I honestly think this is a fools game - maybe it will work in those very very early stages of generative AI we're in now, but there's no way this can work reliably going forward. If anything, I imagine it will soon start yielding false positives against human generated work too.

Probably not. This is gross simplification of how counterfeit photocopied money is detected, but human visual acuity for yellow dots on white paper is pretty weak compared to red dots, so color photocopiers and color printers have/had firmware that would seed money with extraneous yellow dots. Computer vision has no such visual distortion and as such computers see counterfeit money as covered in yellow dots.

In a way its simpler with AI. AI is the ultimate groupthink tool and record keeping is simple. Simply ask the AI if this is the only possible output that could have been generated that's consistent as of March 17 2023 based on political limitations and censorship rules and artistic fads trends memes and styles of that era. The smaller the AI contribution (perhaps an icon bitmap) the (edited: easier) harder it is to hide AI involvement and the smaller the damage caused by AI to the copyright status of the work. The larger the AI contribution the easier it is to detect, but the larger the damage AI is causing, so it balances out.

As a concrete example, today, ask an AI "Please write a Harry Potter book" you will get a story that's extensively filtered and censored and bowdlerized to March 2023 political / cultural standards (hmm who's standards, the book buying public or some other group? Most people do not like our current censors... but they are in total power right now...), it would NOT look like a 1997 book at all, books from that long ago are only suitable for public book burnings now. In theory it should be possible for an incredibly politically incorrect AI to be permitted in 2033 solely for lawsuit discovery purposes to "Please write a Harry Potter book adhering to what we now call the hate filled politics of March 2023", then compare the md5sums, they match. Done, AI detected.

There's not many AIs and there's not many people permitted to grind their axes by censoring those AIs, so its a pretty small solution set.

There's an interesting political aspect to banning copyright of AI production, anything turbo-woke can be assumed to be AI generated and as such uncopyrightable, so the only way to make money in the 2030's will be to sell remakes and sequels of "Birth of a Nation" and "Mein Kampf" because an AI would never be permitted uncensored expression, so those cannot be AI produced and must have been made by humans, and as such only "right wing" content can be copyrighted and sold for huge profit.

right, but what do you do once I've taken my AI generated image, run it through illustrator's image trace, and performed just enough automated manipulation, then copied it through a format or two?

Writing a method to scrub this stuff is trivial.