An AI labeled static as blues, showing it uses shortcuts instead of listening

The blues started out sounding like the blues. A researcher who studies how computers hear music had recorded a basic track with his 11-year-old son, William, on drums, plus his own apologetically rubbish guitar solo on top.

Then he pressed play on what the computer insisted was the same thing. Out came a thin hiss with louder patches in roughly the same places as the original, following vaguely the same shape. A listener in the room said it sounded like the noise some electric cars make when they are going slowly.

That hiss was the whole point. Fed only that scraped down sound, a publicly available music classifier still labeled it blues. A 30-second clip like this contains on the order of 240,000 frequency bins, too many to test one by one, so the researcher had hunted for the smallest set of frequencies that alone would be enough to get a blues verdict, using a divide and conquer search that splits the frequencies into random chunks, tests which chunks matter, and then splits those again until almost nothing is left.

He calls that leftover the sufficient signal. It is sufficient because it produces the blues label by itself, but it is not necessary, because other frequency sets might also do the trick. He goes further and isolates a complete explanation, a set that is sufficient on its own and also necessary, meaning that if those frequencies are erased from the original, the computer no longer hears blues. In effect, as the researcher put it, "we can literally extract the blues from the audio."

The complete version sounded strange. The bass was gone and much of the drums was gone. What remained was mostly his guitar. The machine, in other words, had decided that his guitar playing was the blues.

Then came the turn. He played what was left after that guitar heavy part was subtracted, with the bass back and the drums prominent and the guitar pushed down. To his own ear, he said, it was borderline but still kind of the blues. The model confidently called it hip hop, with 83 percent confidence, though no one in the room agreed.

The model was not listening for rhythm or instruments. It was keying on a narrow frequency pattern, one so narrow that noise could trigger it and real drums and bass could not save it.