Wikipedia talk:Large language models

Chatbot to help editors improve articles

After selecting text, the control panel on the right is used to give instructions. The responses by the AI model are presented in the chat panel on the left.

I wrote a user script called WikiChatbot. It works by selecting text in an article and then clicking one of the buttons on the right to enquire about the selected text. It includes many functions. For example, it can summarize and copyedit the selected text, explain it, and provide examples. The chat panel can also be used to ask specific questions about the selected text or the topic in general. The script uses the AI model GPT 3.5. It requires an API key from OpenAI. New OpenAI accounts can use it freely for the first 3 months with certain limitations. For a more detailed description of all these issues and examples of how the script can be used, see the documentation at User:Phlsph7/WikiChatbot.

I was hoping to get some feedback on the script in general and how it may be improved. I tried to follow WP:LLM in writing the documentation of the chatbot. It would be helpful if someone could take a look to ensure that it is understandable and that the limitations and dangers are properly presented. I also added some examples of how to use edit summaries to declare LLM usage. These suggestions should be checked. Feel free to edit the documentation page directly for any minor issues. I'm also not sure how difficult it is to follow the instructions so it would be great if someone could try to set up the script, use it, and explain which steps were confusing. My OpenAI account is already older than 3 months so I was not able to verify the claims about the free period and how severe the limitations are. If someone has a younger account or is willing to open a new account to try it, that would be helpful.

Other feedback on the idea in general, on its problems, or on new features to implement is also welcome. Phlsph7 (talk) 12:45, 12 July 2023 (UTC)

I meant to reply to this sooner. This is awesome and I'm interested in this (and related ideas) related to writing / reading with ML. I'll try to have a play and give you some feedback soon. Talpedia 10:18, 17 July 2023 (UTC)

Related: see also m:ChatGPT plugin. Mathglot (talk) 07:22, 18 July 2023 (UTC)

Whilst I rather like the ability of this nifty little script to do certain things, I do have some criticism. These functions strike me as extremely risky, to the point that they should probably be disabled:

• "is it true?" - ChatGPT likely uses Wikipedia as a source, and in any case, we want verifiability, not truth. I feel quite strongly, based on several other reasons too, that this function should be disabled and never see the light of day again.
• "is it biased?" - ChatGPT lacks the ability to truly identify anything more than glaring "the brutal savages attacked the defenceless colonist family" level bias (i.e. something that any reasonably aware human should spot very quickly indeed). Best left to humans.
• "is this source reliable?" - Same as the first one, this has so much potential to go wrong that it just shouldn't exist. Sure it might tell you that Breitbart or a self-published source isn't reliable, but it may also suggest that a bad source is reliable, or at least not unreliable.

I don't think that any amount of warnings would prevent misuse or abuse of these functions, since there will always be irresponsible and incompetent people who ignore all the warnings and carry on anyway. By not giving them access to these functions, it will limit the damage that these people would cause. Doing so should not be a loss to someone who is using the tool responsibly, as the output generated by these functions would have to be checked so completely that you might as well just do it without asking the bot.

The doc page also needs a big, obvious warning bar at the top, before anything else, making it clear that use of the tool should be with considerable caution.

The doc page also doesn't comment much on the specific suitability of the bot for various tasks, as it is much more likely to stuff up when using certain functions. It should mention this, and also how it may produce incorrect responses for the different tasks. It also doesn't mention that ChatGPT doesn't give wikified responses, so wikilinks and any other formatting (bolt, italics, etc) must be added manually. The "Write new article outline" function also seems to suggest unencyclopaedic styles, with a formal "conclusion", which Wikipedia articles do not have.

Also, you will need to address the issue of WP:ENGVAR, as ChatGPT uses American English, even if the input is in a different variety of English. Mako001 (C)  (T)  🇺🇦 01:14, 23 July 2023 (UTC)

You can ask it return wikified responses and it will do it with reasonable good success rate. -- Zache (talk) 03:03, 23 July 2023 (UTC)

@Mako001 and Zache: Thanks for all the helpful ideas. I removed the buttons. I gave a short explanation at Wikipedia:Village_pump_(miscellaneous)#Feedback_on_user_script_chatbot and I'll focus here on the issues with the documentation. I implemented the warning banner and add a paragraph on the limitations of the different functions. That's a good point about the English variant being American so I mentioned that as well. I also explained that the response text needs to be wikified before it can be used in the article.

Adding a function to wikify the text directly is an interesting idea. I'll experiment a little with that. The problem is just that the script is not aware of the existing wikitext. So if asked to wikify a paragraph that already contains wikilinks then it would ignore those links. This could be confusing to editors who only want to add more links. Phlsph7 (talk) 09:12, 23 July 2023 (UTC)

I made summaries/translations/etc it so that I gave wikitext as input to chatgpt instead of plaintext. However, the problem here is how to get the wikitext from page in first place. -- Zache (talk) 09:48, 23 July 2023 (UTC)

In principle, you can already do that with the current script. To do so, go to the edit page, select the wikitext in the text area, and click one of the buttons or enter your command in chat panel of the script. I got it to add wikilinks to an existing wikitext and a translation was also possible. However, it seems to have problems with reference tags and kept removing them, even when I told it explicitly not to. I tried it for the sections Harry_Frankfurt#Personhood and Extended_modal_realism#Background, both with the same issue. Maybe this can be avoided with the right prompt. Phlsph7 (talk) 12:09, 23 July 2023 (UTC)

Thanks for setting this up. I've recently had success drafting new Wikipedia articles by feeding the text of up to 5 RS into GPT4-32k through openrouter.com/playground and simply asking it to draft the article. It does a decent job with the right prompt. You can see an example at Harrison Floyd. I'll leave more details on the talk page of User:Phlsph7/WikiChatbot, but I wanted to post here for other interested parties to join the discussion. Nowa (talk) 00:02, 20 September 2023 (UTC)

Thanks for the information. I've responded to you at Talk:Harrison_Floyd#Initial_content_summarized_from_references_using_GPT4 so that we don't have several separate discussion about the same issue. Phlsph7 (talk) 07:44, 20 September 2023 (UTC)

Ran into a brick wall I thought might be helpful to know about. I've been working on the bios of people associated with Spiritual_warfare#Spiritual_Mapping_&_the_Charismatic_movement. GPT 4 and LLama refused to read the RS claiming that it was "abusive". I can see from their point of view why that is, but nonetheless, RS is RS, so I just read it manually. Between that and the challenges of avoiding copyvios I'm a bit sour on the utility of LLMs for assisting in writing new articles. It's just easier to do it manually. Having said that, the Bing chatbot does have some utility in finding RS relative to Google. Much less crap. Nowa (talk) 00:35, 9 October 2023 (UTC)

If we're going to allow LLM editing, this is a great tool to guide editors to the specific use cases that have community approval (even if those use cases are few to none at this point). I found it to be straightforward and easy to use. –dlthewave ☎ 16:06, 23 July 2023 (UTC)

There is no policy or guideline disallowing the use of LLM or other machine learning tools. No need for any approval unless that changes. MarioGom (talk) 17:29, 11 February 2024 (UTC)

Copyright of LLM output

Generative AI output is widely held to be ineligible for copyright (except under UK and Hong Kong law), but just in case, the terms of use for ChatGPT and Claude contain provisions assigning any copyrights to outputs to the user, meaning that the user would have the authority to license those outputs under CC-BY-SA. Meta AI's terms contain no such provision. I suggest that editors who contribute AI-generated text to Wikipedia only use GenAI services that grant them copyright ownership of the output. Qzekrom (she/her • talk) 18:01, 27 December 2024 (UTC)

LLM-generated content

I've noticed links containing "utm_source=chatgpt.com" in some articles (for example in Lucid_Motors). It's quite likely that the content supported by these sources was written by LLMs. While now there are only few such articles, maybe it would make sense to create some kind of a filter or alert? Alaexis_¿question? 23:28, 28 December 2024 (UTC)

If I understand it correctly, this is a UTM parameter to track where the user came from. For example, the article Lucid Motors gives one source as https://marketrealist.com/p/who-makes-lucid-motors-battery/?utm_source=chatgpt.com. If my interpretation is correct, this means that the user clicked on a link on chatgpt.com, which brought them to the website marketrealist.com. In that case, it doesn't say anything about whether the website marketrealist.com was created by an LLM. Phlsph7 (talk) 09:50, 29 December 2024 (UTC)

Probably you're right. Alaexis_¿question? 20:58, 30 December 2024 (UTC)

A filter might still be useful as someone could be having ChatGPT writing a section for them and then citing the sources. E.g. "ChatGPT, write me a summary of X with sources", and then copying and pasting the ChatGPT summary along with the sources. Photos of Japan (talk) 21:56, 30 December 2024 (UTC)

@Photos of Japan: to confirm, I'm accurately understanding what you're describing...

Are you imagining a user experience where Wikipedia is able to detect when someone is copying and pasting content that was "written" by service likes ChatGPT, Claude, Gemini, etc.? If so, what can you imagine happening next?

For context, I ask the above as a product manager at the Wikipedia Foundation who is helping to develop a system that is meant to offer the kind of "in-the-moment" guidance/feedback/awareness I understand you to be describing here. Further, we're in the midst of building a version of – what we're calling "Paste Check" – as I type this! In fact, as part of this work, we investigated whether we could reliably detect pastes from popular LLMs. PPelberg (WMF) (talk) 22:10, 27 January 2025 (UTC)

@User:PPelberg (WMF) If a link includes "utm_source=chatgpt.com" then that means it was copied from ChatGPT, but that doesn't mean any other associated text added alongside the link was also copied from ChatGPT, but it makes it more likely that it was. I would propose a filter flagging such edits so that a user patrolling recent changes who is interested in reviewing content added by LLMs (such as potentially a member of the WP:WikiProject AI Cleanup) is able to review it. Photos of Japan (talk) 00:15, 28 January 2025 (UTC)

@Photos of Japan I don't think it makes it especially more likely (as ChatGPT would most likely link to content it is familiar with from its training data, rather than content someone just created with ChatGPT). To clarify, utm_source=chatgpt.com isn't part of the address itself: https://en.wikipedia.org/wiki/Wikipedia_talk:Large_language_models and https://en.wikipedia.org/wiki/Wikipedia_talk:Large_language_models?utm_source=chatgpt.com are the exact same page.

However, ChatGPT might not be ideal in terms of recommending links (as it hasn't necessarily read the content of the pages), and such a link is good evidence that the edit should be reviewed by a human. I'm pinging @Queen of Hearts (WPAIC's edit filter manager) on this one, and cross-posting this just in case. Chaotic Enby (talk · contribs) 11:22, 31 January 2025 (UTC)

Here's a list of articles that currently have a utm_source=chatgpt.com link. Nobody (talk) 12:06, 31 January 2025 (UTC)

Oof, that's a lot. Checking out the first one, here. The source chat found is about the topic, but doesn't support all its assigned content (that presumably come from somewhere else in its knowledge base), labelled the link as dead for some reason, and got the publisher wrong. CMD (talk) 12:20, 31 January 2025 (UTC)

As someone who regularly uses large language models, but not for the purpose of contributing to any Wikimedia project, I have noticed that large language models can, and often do, pick unreliable sources. Google's AI overviews often list Wikipedia and blogs in its sources. I have noticed that Google Gemini sometimes lists Wikipedia as a source when I use it. If Wikipedia imposes a duty to use reliable sources, in my opinion, we should create an edit filter, and warn, if not disallow, any citations that involve links to large language models, such as ChatGPT, Google Gemini, Claude, and Microsoft Copilot. Wikipedia already lists some AI-generated articles as generally not reliable. See Wikipedia:Reliable sources/Perennial sources#Large language models. Z. Patterson (talk) 22:49, 1 February 2025 (UTC)

page_namespace == 0 &
added_lines rlike "utm_source=chatgpt\.com"

Chaotic Enby (talk · contribs) 23:18, 1 February 2025 (UTC)

@Chaotic Enby: I think this will work. I also suggest writing this.

equals_to_any(page_namespace, 0, 10, 118) & 
(
    llmurl := "\b(chatgpt|copilot\.microsoft|gemini\.google|groq|)\.\w{2,3}\b";
    added_lines irlike (llmurl) &
    !(removed_lines irlike (llmurl)) &
    !(summary irlike  "^(?:revert|restore|rv|undid)|AFCH|speedy deletion|reFill") &
    !(added_lines irlike "\{\{(db[\-\|]|delete\||sd\||speedy deletion|(subst:)?copyvio|copypaste|close paraphrasing)|\.pdf")
)

An edit filter manager can add more LLM links, if necessary. It would work similarly to Special:AbuseFilter/1045. Z. Patterson (talk) 23:43, 1 February 2025 (UTC)

Do we currently have an edit filter for references that cite Wikipedia? There's about 1,500 articles that link to wikipedia.org and once you get past our articles about ourselves most of them are inappropriate. Almost all of our 172 articles linking to the mobile version of our site are inappropriate as well. Photos of Japan (talk) 05:47, 2 February 2025 (UTC)

Filter 1057 (hist · log) Nobody (talk) 08:05, 2 February 2025 (UTC)

Thanks. I noticed that it doesn't include any other wikimedia-related sites such as Wiktionary of which we appear to have a few hundred citations to. Addition of a citation referencing Wiktionary doesn't appear to trigger any filter and Citebot will also add dates to it which I find curious given WP:UPSD highlights it as a 'Wikimedia-related website'. Photos of Japan (talk) 07:28, 3 February 2025 (UTC)

The filter was prettty much created by accident wenn modifying a LTA filter. (Here) They kept it, because it seemed useful for dealing with CIRCULAR issues. It wasn't really a discussion if references to, for example: Wikibooks, Wiktionary, Wikivoyage, Wikiquote, Wikidata, Wikisource should be on the filter too as CIRCULAR or SPS sources. Nobody (talk) 08:33, 3 February 2025 (UTC)

I see. There's few enough of them that I could probably manually remove most of them some day. Photos of Japan (talk) 03:52, 4 February 2025 (UTC)

Demon#Philippines

This section appears to be AI-generated and could use attention from someone more familiar with fixing or removing AI-generated output. Skyerise (talk) 13:56, 29 January 2025 (UTC)

"Wikipedia:PADEMELONS" listed at Redirects for discussion

The redirect Wikipedia:PADEMELONS has been listed at redirects for discussion to determine whether its use and function meets the redirect guidelines. Readers of this page are welcome to comment on this redirect at Wikipedia:Redirects for discussion/Log/2025 March 9 § Wikipedia:PADEMELONS until a consensus is reached. Jay 💬 06:17, 9 March 2025 (UTC)

When is it a crisis? Blocked LLM edit count increasing 10x per year

Data: https://quarry.wmcloud.org/query/92310 Cramulator (talk) 01:31, 30 March 2025 (UTC)

Maybe it could be a crisis when it's an everyday thing, instead of something that isn't even happening once a week?

This isn't really useful data. For example:

• These are decontexualized numbers. About 80,000 registered accounts got blocked last year, and a whopping 11 (yes, just eleven – not even one a month) accounts were blocked with an edit summary that included the letters "LLM". We usually block about 1,500 registered accounts a week and almost as many IPs. We are currently blocking about one (1) a week over LLMs. One a week is not a problem. Compare 1 a week for LLMs against [including registered+IP accounts, from 17–24 March] ~600 a week for spam, 600 for vandalism, 350 for socking, 300 for promotional usernames, 250 for disruptive editing... Overall, it looks like LLM misuse is the least of our problems.
• Last year's LLM rate is 275% as many as the year before (not 10x). It is, if you want to be dramatic, also an infinite increase compared to ten years ago, when LLMs were largely unavailable to the general public. But even if you want to panic about the slope of the curve, it's still a teeny tiny fraction of all blocked accounts.
• Edit summaries are not high-fidelity signals of the reason for a block. An LLM problem can be labeled as vandalism, or as a WP:NOTHERE problem, or as a WP:BLUDGEON problem (false negatives). A problem with POV pushing might be softened to LLM misuse (false positive). And this picks up any mention of LLMs, so it could be that someone has multiple problems, of which LLM misuse is only a minor component. Does the edit summary "Disruptive editing: Copyright violations, disruptive use of LLMs, mangled prose, inappropriate use of speedy deletion, unhelpful answers ate the Teahouse, improper usurpation of an article" give you the impression that the main problem is LLM misuse? That's not what I get from it. (There have also been multiple false accusations of LLM use, though I hope none of those have reached the point of blocking.)
• An edit summary can contain the string 'LLM' without referring to LLMs. For example, any edit summary linking to Wikipedia:Sockpuppet investigations/WTillman29 contains "llm" (though perhaps the script is looking only for the ALLCAPS string).
• Whether an admin mentions LLMs (and not, e.g., "AI") may be idiosyncratic/personal preference. That means that whether the number goes up or down depends on who's doing the blocking, rather than what the editor was doing.

So overall, my answer to "When is it a crisis?" is "not yet, and probably not any time soon". WhatamIdoing (talk) 06:45, 31 March 2025 (UTC)

I appreciate the thoughtful response. I'm suggesting that the trend is the issue here. It's probably best to revisit this in several months to see what that trend does. Cramulator (talk) 12:08, 2 April 2025 (UTC)