A new report from The Atlantic says large collections of music used to train artificial intelligence systems may include unlicensed songs, adding to the growing debate over how generative tools are built and what material they learn from.
The article points to examples in which AI-generated songs echo recognizable elements from well-known recordings. In one case, a performance by Olympic-bound figure skaters featured lyrics that closely resembled lines from the New Radicals' 1998 hit "You Get What You Give," even though the track had been transformed into a different style. The example raised questions about whether the underlying AI model had been trained on the original song and whether it had reproduced parts of it during generation.
The Atlantic's reporting says this kind of copying is not always obvious, but it can surface in ways that sound unmistakably familiar to listeners. The piece describes AI systems as capable of memorizing or regurgitating parts of songs they were trained on, a problem that has become a broader concern in research on generative models.
The article also cites examples involving Suno, one of the better-known AI music generators. According to the report, the service has produced tracks that strongly resemble classics associated with artists including Michael Jackson, Ed Sheeran, Chuck Berry, Bill Haley & His Comets, and B.B. King. The similarities, the report says, have drawn attention to how closely AI-generated music can track existing works even when the output is not an exact copy.
That dynamic has renewed scrutiny of the datasets behind music-generation systems. If copyrighted recordings are included in training sets without permission, rights holders could argue that the technology benefits from material that was never licensed for that purpose. For creators and publishers, the issue goes beyond one song sounding similar to another. It touches on who controls the use of recorded music in machine learning, and what obligations companies have when assembling massive training libraries.
The Atlantic's report does not identify every source used by the systems it references, but it makes clear that the scale of music data involved is large enough to include substantial amounts of copyrighted work. That possibility has become increasingly important as AI music tools move from novelty products to mainstream creative platforms.
The findings arrive at a time when AI companies face mounting pressure from the entertainment industry over how they source training data. Music has become one of the most visible battlegrounds because songs are easy to recognize and hard to disguise when a model reproduces melodic hooks, lyrical phrasing, or arrangement patterns.
The report suggests that as AI music tools improve, the line between inspiration and imitation may become even harder to police. Listeners can sometimes identify a source immediately, but in other cases the resemblance may be more subtle, leaving legal and technical questions unresolved.
For now, the article adds to a widening body of reporting and research indicating that generative AI systems can reflect the copyrighted material they consume. In music, that problem may be especially acute, since even brief echoes of a familiar song can trigger concerns about ownership, licensing, and fair use.
As AI-generated music becomes more common in commercial and creative settings, the questions raised by The Atlantic's reporting are likely to remain central: what data trained these systems, whether that data was licensed, and how much of the original music may still be audible in the output.