These days, the AI we most often use learns from things people have created: images, music, text. What if we took away AI’s ability to learn from human music and gave it only a feedback loop between the sounds it makes and the responses of the pleasure centers in our brains?
Where does this idea come from? From what happened with algorithms that play Go. Earlier versions of AlphaGo initially learned from human games. Then came AlphaGo Zero, which wasn’t given those examples. It was given the rules and started playing millions of games against itself. And it went beyond what the versions trained on human games had achieved. It beat the earlier AlphaGo that had defeated Lee Sedol by 100 games to zero, and later surpassed AlphaGo Master, which had beaten the world’s best players. So it wasn’t just better than people. It was better than the algorithms that had learned from people, too. Later, AlphaZero did something similar in chess as well. And now, even top Go players learn from AI. AlphaGo Zero, AlphaZero, research on professional players learning from AI.
So instead of learning how people play—and people do make mistakes—the algorithm started finding out for itself what works. Of course, the algorithm itself also changed, so it wasn’t just a matter of taking away the human examples. But that’s the moment that interests me: we stop showing it how we do things. We give it a chance to find another way.
With music, there isn’t such a straightforward way for AI to test things against itself. In Go, you know who won. But here, what would it mean for one sound to be better than another? Well, let’s imagine a thought experiment.
There’s a person connected to some very advanced EEG machine, or another device that reads brain activity. Especially what’s happening in the reward system, but in other areas too. And we have an AI producing all sorts of sounds and checking how they affect our brain.
Of course, there are chords, there are frequencies, there are the directions the sound comes from. All sorts of parameters that we probably can’t even fully define yet. The AI changes these things and gets feedback. This has a stronger effect, that has a weaker one, and here something completely different is happening.
And that way, the AI starts creating sounds that might eventually turn into something we could call a kind of music. Something that instantly makes us laugh or cry. Not because it learned from us how to write a sad song, but because it tested how particular sounds affect a particular brain.
My first thought is: surely this must be possible. But is it? And could we read those responses well enough? Today’s EEG isn’t simply a device you connect to someone and immediately know how much pleasure they’re feeling. There is, however, research into indirectly estimating activity in parts of the reward system from EEG, with the help of fMRI. So there’s something to work with. A 2023 study.
When I started looking into this, it turned out that people had already tried similar things. Back in 2012, there was an experiment called DarwinTunes. Short sequences of sounds changed and combined, and people rated what they liked. The better-rated versions went on to the next generation. And out of the initial noise, something began to emerge that people preferred listening to. There were no ready-made melodies to imitate, although the creators had set certain things, such as the tempo and tuning system. DarwinTunes.
Well, that’s already part of the idea. Except people were rating it consciously, rather than through readings of their brain responses. And this was 2012. The algorithms available then were much less capable than the AI we have today.
Then, in 2019, researchers built a system that connected EEG with music generation. It read signals associated with emotional states and changed the music accordingly. The person could also try to influence it by bringing on emotions, for example through memories. It was a small pilot study. The 2019 study.
And that’s essentially the loop I’m talking about. Except the generator already had certain musical relationships programmed into it. And again—this was 2019. Today, we have much more sophisticated AI. We could train a model in a completely different, more advanced way. So I think there’s a reason to revisit this. Would that be enough? We don’t know yet.
There’s also a more recent study in which in-ear EEG was used to select existing songs to increase pleasure and bring on musical chills. Interestingly, the researchers couldn’t demonstrate a statistically significant advantage for pleasure-focused selection updated through EEG over pleasure-focused selection without those updates. The C-BMI study.
That gives me some hope because, honestly, these possibilities scare me a little. If someone managed to tailor sounds to a particular brain that closely, wouldn’t that give them a way to control our emotions?
And this starts to remind me of the Voice in Frank Herbert’s Dune. The witches of the Bene Gesserit order could use their voices, adjusting them to a particular person and making that person act because they couldn’t resist.
Well, that’s science fiction. But that’s exactly why I find it reassuring, in a way, that things didn’t work so easily in this study. Perhaps even today’s AI can’t do this yet.
What particularly excites me, though, is a project called MindMelody, described in a 2026 preprint. Here, we already have music generation, EEG readings, and adjustments to the parameters based on the listener’s ongoing responses. So, very close to what I’m thinking about. But again: the system uses a pretrained music generator. We haven’t taken away its knowledge of our music. And for now, it’s a preprint. MindMelody.
So I have two reactions at once. I’m glad someone is trying this, and I’m a little afraid of what might come out of it. Well, we’ll see.
There’s another question, too: wouldn’t this music be tailored to just one person? Would it sound like a jumble of strange noises to the rest of us, while triggering particular, unexpected responses in that one person?
Because even if the AI didn’t know our music, we do. We’ve already been learning it, listening to it. We have our experiences, our memories. Everyone’s musical taste has been shaped a little differently. So, in a way, it’s only natural that the result might be different for each person.
Even in studies of ordinary music, the same passage can be very pleasurable for one person and neutral for another. But I’m curious how far that kind of tailoring could go. A study of responses to individually selected music.
I’m not trying to scare anyone about the future here. I’m asking at what point this might become possible. If we can already adapt algorithms to our responses, and AI is getting better and better at it, then I’m curious what it could do with the human brain and sound.
Especially if it had never heard our music before.