Inc.
AI is transforming workplaces, but government employees are struggling to keep up, with 51% of surveyed trainees having never tried generative AI.
YouTube Just Started Labeling AI Music. Have You Been Misled?
Author: Nick Hobson
Two streaming services spent last week trying to figure out whether their listeners should be told that music was machine made. Deezer said it will soon let listeners filter entirely AI-generated tracks out of search. Last Wednesday, YouTube Music started putting AI labels on songs.
The timing isn’t hard to explain. Deezer says it now receives nearly 90,000 entirely AI-generated tracks a day, more than half of everything delivered to it on peak days. YouTube says its listeners “have told us they want transparency when they’re hearing an AI-generated track.”
If you ship anything made with AI, you face the same decision, and it’s a product decision as much as an ethical one. In experiments with piano music and paintings, telling people a machine made the work changed how they judged it, even when the audio or the image was the same.
There’s an important element of human psychology that’s worth examining.
Same audio, lower ratings
The cleanest test I know of comes from a team led by Alessandro Ansani at the University of Jyväskylä in Finland, published in Empirical Studies of the Arts. They filmed a grand piano that played itself, replaying keystroke data from human performances of classical pieces. In one version, a professional pianist sat at the keys and mimed (people believed the pianist was playing). In the other, the bench was empty and viewers were told an AI was playing.
The audio was identical. That’s an important part of the study.
The participants watched both versions and rated the “AI” performances lower on liking, engagement, emotional valence, and overall quality. The penalty showed up regardless of musical expertise and was larger among people who already felt worse about AI.
Afterward, 79 percent said they had noticed differences, and many described them in detail, from timing that was “too precise” to pedal blurring. But none of it was in the audio. They made it up in their heads. The label shaped what people believed they heard, and they built explanations to match the intuition.
This effect has been found elsewhere, including with paintings. In one study, every image was AI-made, and only the “human-created” or “AI-created” label was randomized. They found that the AI penalty hurt the liking of the art just a little, but much more on how much money people thought a painting was worth.
Two bets about the same customers
YouTube’s help page asks music partners to declare each track as Fully Gen AI, Partly Gen AI or No Gen AI. Auto-Tune counts as No Gen AI, while an AI-generated bassline under live vocals counts as Partly. The label sits on the player page, and tapping it explains AI’s role. YouTube’s bet, in other words, is that informed listeners will make their own call.
Deezer allows for a chosen exit. Its toggle, due before the end of 2026, removes tracks it identifies as entirely AI-generated from search results. The company points to an Ipsos survey it commissioned of 9,000 adults in eight countries, in which 45 percent of streaming users wanted that filter. The same survey found 80 percent supported clear labeling and 97 percent failed to pick out the AI tracks in a three-track blind test.
If listeners can’t hear a difference (and this study at least shows that’s likely), the label is most of what they have to go on, and the piano study shows how far the belief that an AI performed can move a judgment. Because the penalty was smaller among people with warmer views of AI, the cost of disclosure likely depends on who your users are and their views on AI generally.
None of this is an argument for hiding AI use. In fact, 80 percent of the people Deezer surveyed said they want the labels. My prediction is that customers want to know ahead of time, and if you try and hide it, it will come back to bite you.
What changes at scale
At Deezer’s volume, nobody is checking labels by hand, and YouTube relies on partner declarations plus its own detection tools (AI trying to find AI), treating undeclared tracks as unknown. Detection makes mistakes, and the piano experiment shows the AI framing works whether or not it’s accurate, since no AI played a note. A human artist wrongly flagged would likely pay the same penalty, so labeling at volume needs a way for creators to contest it.
We’re seeing an arms race unfold in real time. As generators get better at sounding human, platforms like YouTube lean on detection tools to catch them, and every improvement in detection gives the next generator something to beat.
I don’t expect that to settle anytime soon, as the impulse behind it isn’t going anywhere. We (the humans) want to know who made the thing we’re listening to or looking at, and we use that knowledge to decide what it’s worth, which is what the piano and painting studies captured. The optimist in me thinks that impulse will shape how AI gets used in music and art. The same tools sending Deezer 90,000 tracks a day can help someone with no formal training get a melody out of their head, or show a listener what a performer is doing with timing and dynamics they’d never otherwise notice.
Used that way, AI handles the work around the art, and the person writing the song and the person listening to it stay at the center. The labels YouTube and Deezer launched this week are an early attempt to protect that, and the artists and companies who disclose plainly are the ones listeners will trust when they get to choose.
Credits: TCA, LLC.