will not know that the AI got things wrong, and everyone else might assume that the incorrect content is what the editor intended. Unless there's a proper disclosure that AI was used, in which case readers will know that hallucinations / bullshit / weird mistakes are part of the mix, and not necessarily the posting editor's intention.

Get rid of the exception for translation, and I would be more copacetic with this alternative. ‑‑ Eiríkr Útlendi │^{Tala við mig} 18:32, 28 April 2026 (UTC)

If an editor participating in a discussion doesn't know a language well enough to write in it natively, and is using an AI to communicate, and that AI gets things wrong, the editor likely will not know that the AI got things wrong, and everyone else might assume that the incorrect content is what the editor intended. Unless there's a proper disclosure that AI was used, in which case readers will know that hallucinations / bullshit / weird mistakes are part of the mix, and not necessarily the posting editor's intention.
Because it's actually useful for writing in foreign languages and the editor cannot do a review in that case. But yes, it follows from the current version that neither disclosure nor review are required, while I meant only the second. Well very well (talk) 02:09, 29 April 2026 (UTC)

@Eirikr @TheDowningStreetCat Fixed that. Well very well (talk) 02:17, 29 April 2026 (UTC)

@Well very well, @TheDowningStreetCat, thank you both. Amending my vote. ‑‑ Eiríkr Útlendi │^{Tala við mig} 19:04, 29 April 2026 (UTC)

Neutral, as worded after the phrasing update by @TheDowningStreetCat. ‑‑ Eiríkr Útlendi │^{Tala við mig} 19:04, 29 April 2026 (UTC)

•  Oppose per Grufo. Лоття (talk) 00:41, 29 April 2026 (UTC)
•  Support but I would remove the exception for translated discussion messages. Those can still have major errors. fixed now. TheDowningStreetCat (talk) 01:05, 29 April 2026 (UTC)
•  Comment - this is a step in the right direction, but I'd like it to be more explicit that it's a recommended policy that any project can adopt on the opt-in basis. So rather than "global policy" it should be a "best practice" or "framework". Alaexis (talk) 05:53, 29 April 2026 (UTC)
•  Oppose Admins routinely delete the garbage that some people think absolutely needs to exist in all wikis. With this wording, every such deletion could be challenged for lack of "strong evidence". Huñvreüs (talk) 07:51, 29 April 2026 (UTC)

  With this wording, every such deletion could be challenged for lack of "strong evidence".
  I don't think that's really what happens/would happen, this would borderline be w:WP:POINT. Well very well (talk) 08:58, 29 April 2026 (UTC)

  A policy maker who writes When there is strong evidence should expect users to require strong evidence, and while it can be easy for a human to be convinced that a text is AI slop, whether this alone constitutes strong evidence is not obvious at all. This is really shooting oneself in the foot. Huñvreüs (talk) 10:06, 29 April 2026 (UTC)

  No policy is as strict as a legal law. All Wikimedia policies are sufficiently vague to allow applying common sense and consensus in non-obvious situations, but I believe that by most projects "multiple people are convinced that text is AI slop" would be considered a pretty strong evidence. Well very well (talk) 12:35, 29 April 2026 (UTC)

•  Support Alfa-ketosav (talk) 15:58, 29 April 2026 (UTC)
• Question: how does "default policy, being possible to be override on any specific wiki" differ from "opt-out policy"? I thought that was what "opt-out policy" meant. Am I wrong? - Jmabel (talk) 18:11, 29 April 2026 (UTC)

  That is an opt-out policy. By different I meant that it's different from the two proposed ones. Well very well (talk) 08:09, 2 May 2026 (UTC)

•  Weak support Judging the text on merits - and not the tools used to create it - is still the best approach IMHO (yes, I use dashes frequently). --Викидим (talk) 21:28, 29 April 2026 (UTC)
• Support as a first start, although I would disagree with the proposal of making it only a "best practice" despite it being much weaker than what is already being done on many wikis. Also seconding Jmabel's question, as this seems to be a new, weaker opt-out policy? Chaotic Enby (talk) 01:09, 30 April 2026 (UTC)
•  Oppose Every project should set it's own rules. There's nothing gained by having a global rule. Different projects experimenting with slightly different policies and learning from each other is the way content policy was always developed in Wikimedia and it should stay that way. — The preceding unsigned comment was added by ChristianKl (talk) 08:43, 30 April 2026 (UTC)
•  Support I would like to see a much more stringent default/opt-out-able global policy, but this is a pretty good compromise with the opposition to a stringent regulation. --Dan Polansky (talk) 11:16, 30 April 2026 (UTC)
•  Oppose this concept of what global policies are for. We should have global guidelines (see Artificial_intelligence/Guidelines for a separate constructive proposal), and should limit the creation of 'stick' policies ["don't do this or you will be punished"], which by definition divide and exclude. Also, global policies should not run counter to the canonical wiki workflow: publishing early and often, editing and revising in-place. The idea of "review and correct before 'publishing' " seems unwiki in three ways: not using the wiki for reviewing, not using the wiki and its edit history for correcting, and reframing 'publishing' as a big deal. A collaborative approach to the same idea might be: draft spaces for certain workflows. –SJ talk  11:49, 30 April 2026 (UTC)

  •  Oppose The AI can be false or getting an "error", might not suitable for any workflows and provided to any wiki

  AcoolDudeMan-editer (talk) 12:07, 2 May 2026 (UTC)

•  Support It doesn't make sense for humans to waste time reading probable spam AI-created discussions and talks. We already wasted our time reading a contemptuous AI-generated closure request for the Egyptian Arabic Wikipedia. Such cases shouldn't exist. Humans are expected to read other human-authored texts, particularly in discussions which need a human response. --Esperfulmo (talk) 18:58, 4 May 2026 (UTC)
•  Oppose, per above. IdanST (talk) 10:57, 5 May 2026 (UTC)
•  Oppose, I repeat myself: let communities decide on their own and do not force policies down their throat. Sannita - not just another it.wiki sysop 11:27, 5 May 2026 (UTC)
•  Strong oppose: 1) AI is under strong and fast evolution, any rule could be obsoleted in short time, 2) Wikipedia is growing with freedom, without too many generalist top down policies witch are not considering local condition.--Bramfab (talk) 12:29, 5 May 2026 (UTC)
•  Oppose, per above.--Superspritz^{tell me} 20:51, 5 May 2026 (UTC)
•  Oppose I suspect this proposed policy might result in editors casting aspersions on each other. w:en:Wikipedia:Casting aspersions#Selected examples 2008–2015 describes the problem of aspersions that resulted from rules about UPE, which have the same problem of requiring proof of off-wiki behaviour that is not recorded by diffs and is therefore difficult to prove. James500 (talk) 20:27, 6 May 2026 (UTC)
•  Support. But even for translation it should be disclosed if used significantly, and it must be clear that unpolished, AI-generated translations are not acceptable, either. --Paloi Sciurala (talk｜contribs) 09:26, 8 May 2026 (UTC)
•  Support. I am not sure about the disclosure bit, as your editor may use AI for spelling correction (or AI may be involved in other such ways), where the text is always "reviewed" and the editor may not even know about the AI input – and I would like to distinguish minor from major AI assistance, for which "AIgen" isn't enough. I also think that one shouldn't take part in discussions in languages one doesn't know well enough to do a best-effort review. However, this is the most reasonable of the proposals (Artificial intelligence/Guidelines has another focus), and I think the opt-out policy approach is the one to go for (as defined in the context of the original proposal). –LPfi (talk) 15:04, 10 May 2026 (UTC)

Minor edit focused policy

Strategy: since the translation topic produces controversy, attack the problem via the minor/major edit distinction:

The use of LLMs models to generate or rewrite content must be clearly disclosed, except for minor edits. Any LLM-generated content must be subject to human review before publication. Editors using LLMs for any purpose take full responsibility for their edits.

Tmagc (talk) 02:29, 2 May 2026 (UTC)

Discussion

 Oppose. Minor edits are a technical feature of MediaWiki; there is no universal standard for how they should be applied. Authorizing the use of LLMs to generate content, so long as the editor checks the "minor edit" checkbox before submitting that content, just seems like it'd encourage the misuse of that checkbox. Omphalographer (talk) 17:38, 2 May 2026 (UTC)

 Oppose Explicitly excluding minor edits doesn't make sense, since they only cover things like typos and formatting. Deltaspace42 (talk) 20:36, 3 May 2026 (UTC)

 Oppose. Minor edits are not really defined, and even if they were, they are still meaningful, and they must be human-verified just like all other edits. Amir E. Aharoni (talk) 01:42, 5 May 2026 (UTC)

•  Oppose, per above. IdanST (talk) 10:58, 5 May 2026 (UTC)
•  Oppose per Amir Aharoni Sannita - not just another it.wiki sysop 11:27, 5 May 2026 (UTC)
•  Strong oppose: 1) AI is under strong and fast evolution, any rule could be obsoleted in short time, 2) Wikipedia is growing with freedom, without too many generalist top down policies witch are not considering local condition.--Bramfab (talk) 12:29, 5 May 2026 (UTC)
•  Strong oppose per Amir Aharoni. --Torque (scrivimi!) 12:52, 5 May 2026 (UTC)
•  Oppose per Amir Aharoni. Authorizing the use of LLMs to generate content, so long as the editor checks the "minor edit" checkbox before submitting that content I don't think that the proposal should be interpreted in this way, but it certainly lends itself to be wrongly exploited. --ZandDev (talk) 13:58, 5 May 2026 (UTC)
•  Oppose per Amir.--Superspritz^{tell me} 20:53, 5 May 2026 (UTC)
•  Oppose Linh Huynh (talk) 06:46, 7 May 2026 (UTC)

 Strong oppose This would open a loophole that would increase the misuse of the "minor edit" checkbox. --Minoa (talk) 20:44, 11 May 2026 (UTC)

•  Oppose- as much content as is sometimes added by editors checking of "minor edit" I absolutely oppose this.--Wuerzele (talk) 12:08, 17 May 2026 (UTC)
•  Oppose -- Umarxon III (talk) 00:24, 18 May 2026 (UTC)