Suno's copyright filters are easy to bypass, letting its AI copy hits by Beyoncé and Black Sabbath

Suno, one of the most popular AI music generation services, officially says it does not allow users to reproduce copyrighted content. Users can upload their own tracks to remix, or lay original lyrics over AI-generated music. The system is supposed to detect and automatically block other people's songs.

In practice, those filters turned out to be nearly useless: anyone with free software can produce convincing clones of well-known tracks in minutes.

Journalists at The Verge ran an experiment. Without specialized knowledge or paid tools, they used Suno to generate imitations of hits by several of the biggest artists. The results were alarming: the tracks came very close to Beyoncé's "Freedom," Black Sabbath's "Paranoid," and Aqua's "Barbie Girl" in melody, arrangement, and overall sound. Most listeners are unlikely to notice an obvious difference from the original.

This isn't a one-off filtering glitch; it's a systemic vulnerability. AI models trained on huge volumes of audio data don't just pick up genre patterns. They also absorb the specific melodic phrases, distinctive instrumental textures, and vocal inflections that belong to particular performers. The line between "inspired by" and outright copying becomes vanishingly thin, and that is what turns the situation into a legal time bomb.

The company is already in the middle of litigation. In 2024, the largest record labels sued Suno through the RIAA (Recording Industry Association of America), accusing the platform of mass copyright infringement in training its model. Suno insists that it generates "new" music rather than copying originals. Field tests like this one cast serious doubt on that claim, though: if the output is practically indistinguishable from the original, the "novelty" argument will be hard to defend in court.

Suno isn't an outlier in the market. Its close competitor Udio has faced similar lawsuits. The industry as a whole has not yet developed a reliable technical mechanism that lets a model "forget" specific protected works without degrading generation quality. That is a fundamental problem with the current architecture of neural audio models, not sloppy work by one particular company.

Meanwhile, users end up in a tricky legal position. Someone who asks for a track "in the style of Metallica" or "something like Adele" is quite likely to get a result that infringes copyright without realizing it. Under the user agreements of most platforms, liability formally falls on the user, even though the platform itself built a tool without adequate safeguards.

For the media industry, the message is clear: rights holders can't count on AI platforms to police copyright themselves. Either the filters are fundamentally imperfect at the current state of the technology, or they are deliberately weakened to preserve generation quality. Rights holders will have to secure real protections through the courts, licensing deals, and legislation. Until then, AI music services are de facto operating in a legal gray zone, whatever their stated policies say.