Talk:Artificial intelligence/Policies by project

Wikivoyage

There is an ongoing discussion on Wikivoyage on AI, on a future guideline there.

Thanks, Globetrotter30 (talk) 15:12, 19 May 2026 (UTC)

LLM Paste Check - Question about the default wording to show in the global feature

The Editing team is working on an "LLM Paste Check" feature, for the Edit Check system. Our question for you, as experienced editors familiar with the broad variety of local policies, is how to word the default versions of the interface messages that are shown to editors (via the translatewiki process).

For context: the LLM Paste Check feature can detect when someone is copying and pasting LLM generated text directly into a VisualEditor edit session, because many popular LLMs automatically add specific hidden metadata. As currently designed, when LLM Paste Check detects this metadata, it will show a box warning the contributor about using a direct paste, and it will link to generic (or local) policies, and ask them to decide whether to keep the text they have pasted or to remove it. In addition to this in-edit feedback for the contributor, LLM Paste Check will also add Tags to all edits in which it was shown, in order to assist experienced editors in patrolling.

Note: As with all Edit Checks, the text strings and links can (and should where needed) each be locally overridden to adapt for local variations in policy, and any existing local guidance pages (instead of the generic default); but we need a default that works at all other wikis.

TLDR: We need your feedback about what exactly should be written in the default version of the (translatable) interface text strings, especially the "Title" and "Description", that are shown to the user who has directly copy & pasted content from an LLM. Two of our current drafts are shown in the screenshots and the text-examples below:

Default card content, draft text 1:

• Title: "Potential AI-generated content"
• Description (≤2 sentences): "What you pasted may include AI-generated content, which may be prohibited by Wikipedia policy. Help keep articles trustworthy by writing this text in your own words and confirming that references are reliable and support the claim(s) being made."
• Button1: "Keep it"
• Button2: "Remove it"

Mobile and Desktop draft text 2

When someone clicks Button 1, it will show them a survey asking (draft text 1):

• Survey description: "Please help other editors understand why you are keeping this text."
• Survey options:
  • This text is written in my own words and I've confirmed reliable sources support the claim(s) being made.
  • I wrote this text myself. It is not AI-generated.
  • None of the above applies.

Survey draft text 2

These survey options are primarily meant to help guide the editor to be sure about their decision. The results are stored, but are not currently available to other editors. However, each published edit that results in the user seeing the Edit Check will be Tagged.

Length reminder: Translations affect length, plus it needs to fit within a mobile-screen's Check card, and some newcomers might ignore/be overwhelmed by too many details and links. Please keep this in mind.

Mockup design of the UI to show how translations can affect length

Try a prototype:

You can test the LLM Paste Check feature by:

1. Open this patchdemo page.
2. Copy at least two paragraphs of generated text from the web interface of ChatGPT, Claude, or Gemini. (Note: It will not work with output from Google Search's generic "AI Mode" tab.)
3. Paste that content into the VisualEditor edit session.

Please share your ideas (and reasoning) for improving the wording used for any of the specific elements, especially the "Title" and "Description, or let us know if either of the current drafts are good. Thank you! And sorry for the length in my attempt to be clear. Quiddity (WMF) (talk) 00:09, 20 August 2026 (UTC)

Sorry if this is obvious but the flaws with the plan are that the wrong person receives the information and that the people who post slop cannot be incentivized to write anything. If you tell the user "we know you were naughty" they are just gonna try to find a way around that.

We need to tell the community "this guy is posting from an LLM", not (just) the editor. Also @Tamzin: explained all this already with the puffery check IIRC, people will just do the minimum effort required to no longer get the popup instead of actually doing the work. Because the people who paste from LLMs are generally speaking too lazy to do the actual work, I don't think there is anything you can tell em that would encourage them to actually write their own text. We'd need a central location were we'd keep track of the people who pasted from LLMs and then do a full investigation of every edit they ever made which makes it completely unworkable. Or just indef them automagically. Since neither is workable, we are stuck. Polygnotus (talk) 23:41, 21 August 2026 (UTC)

Well the edits will be tagged, which makes it easy to patrol/check them. InfernoHues (talk) 23:48, 21 August 2026 (UTC)

Anecdotally it influences some editors, and the tag helps cleanup.

@Quiddity (WMF) in practice the existing paste check for any multi-paragraph paste saying "don't paste" is quite counterproductive. It dilutes attention to the real warnings, and (as someone who does appreciate appropriate warnings) makes me want to aggressively do away with all notices. –SJ talk  00:53, 22 August 2026 (UTC)

Thanks for this! Some thoughts:

• "This text may include..." is probably not the best wording -- the person making the edit knows whether they were using AI or not, so best case scenario this does not tell them anything they didn't already know, worst case scenario it may confuse them. (Something like "This text appears to include"?)
• Not sure why there are two options that both convey the same thing ("I wrote this myself in my own words"), unclear what the distinction is
• possible to have a tooltip clarifying that Grammarly, DeepL, etc., are also AI tools? as this also confuses many people who are genuinely not aware
• Ideally have a substitutable link to the guideline for a given wiki, if possible

Gnomingstuff (talk) 02:36, 22 August 2026 (UTC)

Re: "substitutable link" - Absolutely. A core part of Suggestions and Edit Checks, is they can (and should) be adapted locally, especially the guidance-links within. There's a listing at mw:Help:Suggestion mode#Localize the links for admins to easily locate those parts. Quiddity (WMF) (talk) 18:11, 22 August 2026 (UTC)

Looking at m:Artificial intelligence/Policies by project, I think the default should probably be something like:
What you pasted appears to include AI-generated content. Please make sure that you are following this wiki's policy on AI use, as it may be subject to requirements or prohibited.

For Button 1, would we be able to see what they answer with? I think the following covers most of the bases:

• I wrote this text myself. I did not use AI.
• I wrote this text myself, and have used AI to make minor edits or corrections to it afterwards.
• This text is partially or completely AI-generated, and I have reviewed it for policy compliance and disclosed this.
• This text is an AI-assisted translation from a Wikipedia whose language I speak, and I have reviewed it for policy compliance.
• None of the above

Kowal2701 (talk) 15:11, 22 August 2026 (UTC)

apologies for not looking this up myself but I'm about to head out the door, are the cross-wiki AI translation policies uniform enough that we can mention something about fluency in both source and target language to bullet point #4? (Since as worded it would violate LLMTRANSLATE on enwiki, but the fact that it is the only translation option implies that that's ok) Gnomingstuff (talk) 19:31, 22 August 2026 (UTC)

Maybe This text is an AI-assisted translation from another Wikipedia. I speak both wikis' languages, and I have reviewed the text for policy compliance. Kowal2701 (talk) 17:21, 23 August 2026 (UTC)

also maybe condense "subject to requirements" to "restricted," shorter and clearer Gnomingstuff (talk) 19:32, 22 August 2026 (UTC)

yeah I considered that but thought it might be misleading as requiring disclosure might not be thought of as a restriction, though it says "may" anyway so that's probably fine Kowal2701 (talk) 17:22, 23 August 2026 (UTC)

I also think there should be a pointer to m:List of Wikipedias. LLMs are frequently used to compensate for a poor command of the language (especially at enwiki), and we should be directing people to wikis in languages they speak, where their participation would be more consequential and effective at furthering the movement's goals due to such wikis often being in relatively-early stages Kowal2701 (talk) 07:39, 31 August 2026 (UTC)

+1 Dw31415 (talk) 01:30, 2 September 2026 (UTC)

Thank you all for the comments so far.

I will boldly ping a few of the contributors to the listing page that I can see are established and multilingual users, in case you have any additional suggestions (but no pressure!). @Aca, @Krol111, @Pxos, @Nemoralis.

In a week or so, the Editing team will work to consider all these suggestions for the default, and followup here. Thanks again. Quiddity (WMF) (talk) 21:41, 26 August 2026 (UTC)

Thanks, interesting feature. As both survey options in draft text 2 indeed convey almost the same thing, I suggest maybe changing at least This text is written in my own words and... to This text reflects my own words and.... Krol111 (talk) 22:39, 27 August 2026 (UTC)

I am useless here, but I mentioned this discussion to a person (Fisuaq) who has been active on the Finnish Wikipedia on these kind of questions. They are busy in real life and replied only in Finnish here but they asked me to relay their opinion here. Fisuaq basically agrees with user Polygnotus and suggests that there should be established a Meta-Wiki-wide Noticeboard, where people could report at least the most obvious cases of large language mobsters, and that the N-board should also be linked to a Wikidata page. (If I have misunderstood something or quoted someone wrong, it's just a happy coincidence.) --Pxos (talk) 11:12, 4 September 2026 (UTC)

Maybe say "This text may include content generated by an AI tool such as ChatGPT, which is not allowed on Wikipedia. Help keep Wikipedia trustworthy by rewriting it in your own words and make sure the references support the claims." ~2026-51567-13 (talk) 22:37, 24 September 2026 (UTC)

Conclusion: We settled on the version given in the description of phab:T438521 (in the "Desired" column) for the default version, which went live with last week's deployment train, and is available for translation in translatewiki (and can & should be locally altered on each wiki to adapt for any specific local guidance details/links.) Thanks to all for the feedback.

Further feedback about the overall tool or feature is welcome at the relevant project talkpages (mw:Edit check or mw:Edit check/LLM Paste Check), or tasks in Phabricator, or other logical communication channels. Quiddity (WMF) (talk) 22:23, 27 September 2026 (UTC)

French Wikipedia update

They had a survery with changes. See fr:Wikipédia:Sondage/Évolution de la recommandation sur l'IA générative kulturgreek.fr. Thanks and regards. KurtR (talk) 03:41, 18 September 2026 (UTC)