If you've been following AI copyright litigation by tracking who settled with whom, you've been watching the wrong scoreboard. Settlements produce licensing deals, not precedent. And right now, the absence of binding precedent is the whole ballgame.
After Warner Music and Universal Music both settled with Suno and Udio in late 2025, complete with licensing deals attached, it looked like the music industry was choosing revenue over principle. Sony Music did not follow. Its federal case against Suno is still pending as of this summer, and it is the last realistic chance to get a binding fair-use ruling on AI music training from a U.S. court. If that settles too, the industry will have extracted money without ever establishing whether training on copyrighted recordings was legal in the first place. Everyone gets paid, nobody gets clarity.
The Fair-Use Question Is Actually Two Questions
The central legal dispute in U.S. courts is whether training AI on copyrighted material constitutes 'transformative' fair use. AI companies argue they are building tools, not copying expression. Rights holders argue that when your model reproduces the style, feel, and commercial value of their work, transformation is a fig leaf. Both arguments have real merit. That is precisely the problem.
The U.S. Copyright Office weighed in with a 108-page report in May 2025, concluding that training on expressive works that compete in existing markets probably exceeds fair use. It is a significant institutional statement. It is also completely non-binding. Courts will decide this independently, and depending on the circuit, they might decide it differently from each other.
Meanwhile, Germany's Munich Regional Court already decided, and it went the other way from where U.S. tech companies are hoping to land. In November 2025, the court ruled that OpenAI's storage of song lyrics inside GPT-4's model parameters, not just its outputs, constituted copyright infringement under German law. That ruling treats the training process itself as the violation. If that logic ever crosses the Atlantic, the fair-use defense doesn't just weaken, it collapses at the foundation.
The Output Side Is No Cleaner
Set aside training entirely. Even what AI produces exists in legal limbo. The U.S. Court of Appeals affirmed in March 2025 that fully AI-generated works cannot be copyrighted. Fine. But neither the court nor the Copyright Office has defined how much human creative input is 'sufficient' to qualify for protection. A producer who uses AI to generate a beat and then edits it. A songwriter who prompts a model and rewrites the bridge. A visual artist who steers a diffusion model through fifty iterations. Which of these clears the bar? Nobody knows. The line is somewhere, undefined, and every creator using AI tools commercially is working on the wrong side of legal uncertainty.
The charitable read on all this is that law always lags technology, and the system will eventually catch up. The less charitable read is that AI companies have a structural incentive to keep the law unresolved for as long as possible, because ambiguity is cheaper than licensing, and settlement is cheaper than losing. Sony Music refusing to settle is, whatever else you think of the record labels, the only thing standing between the public and a future where nobody actually knows what the rules are. That is not a comfortable thing to say about a major label. It happens to be true.
