Editor’s note: This piece is a guest submission from Rose Qianyi Sun, a PhD student in computer music at UC San Diego.

A little while ago, I gave a talk at a local San Diego bar as part of a series that hosts academic lectures over drinks. I played the audience short clips of melodies — some composed by humans and some generated by AI — and asked them to vote on which was which. A room full of curious, attentive music lovers got it wrong about as often as they got it right. People who were sure they had spotted the AI clips picked human composers. The opposite happened, too.

We are often fascinated by these tests, and music perception researchers have long found pleasure in designing them. In the music department at UC San Diego, I research how people listen to machine-made music, and the more I explore this field, the less interested I am in whether we can distinguish AI from human music. As the audience demonstrated, we mostly cannot. What I have come to think matters is the guess itself. Our lived experience influences our judgment before the music even begins. There is no neutral ear. Listening to any music, human- or machine-made, is partly an act of expectation, and that expectation is now setting the terms for the law and the music industry.

Ever since AI music tools became widely available, music perception researchers have reached a consensus. Researchers from the Missouri University of Science and Technology asked participants to guess whether short excerpts of electronic and classical music were composed by a human or an AI. Every excerpt was composed by a human. They found that listeners more frequently mislabeled the electronic pieces as AI because they expected AI music to sound repetitive and loop-driven, while classical pieces were typically identified correctly as human-made. Participants reported using features like dynamic range, tempo, and rhythmic clarity to inform their judgments, none of which were actually helping with discernment. The ear was hearing its own assumptions. 

In 2023, I conducted a disguised Turing Test using AI-generated and human-composed jazz and folk tunes and found something similar. Listeners confidently named complexity, repetition, and melodic structure as the features driving their guesses, but those features did not consistently track the actual source. It turns out that we not only hear with assumptions, but we also construct post-hoc explanations for our judgments, regardless of whether they were correct.

These assumptions also operate across the senses. Researchers from the North China University of Technology found that the more humanoid a digital musician looked, the higher the listeners’ expectations for the artists’ music were. In contrast, cartoonish AI musicians were more warmly received because they were held to lower standards. 

These studies tell a slightly humbling story about how listening works: What we hear is largely ourselves. We listen with expectations about what classical music sounds like, what AI music sounds like, and what a humanoid robot should be able to do. But those categories themselves have been shifting — human-made music increasingly optimized for function and flawlessness, machine-made music for eliciting emotions. The more the two converge, the less stable our expectations become, and the less reliable our ears are.

Our perception can easily be biased, but that’s not new. AI music is only making it newly visible, and the stakes are beyond the lab. These perceptual biases will shape music markets, music credits, and law. 

Major record labels have spent the last two years in court with leading AI music generators over whether those companies can lawfully train on copyrighted recordings. By late 2025, Universal Music Group had settled with Udio, and Warner Music Group had settled with Suno, and a fair-use ruling in the Sony-Suno case is expected this summer. At the center of these disputes is whether the resulting AI outputs are distinctive enough to count as new works. And that is also a perceptual question: transformative to whom, perceived how, and in what context?

A 2024 study from Utrecht University suggests that when listeners were told a piece was composed by a human, they rated it significantly higher on both artistic and economic value than when the same piece was labeled as AI-generated. Labels move price tags. That finding lands directly on the policy questions now being asked at the federal level: Should AI-generated music disclose its origin? How should it be tagged in metadata? What is considered a “human track” when human and machine labor are increasingly entangled? Whoever decides what gets labeled what, and where in the listening flow that label appears, is making decisions about money and artistic credit that the listener’s ears cannot independently verify.

Back in the bar, my audience was not at all embarrassed when I revealed that they had mixed up AI and human music. They were curious and wanted to know what they had missed. What I ask is that we remain careful, curious listeners. In the last decade, more and more music has been created for utilitarian listening: music for work, for mood, for sleep, for the gym. AI is going to keep pouring into that pipeline, whether we engage with it or not. As musician and producer Brian Eno put it, art is not a thing but a name we give to a certain kind of engagement. Noticing what you expected to hear and asking where that expectation came from is something no algorithm can do for you. Composer Pauline Oliveros called this kind of practice a discipline of attention. The small discipline of noticing may be one of the few things in the pipeline that cannot be automated. Both the word listening itself and what it refers to as a practice are being redefined by the conditions of the moment. The baseline mode of meaningful listening should be an active, deliberate process.