Sonic Truth: Can Artificial Intelligence Make Meaningful Music?

Hey everyone,

Welcome to the very first instalment of Sonic Truth! Evan writes:

Do you think that artificial intelligence can create meaningful music? Will people like the music AI creates if it’s missing the “human element” to it?


Dear Evan,

So I reckon you can divide the music writing process into two rough chunks. One is applying knowledge and problem solving; the other is injecting an emotional quality via the imagination or personal experience. This first chunk is the kind of thing computers are really, really good at.

We’ve been able to teach computers the basic theory for music composition, but would it pass a Turing Test for ‘good music’? This is a test where the intelligence of an AI is measured by how convincing it appears to a human observer. Normal meatbags with squishy analog brains have an emotional relationship to music, and that’s the part where the computers fall down. While Prof. David Cope’s AI has been able to think up some decent classical pieces, this is a genre that is highly formal and rule-driven, rigging the odds in favour of the robot. Similar attempts in rock and pop formats have struck my ears as stilted and dinky.

Could one train an AI to know, like Billy Corgan did, that if you detune the guitar solo a quarter-step in ‘Behold! The Night Mare’ it will not only be shockingly dissonant but also achingly gorgeous? I guess you could in theory — sit the AI down, have it take notes on the musical features of literally hundreds of thousands of songs, how humans have rated those features, aggregate it all into a big database and then generate new musical output based off it.
Computationally, it’s staggering the think about but comes fairly easily to humans. Hubert Dreyfus has argued convincingly since the 60s that the human mind is not wholly or fundamentally computational anyway. AIs will always miss the key component of intuition that drives musical decision-making.

Could we perform some Blade Runner–level memory implanting to make a computer write and perform something like Nina Simone’s ‘My Man’s Gone Now’? The research indicates that we will be able to teach AIs many things but that good ol’ tangle of romantic projection, lusty desire and fear of death that fuels so much good music is going to be forever out their grasp. Which is exactly why Roy Batty was so homicidally pissed.

Those pesky emotions will always pose a problem for our budding AI composers. And while I’m 100% positive that massive leaps are going to get made in the problem-solving aspect of artificial music composition, I wonder if the mere accumulation of huge reams of musical ‘knowledge’ will ever be enough to create a Turing-certified album. The point of music has never been perfection anyway. That’s the reason why we don’t all sit around enjoying Chick Corea Elektrik Band, unless we have a sudden craving for one of the most unintentionally awkward and hilarious video clips of the 1980s.

We’re definitely heading towards a world of increasing automation, and our expectations about what the AI future will look like can’t be accurately determined by our current knowledge. But I really doubt Skynet will be able to write us a five-star album.

Cheers,
Alex

If you’d like an answer to any of life’s great mysteries, get at Alex on advice@lifeisnoise.com.