It won’t be easy to detect just from single posts, which is why I suggested keeping statistics on how often users do it — possibly by things like ctrl-v, right-clicks, tab focus, and text insertion (separate scores). The goal would be to see patterns from users, so when it comes time to confront them, there is evidence instead of just guessing.
The final moderation decisions would still be made by humans, but the main difference is that those humans would have more data to make those decisions faster and more accurately. Not every case needs to be confronted, but some do.
There could be ways to detect different kinds of motivations. A post where the composer is open for a long time, and there are many pastes after the tab loses focus, probably isn’t AI text injection.
The data that Discourse provides already has made moderation decisions much easier — even just basic things like location of IP address, combined with username, first post content, whether they typed too fast, typing_duration_msecs, composer_open_duration_msecs, etc.
I think a few more fields there would be useful, if Google’s feature takes off. Even if Discourse doesn’t build a sophisticated detector for AI, collecting that data in the database could make it easier for other people to develop plugins.
I didn’t notice it before, but just saw an “AI” thing pop up here that translates to English. I personally wouldn’t do it like that in my forum. I’d have the users write in their own languages as the raw content, and then run a translator on that before presenting it to the other users. That’s what I meant above in my comment about future translation software being better than today’s. By storing the user’s native language, the better translation software of the 2030s and 2040s could be used on it, and the content could be read by users who speak that language without translating it into English and back again.