If you strip a song down to just its rhythm, no pitch, no timbre, people can still tell you what emotion it's going for. Add the pitch back and it barely moves the needle. Timing between events does most of the heavy lifting. The actual "sound" part is almost window dressing on a temporal skeleton.
I think backlash against the stochastic parrot idea has always been based on a bit of a strawman — what's being parroted can be pretty abstract and broad in scope, down to reasoning strategies and personality. But with art in particular, I always feel like it's broadly imitating something from elsewhere.
In those AB tests in the linked essay for example, the options are all pretty prototypical genre music samples being compared against one another. Even the fact that they have pretty accurate prototypical labels says something about how prototypical they are, and you can always find something pretty stereotypical of a genre to compare against to play "gotcha" or make it more difficult. It's just not interesting or relevant to me to test whether you can tell the difference between cliched music samples from a human or cliched music samples from AI.
Real art is often like that — you have writers and musicians who follow genre conventions for a variety of reasons — but the most interesting art happens in the margins in ways that are unexpected.
The essay is worth reading for its argument that music is fundamentally a participatory activity — this is something else on my mind lately about live performance art in general, and the implications of that for understanding certain societal trends. But sometimes I think the discussion of AI and art is really missing the boat in other ways as well, ways that apply to AI in general.
alhwyn•1h ago