Talk:Strategy/Multigenerational/Artificial intelligence for editors

Opt-out

There ought to be a way for editors to opt-out of interacting with "AI". Vexations (talk) 15:42, 30 April 2025 (UTC)

Relevant: A way to filter edits [by various bots] from diffs [and the Watchlist] – this would allow excluding changes made by ClueBot (which uses AI). Prototyperspective (talk) 18:31, 2 May 2025 (UTC)

They just shouldn't add this feature, at all. Wikipedia is the last remaining major resource for reliable information on the Internet. Adding slop to it, will destroy the entire project. --WizWorldLIVE (talk) 01:42, 4 June 2025 (UTC)

Which "feature" are you even talking about? What are calling "slop", a word like "trash", and roughly why? In case you misunderstood, articles aren't intended to be written by LLMs. Prototyperspective (talk) 11:23, 4 June 2025 (UTC)

I didn't misunderstand. LLM slop doesn't belong anywhere on Wikipedia, in any form. No summaries, no images, nothing. WizWorldLIVE (talk) WizWorldLIVE (talk) 07:06, 7 June 2025 (UTC)

"Large language models (LLMs) capable of summarizing and generating natural language text make them particularly well-suited to Wikipedia’s focus on written knowledge."

This claim, "Large language models (LLMs) capable of summarizing and generating natural language text make them particularly well-suited to Wikipedia’s focus on written knowledge." is just literally wrong.

LLMs do not summarise text. They shorten it. The meaning is not preserved, because facts are not a data type in LLMs. And they get it badly wrong a lot.

See e.g. the ASIC report on LLM summaries (PDF) - the AIs were worse than humans in every regard.

In similar tests, LLMs will happily reverse the point of a paper.

English Wikipedia doesn't have an explicit policy against LLM usage, but LLM text is widely shunned as absolutely the opposite of a compilation of factual information. The LLM text-shortening use case seems monstrously ill-suited to Wikipedia.

I'm sure people will still do it - but encouraging it is a ghastly error.

I would like whoever wrote this line to explain to the community their basis for this claim, in detail, showing their tests - David Gerard (talk) 16:16, 30 April 2025 (UTC)

The whole statement reads like vapid LLM-generated corporate waffle, there isn't a single concrete "plan" in there. Though using LLMs for summarizing content (or WP PAGs, as suggested elsewhere!) in any formal capacity is so obviously inane and time-wasting that I wouldn't be surprised if the WMF had already created a task force dedicated to developing precisely that tool... JoelleJay (talk) 20:42, 30 April 2025 (UTC)

I agree this document has flaws (including the lack of concreteness observed by JoelleJay). But David's literally wrong accusation is, well, literally wrong. To avoid splintering conversations: I have followed up at en:Wikipedia:Village_pump_(WMF)#WMF_plan_to_push_LLM_AIs_for_Wikipedia_content (where David posted a version of the same comment) with a reliable academic source for the capable of summarizing claim that David finds objectionable, and other remarks.

Regards, HaeB (talk) 03:17, 1 May 2025 (UTC)

If you ask an LLM a question, it doesn't give you an answer. It gives you something that looks like an answer.

For some questions, "looks like an answer" is good enough. For others, it very much is not. DS (talk) 02:04, 6 May 2025 (UTC)

Agree but you're talking about a different subject. The subject of this thread seems to be summarization, not asking a normal chatbot factual questions (Q&A). Thanks HaeB for the constructive answer there that clears up a few things. Prototyperspective (talk) 10:27, 6 May 2025 (UTC)

Concrete proposal for introducing editors to the actual capabilities and drawbacks of LLMs

[Crossposted from en:WP:VPW#WMF plan to push LLM AIs for Wikipedia content.]

About a month ago, I ran an extremely bold experiment where I took the top 68 articles with {{technical}} tags by pageviews per month, used Gemini 2.5 Pro to generate a paragraph of text to address their tagged sections or entire article, and posted it to their talk pages with full disclosure including source code asking human editors to review and revise the suggestion to address the tag. Objectively the project was a huge success, going by the number of fully human editors who have been addressing over a dozen of these tags, amounting to solutions of longstanding requested improvements for over a million readers per year so far. But the opposition was overwhelming, probably because I started with fifth grade (ages 10-11 years) reading level summaries without any source citations, which is well below the target reading level for STEM articles on the English Wikipedia. I feel strongly that if I had started with 8th grade reading level summaries will full source citations the outcome would have been very different.

One observation which was clear from the VP/M discussions is that some of our most respected, senior, and knowledgeable editors have very heterodox opinions on both the capabilities and drawbacks of recent LLMs. I am not sure what to do about this issue. When one of the most respected senior editors claims something like "LLMs just predict the next word," without regard to the world modeling in latent space and attention head positioning that accurately making such predictions require, I just don't know how to respond. However, I think there is one way in which the Foundation's R&D team could help introduce editors to the capabilities of LLMs in a way which wouldn't involve even the mere suggestion of content improvements, but would help one of our most important core pillar workflows for all edits to all articles.

Let's re-imagine ORES away from random forest classifiers of simplistic and easily gamed features, into a full LLM analysis of each watchlisted edit or new page being patrolled for quality, including a full attempt to verify both the existence of offline source citations and the correctness of online sources, as to whether they support the article text after which they are cited. This might require an extra click to save resources, but it might not, for example with Foundation self-hosting some of the new low or zero-cost for models capable of this task. Let's compare the results to legacy ORES to show what LLMs can do to help uphold WP:V. Cramulator (talk) 23:35, 1 May 2025 (UTC)

This entire page already reads like it was LLM generated ...

"AI excels at handling tasks such as information retrieval, translation, and pattern detection."

No, no it does not.

"Automating the translation and adaptation of common topics"

You are planning to do the exact thing you say not to do elsewhere to avoid alienating your editors. The actual human first approach you are mindlessly repeating throughout would be shifting funding towards these under-represented editor groups. Not giving them translated garbage to proof read.

"AI offers opportunities for generating valuable types of suggested edits"

And again. You say not to generate content elsewhere, because it is proven to be wrought with mistakes not acceptable for an encyclopedia - yet you try to sneak it in through generated "suggestions". This is no different from the existing and already mentioned problem of validating LLM edits. You are just making it part of the process. Why would you do that.

2A02:8109:9D9E:9F00:844F:4C0D:5861:438 17:40, 3 May 2025 (UTC)

It's pretty good at machine translation already and for pattern detection you may want to try one of those apps that detect the plant species from a photo or voice recognition or the anti-vandalism Wikimedia tools etc.

Do you even contribute to Wikimedia yourself or are you just here to complain without explanation in overgeneralized nonnuanced ways and without consideration of scarce volunteer time and efficiency etc? Suggestions have already proven to be quite useful to help people get started, motivate, and find more things to do. Prototyperspective (talk) 11:17, 4 May 2025 (UTC)

Use LLMs to root out LLM-generated content

I think improvements to the moderation process should be the first focus of any LLM implementation to the projects, which as has been pointed out by others, would basically mean an improvement to Wikimedia bots. I believe one of the enhancements that could be made to them would be an ability to cross compare additions by new/IP users to passages generated by LLMs, and if they have closely match the generated text the edit and account would be flagged for manual review. After an extended period of time, i believe the data accrued would be useful in developing a more automated process.

LLMs and AI in general are still in their culture shock phase by my reckoning, so i believe Wikimedia should take it slow, boil the frog so to speak. Going all-in on such a novel addition poses numerous risks that may not be evident until time has passed, and i'd rather not participate in a project that forces us to eat sunk-cost arguments all day. Anthropophoca (talk) 03:13, 4 May 2025 (UTC)

Proposal of research team to focus less on research? ;D

Hi, thanks for sharing your thoughts on AI and a proposal for a strategy for the Wikimedia Foundation. It can have impact on Wikimedia projects but it's still the community processes that define the amount of support through other means than volunteers (e.g., German Wikipedia doesn't have any of these, with good reasons for each of them: “support editors to detect vandalism on all Wikipedia sites, translate content for readers, predict article quality, quantify the readability of articles, suggest edits to volunteers”, except for the latter which was introduced through the new editor page without community consensus is is highly disputed). But I try to understand this strategy as a blueprint for WMF what to focus on in future developments that can be used by community.

Where I personally had a laugh was in section “Create more time for editing, human judgment, discussion, and consensus building.”. — This strategy also seems to be defined by the Research team at the Wikimedia Foundation that, if I read that correctly, suggest that volunteers do less research by themselves and focus on editing and post-processing. Would you make the same proposal when it would affect your research? ;D

Personally speaking, I am not a good editor on Wikimedia projects. I have spent hundreds of hours for my featured article on German Wikipedia. I'm much better at other means (communications, administration, organization). But why did I still do this? Because I loved to dig deeper into a topic, because I loved to learn new things, because I loved to write bibliographies, searching for books in libraries, and finding that one thing that was still missing. — And i do have the feeling that there are many more community members who edit Wikipedia because of that: research, summarizing the stuff they have learned into a readible article. Instead, it annoys me (having a MA in literature, translations and philology) a lot when I have to correct bad translations, poorly researched, summarized or unintelligibly written content (be it through AI-supported translation tools or because of other means) that made its way into articles or other pages. I am a volunteer. I can decide for myself with what I want to waste my time with (or not, as I said, that kind of research makes me smarter!). And if being asked, I would always respond that the pre-editing parts are essential for writing a good article, if not the most important part. And while gamification is an emerging topic (I love to play games myself and see easy success), I do have the feeling that also the next generation will be interested in good own research instead of using AI for likely partly rehashed content.

I could tell a similar story with regards to the “Engage new generations of editors with guided mentorship, workflows and assistance.” section where I personally love to mentor new editors in person, onwiki, and beyond where I am grateful for more support but also have the feeling that—without much need—some activities I love and I am good at are suggested to be supported or even partly replaced with AI. And I know countless of community members who focused on such activities, have established groups for that, etc.

To sum it up, for that reason I am interested in your research within the community about what they actually ask for support with AI and where I can find the results of this. Because I'm seeing a long list of WMF staff that was interviewed but not much interaction with the target of this strategy, volunteers and their interests, needs and wishes. Pinging my (disclosure!) friend @LZia (WMF): for understanding better what this is all about. —DerHexer (Talk) 16:02, 6 May 2025 (UTC) PS: I have seen the Phabricator ticket. PPS: I have also seen Leia's statement here, but it's not so much about the research nor mentoring.

This section is a good thing to take a closer look at but I think you may have misunderstood it while it is indeed written in a slightly unclear way so one can't say for sure. I think with Part of this time is invested in finding the information they need for their editing, discussion, or decision making. AI excels at handling tasks such as information retrieval, translation, and pattern detection. By automating these repetitive tasks they didn't mean using AI to automate research that should be done by humans but instead for example suggesting relevant templates, Wikipedia meta/help pages and possibly which section could be missing in an article. If there are further types of things, I still doubt what they meant is the main research that goes into writing articles – that is in part because they write repetitive tasks which such isn't. See also Wishlist focus area Repetitive tasks.
Moreover, I think with Create more time for editing what is meant is the whole process you described, rather than just the ultimately wikitext writing. This could become a difficult area and currently it's kept broad. the next generation will be interested in good own research Moreover, I doubt it's about 'taking away' anybody's tasks/activities such as mentorship – there's too few users doing that and it can be combined where users can quickly ask simple questions to a tool which e.g. quotes things from help pages and if that doesn't solve their issue or it's something more complex they interact with mentors. Prototyperspective (talk) 17:33, 6 May 2025 (UTC)

@DerHexer: Thanks for writing and I'm happy to expand and clarify:

• For me one of the key points about this strategy is when we talk about "choice" and "autonomy". So when I reflect about the points you raised, what I think about is: there are editors who really enjoy mentoring other editors and they're doing that work now, there are editors who really enjoy it and they wish they could do more of it and can't scale, and there are editors who don't enjoy doing it and yet there are newcomers (to the projects, or to new areas of work) who need mentorship. So in this case, the thinking is that we know AI does reasonably well on basic mentorship support and we can offer additional support to newcomers to new tasks or the projects through AI.
• On the topic of research ;): again, first and foremost I'd go to "choice" to explain the intention. What if you, who really enjoy doing some part of the research, had another option to do less work to get to some of the material you may need for writing an article (think about an example we take for granted: a lot of the information in the libraries is now available on the internet, and you don't need to check every single branch to know if they have a book or not; or you have access to search on the web to find resources you could not easily find if you were to write a similar article 25+ years ago.). What if you had the option to send some agents through AI to do some of the retrieval of knowledge for you and then you could spend more of your time digging through that material. If you have this option, you can still choose not to use it and find your information in another way. This option will also help some other editors who have more limited access to information to be able to pull information with less technical/physical work. The other way I think about is closer to what Prototyperspective shared above. Indeed there are a lot of "technical tasks" that are not about the research itself but about getting to the research that you can have the option to remove from your workflow (that's the intention and hope).

I hope this is helpful in clarifying the intention and scope. --LZia (WMF) (talk) 00:32, 13 May 2025 (UTC)

Thanks for the response!

• Maybe it would make sense to make it more clear that you are thinking of any kind of automation and not particulary of generative AI. When I'm reading about AI-supported mentoring, I am thinking of horribly impersonal messages from chatbots on one of the big platforms. I'm getting the feeling that you are thinking of other means of AI-supported mentoring/onboarding.
• Wrt research, I agree new means of finding sources are often very helpful. (Nevertheless, I found a very old book on my main topic Laocoön because I was browsing through the library branch by branch that was never digitalized nor usefully tagged. ;P) However, I have raised some concerns about the other parts of the section “AI excels at handling tasks such as information retrieval, translation, and pattern detection.“ (my mark). I fear again generative AI that will automate some of these repetitive tasks in a way that does more harm than good (extra checks, skipping important parts, wrong summaries, bias, halluzination, etc.). When you are more thinking of other ways, it would be good to express this more clearly, too.

Best, —DerHexer (Talk) 20:22, 19 May 2025 (UTC)

Suggestion: AI-assisted multilingual article enhancement

Hello,

I've read this proposal with great interest, and I would like to contribute an idea that aligns closely with its goals.

A major challenge I often notice as a Wikipedia reader is the disparity in article quality across languages. Some versions (e.g. English) are comprehensive, with images and detailed context, while others (e.g. Catalan or Spanish) may only have a few lines on the same topic. This unevenness limits access to knowledge for speakers of less-resourced languages.

I would like to suggest a system where AI tools (such as language models like ChatGPT) could assist editors by:

• Automatically identifying the most complete version of an article across all languages
• Using it to generate enriched drafts in other languages, including translated descriptions, related links, and appropriate media
• Presenting these drafts as suggestions that human editors could review, improve, and approve

The goal would not be to auto-publish content, but to support and accelerate human editing efforts — especially in smaller language communities.

Would this kind of initiative fit within the scope of this project? If so, I’d be happy to help refine the idea further or discuss potential pilot applications.

Thank you for your work and for considering this!

-- Tuamir (talk) 11:53, 13 May 2025 (UTC)

the disparity in article quality across languages […] while others (e.g. Catalan or Spanish) may only have a few lines on the same topic. This unevenness limits access to knowledge for speakers of less-resourced languages. Agree. I think there is huge potential in this area.

You may be interested in this proposal which is very similar but has some differences and there can be (and are) multiple approaches regarding this subject (the learning from the postediting and flagging could also be used in an approach like yours). For example, while there are some potential upsides, downsides of your approach include low scalability, in practice it doesn't work well when the article in target language is not super short because it would have to kind of inject content and alter existing sections rather than just e.g. adding entirely new sections, and that when flaws like mistranslations of ambiguous words or names are fixed, that only benefits that one particular article instead of also flagging similar phrases in other articles. For suggesting appropriate media, media set in Wikidata items could also be used. Prototyperspective (talk) 12:45, 15 May 2025 (UTC)

However, I don't think ChatGPT and similar models can be used for this; it's largely machine translation models I think; I don't know if there already is a model to identify the likely highest-quality comprehensive Wikipedia article and that would be a new system that is specific to Wikipedia. Prototyperspective (talk) 12:48, 15 May 2025 (UTC)

Mixed impression, some comments

This sentence left a mixed impression. On the one hand, these are correct and even pleasant words that you want to hear and that are hard to argue with. On the other hand, there are serious doubts that a strategy is possible in the field of AI, since everything is very turbulent. In addition, the Wikimedia Foundation is not a corporation or a state that would really need strategies (and some of the text is more aimed at editors than at the Foundation's employees). Overall, the "strategy" is more conservative than I would like. Some comments on the text:
1. "AI excels at handling tasks such as information retrieval, translation, and pattern detection." It depends on the topic and the depth of immersion in the topic and also the language. In general, machine translation is not the best option and also requires careful human control. en:WP:MACHINE. Even if the translation into a given language is of fairly high quality, then, in many cases, on average, readers in this language are more educated and more demanding of the quality of the text (AI and automatic translators are trained on a corpus of texts, but if there are a lot of good texts in a language, then readers often have a high level of education and skills). But for some reason, the problem of translation is not even mentioned. As for the search, AI searches only in digitized literature and, of course, makes many mistakes. In general, in order to use AI for anything, you need to be able to do it anyway and human experience is needed.
2. The paragraph about "Create more time for editors to share local perspectives or context on a topic" seems strange. If editors face pressure to increase content (this is true), then what is the quality of this content? See also en:Goodhart's law and en:Campbell's law. If the editor simply translates from another Wikipedia, then what does the adaptation of general topics have to do with it? The article will simply be translated, possibly without even checking it. Then, with high probability, it will remain that way for many years without adaptation, updating, or significant improvement. Adapting topics and updating articles is also a lot of work, and other languages ​​may not even have the correct terms for some concepts or objects.
3. The paragraph about "Content generation vs. content integrity" also raises questions. It says about "moderators (?) and patrollers", but the problem is broader: articles need to be updated, improved, or rewritten, which means there needs to be a sufficient number of editors working in the subject and ready to support the articles.
4. "Strengthen Wikipedia's position as the most trusted source of encyclopedic knowledge in more languages", "Wikipedia can be that backbone of truth", "they will continue to seek the truth", "to be the arbiters of knowledge" and other such rhetoric. According to Wikipedia's recommendations, it and its sister projects are self-published sources. It is also unclear how well Wikipedia can, if it should, establish the "truth" or be an arbiter of knowledge in a changing world. When comparing the level of trust in different encyclopedias, it is also not at all necessary that Wikipedia is now or in the near future the most trusted (although it may be so now).
5. "AI offers opportunities for generating valuable types of suggested edits that make sense for a new generation." Suggested edits have previously caused great discussions. Many new editors perceive such edits as an obligation, like "passing a level in a game" in order to be allowed to continue editing. Therefore, they often make useless or even harmful edits, not realizing what is expected of them. Even if a person understands that this is training, it is not a fact that he is ready to actively participate in it. This interactive training can be really interesting, like a beautiful toy, but is there any benefit?
At the same time, experienced editors do not want or cannot spend time checking such edits. For example, they do not want to offend. I'm not sure AI will help much with this problem, especially since new editors even of the same generation are often very different. Possible questions and interests differ, for example Proeksad (talk) 18:24, 21 May 2025 (UTC)

Comment from German Speaking WP

"Humans First" as strategy of Wikimedia Foundation

It is very necessary to start a broader discussion about this topic so as not to be overrun. Strategy is a start point, but what's to do should become more clear means the Go's and NoGos. In my opinion, there have already been some approaches and projects funded by the Foundation, but they lacked a strategy and therefore did not really achieve the goals (e.g. Automoderator) or were already seen as problematic and incompatible with the strategy at the planning stage (e.g. Abstract Wikipedia).

In our German project AI and Wikimedia we have collected, that the strategy was taken up with interest. But the boundary to use as an aid must be defined more precisely. In German we try to specify this. We assume that nothing can be implemented without consensus in the communities. Therefore we prepare a survey for our WikiCon in October. It would be great to get a little further until then - and I have proposed to the program group, that somebody of you explains the strategy there. I am sure that your browsers can also translate German into English. Let's go ahead to protect our hobby and the world knowledge as well. --Wortulo (talk) 07:19, 1 June 2025 (UTC)

• Sometimes not over-specifying things and keeping it broad is a good thing instead of trying to specify things prematurely too early; I don't see why
• I don't see how that image on the right is relevant or useful here
• Don't know why you think that about Automoderator; I thought it was fairly successful and is about to be scaled up, improved and implemented more widely
• "must be defined more precisely" is not true and also not by WMF at global scale or via some process that is not based on rational discussion
• I care not egoistically about my activity as a hobby I enjoy but about the success of open knowledge and the importance of it such as when people can freely educate them in their own language on important issues like say Water scarcity

Prototyperspective (talk) 13:33, 1 June 2025 (UTC)

It was more an information for the Foundation, what happens in German WP. Strategy is really a good start, but it must be implemented, integrate all other activities and show the way how to do this - at the end. Otherwise she will be forgotten. I want to keep the debate alive about this - next point could be our WikiCon in October.

Regarding Automoderator we had this presentation of Sam Walton in English - translated with discussion also before and after (see project archive). We will not use, seems not to be ready and good enough, but sounds also like further development has been ended. No majority also for too much automatism. The goal itself suits the strategy and we need something to help, but as a partner for authors and admins. A hobby (that means voluntary) can also be useful and the picture is my illustration of strategy's intention - generated by AI, but the human is the "boss" with the good slogan of the Project "Humans First" - no cyborg :-) Furter discussion between us better in the German project? --Wortulo (talk) 17:06, 1 June 2025 (UTC)

Ok, I haven't watched that video by Walton so I don't know why you conclude Automoderator would be more or less a failure from it. What I've heard or thought about it is that the results are relatively good and it's already deployed at 8 Wikipedias with plans for more after some improvements and adjustments. Yes of course this hobby is useful and it's not what I addressed, I just wouldn't think of it in terms of it being for our personal pleasure as goal or value but for the reasons why we think it's important and worth our time; and there's way too much work to not improve time-efficiency and use tools where well possible. If editors don't need to spend as much time anymore on patrolling recent changes for example, they can do other things and world knowledge is protected by enabling them to do so, especially given that human resources here are relatively very limited even within ENWP. Prototyperspective (talk) 17:29, 1 June 2025 (UTC)

I am not the "judge" regarding Automoderator, summarise only our conclusions and argue why we do not use it. The 8 projects using it are not our biggest ;-) (no EN, DE, ES, FR, IT) A discussion in advance would probably have helped to better align this with the needs - and with the new strategy, this should happen first (identifying the needs of the communities and tailoring them accordingly). We don't need to talk about the targets, they make sense. It's about whether it really works that way and works interactively, not autonomously (main critics). But look at this discusion (unfortunately in German) regarding Abstract Wikipedia. Here I ask myself whether this is at all compatible with the strategy - or even runs against it. Wikidata is nowhere near ready to be used to create reliable articles and we loose quality. --Wortulo (talk) 18:13, 1 June 2025 (UTC)

Just briefly regarding Abstract Wikipedia I agree; I think it's something that is not viable, maybe a pipedream one could say and I don't see an issue with such experimental project if people think it's worth their time (and Wikifunction is a separate project that could be useful) but there is a problem if people think other projects with the same or a similar goal but a far more viable method to achieve it shouldn't be tried since there already is AW; in that way it could curtail innovation/progress. It's not as much about reliability though than about that an article that is 3 sentences long, consisting of pretty plain data, isn't really useful when the English Wikipedia article is 3 pages long. Lastly Automoderator may not work well on DEWP but that could be an issue of required adjustments or training data or sth of that sort. Prototyperspective (talk) 19:42, 1 June 2025 (UTC)

Wikipedia and AI circle

If we agree that a strategy should guide and coordinate all future AI activities: deciding what we need and who can best develop it, then that’s a good approach. We need strong solutions now, especially as generative AI is advancing quickly.

However, we also have to consider funding and available resources. As long as we receive donations, we should use them to create something valuable. If we have permanent staff, we must use their time and skills efficiently.

I’m not focusing on fears like the potential impact of a Trump administration or how new chatbots might affect Wikipedia’s visibility and donations. Instead, I’m thinking long-term, which is the purpose of a strategy. The strategy will be reviewed and improved every year, which is positive.

The process for discussing and updating the strategy should be clearly defined and include input from the community. I’ve added another diagram in English because we need to analyze this complex change process carefully. Wortulo (talk) 07:42, 2 June 2025 (UTC)

1. We adopt a human-centered approach. We empower and engage humans, and we prioritize human agency.

Please specify this by saying that our AI models "....respect the subsidiarity in the individual Wikimedia community projects."
I know it's tempting, but don't try to impose a one-size-fits-all AI approach on the communities. Please keep subsidiarity of the communities as a top-priority, also when working on improving Machine Learning in our movement. Ciell (talk) 08:51, 8 June 2025 (UTC)

Copyright position

I would like to raise a point here that I think Wikimedia (and the general community) should consider very carefully before rolling out any GenAI-based features, especially user-facing.

Current models trained on Wikimedia data are restricted by copyright (since the license is CC-BY-SA). That means that if a model like ChatGPT spits out something that is substantially similar to Wikipedia content, it has to credit Wikipedia or it is in breach of copyright. What's more, that output then needs to be licensed CC-BY-SA. Most AI systems have worse copyright concerns than this, but it is a possibility that this is something WM would like to enforce in the future.

This also applies to the models trained on WP data. If the model is distributed at all, it should be distributed CC-BY-SA, since it's a derivative work. If the model is released under a more permissive license (which is true for almost all open models), the copyright is violated.

The problem is that if WM also rolls out lots of its own features based on LLMs trained on copyrighted content (case in point, Cohere's Aya, which was trained on CommonCrawl data), it paints itself into a corner. I'm not a legal expert, but I think it would be a complex court case where you try to get somebody to stop doing something that you yourself are also doing at the moment (the WP content may be covered, but the same principles apply for any, say, StackExchange content that the model spits out). Even if it would be legally possible, it would look very bad.

In short, Wikimedia should be very aware that by rolling out AI features based on models trained on unlicensed data, they are very much picking one side and not the other, and possibly committing to not ever enforcing the CC copyright in GenAI use. PeterBlm (talk) PeterBlm (talk) 18:01, 11 June 2025 (UTC)

Wikimania 2024 session

Hi! In the Acknowledgements you mention a Wikimania 2024 session with some of the affiliates and volunteers where perspectives were shared and feedback was given. Do you mean this one: https://app.eventyay.com/talk/Wikimania/talk/UUCZXT/?

If yes, I think it may be useful to link to it from the text, as it wasn't obvious to me what session you were referring to. I wasn't sure about doing it myself directly though, because I didn't know if the document was meant to be collaboratively edited. Diegodlh (talk) 23:51, 26 June 2025 (UTC)

AI Tools to identify errors and outdated articles

In the German-language Wikipedia, an investigation by a newspaper that used generative chatbots to search for incorrect and outdated articles has caused quite a stir and - not surprisingl - found a few. This gave rise to the idea of developing a tool that authors can use interactively as an aid. The proposal and more information can be found here:

• wish (English)
• Wish (German)

This leads me to 2 general questions:

1. Is it planned to develop also tools asap? (Giving Wikipedia’s editors time back by improving the discoverability of information on Wikipedia to leave more time for human deliberation, judgment, and consensus building) As discussed with WMDE there are no capacities to do this soon because of a long planning process.
2. Is it planned to discuss the strategy with communities in general, because we have the problems and must solve them :-)

F.e. for me to develop an own AI model will not be as good as to use the best existing models (reasoning, deep research), but interactive by authors.

• Input: article or articles by categories or date of last edit, or...; specific prompting to receive the severe errors or outdated articles for the domain, the author works with.
• Output: a list of articles, ordered by any priority.

This is a question of interface and may be, we have to get in contact with the developers of such. We have in August a meeting with one author of the newspaper. We will ask them about howto and prompting. @M2k~dewiki, Sinuhe20, Johannes Richter (WMDE), DaWalda, and Salino01: Thanks --Wortulo (talk) 08:38, 16 July 2025 (UTC)