This task involves the work of implementing the tracking necessary for patrollers/reviewers to identify edits that involve someone pasting content in an edit session.

Stories

As an experienced volunteer motivated to maintain the quality of content available on Wikipedia, I want to be able to identify edits that involve people adding new text they pasted from another source, so that I can review these edits aware of the risks they might carry and actions I might take in response.

• Where "risks" here includes, and is not limited to, COPYVIO, content copied/pasted from generative AI (e.g. ChatGPT, Claude, etc.) [i]

Open questions

• How/where might this information be made available? E.g. via edit tags? Another system/view?
  • Related: to what extent – if any – will it be possible to aggregate edits based on this information? Instead, might this information only be available on a per edit basis?
• What – if any – rights/permissions/account characteristics might someone need to "possess" in order to gain access to the information this ticket ventures to introduce?
• What conditions would need to be met for the software to label a published edit as including pasted content? E.g. might this label be reserved for published edits we can prove (in some way) includes posted content?

Considerations

Making information of the sort this ticket describes available ought to be considered with care.

Reason being: offering this information via an edit tag would involve us labeling edits based on what happens within an edit session rather than on the basis of the contents of a published edit, as has historically been done. [ii]

References

Volunteers who see value in "labeling" edits that include pasted content for the purposes of detecting potential AI writing.

• Example of a volunteer (en.wiki) suspecting a talk page comment to be written by/using AI: source
  • Thank you to Kelsi Stine-Rowe for spotting.
• "Check this out, someone says that I'm an LLM. I spent an hour writing this with no LLM assistance." source: en.wiki
  • A volunteer sharing a case where another volunteer labeled/assumed/concluded the comment they published on a talk page was written by AI
• From @Venuslui "Do you think that the Paste Check Tool can be used for fighting against AI/ChatGPT writing? Some admins from the zh community are asking as they found out that some new users used AI-writing when they edited on Wiki, which is not ideal for our writing style."
• From @IKristiani-WMF "The Indonesian Wikipedia would like this. There was a discussion about the rising amount of AI generated content there that someone suggested several tags for this. The paste check feature will help!"
• From Sonja Perry: "I know we had some arguments against that in the past, but I've been speaking with [English Wikipedia community] about their Articles for Creation process and they are voicing serious concerns around detecting AI, and I see that as an opportunity to get buy-in from the en wiki community on article creation early."

i. Volunteers at en.wiki (and perhaps others) are already compiling signals they can use to identify "AI writing"
ii. To date, edit tags have been reserved for edits based on the contents of a published edit or based on the contents of an edit someone explicitly attempted to publish (as in the case of Abuse Filter).