I feel like audio-level heuristics will be easier, but ultimately who's to say?

> Generative models synthesize sound mathematically. These synthesis methods leave unnatural dips, specific spectral noise profiles, or phase alignments that rarely occur in real, human-recorded audio

Then the slop merchants will simply move to controlling a DAW with AI and use the same software synths that everyone else does. It's a little more involved and slower, but far from hard.

Ultimately this isn't really solvable without a way of marking audio with a verifiable signature that it was produced by a specific human, with some kind of reputation algorithm.