Comment by dragonwriter

> How can they distinguish from real people exploited to AI models autogenerating everything?

Watermarking by compliant models doesn't help this much because (1) models without watermarking exist and can continue to be developed (especially if absence of a watermark is treated as a sign of authenticity), so you cannot rely on AI fakery being watermarked, and (2) AI models can be used for video-to-video generation without changing much of the source, so you can't rely on something accurately watermarked as "AI-generated" not being based in actual exploitation.

Now, if the watermarking includes provenance information, and you require certain types of content to be watermarked not just as AI using a known watermarking system, but by a registered AI provider with regulated input data safety guardrails and/or retention requirements, and be traceable to a registered user, and...

Well, then it does something when it is present, largely by creating a new content gatekeepiing cartel.