The detection problem is genuinely hard. Even desktop AI agents I've been working with recently can control Spotify, fill forms, navigate apps — all indistinguishable from human interaction at the OS level. If that's hard to detect at the application layer, detecting AI-generated music at the audio layer seems like a cat and mouse game that Tidal will struggle to win without self-reporting from uploaders.

I mean you can outsource it to users, which also allows you to organically make exceptions for AI music that is actually popular.

You really just want your users, who hate AI, to see a big "REPORT AI" button they can click. The problem you are trying to solve is the perception among your users that your platform is dominated by AI slop. So at the end of the day the only thing you actually have to figure out is what your users think is AI slop, have a quick trigger on un-popular stuff, and basically never enforce on popular stuff unless there is actually some controversy.

Maybe if enough AI produces self-report their work as AI, and enough non-AI producers are honest about uploading non-AI work, they'll quickly have the necessary amount of good-enough data to train good classifiers?

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

You are correct, but I think having a good policy - and trying earnestly to enforce it - is a good start, even if that enforcement is very imperfect.

Let's go with impossible.