Here’s a strange thing to sit with. A machine trained on other people’s music can now hand you a finished song in thirty seconds, and nobody, including the company that built the machine, can tell you for certain that it’s yours. The AI music copyright lawsuits now moving through US courts are what happens when that uncertainty finally hits a courtroom.
On one side, Anthropic stands accused of feeding pirated lyrics into its models. On the other, Suno and Udio have started paying the labels instead of fighting them. Two very different strategies. Same underlying question: what does “ownership” even mean when a song is co-authored with a model trained on someone else’s catalog?
The answer isn’t being written by philosophers. It’s written deal by deal, filing by filing, before any court doctrine settles. And the person with the most to lose is the one making the music.
What Anthropic is accused of: piracy at training time
Start with the messy side of the ledger.
In January 2026, Universal Music Group, Concord, and ABKCO filed a second, larger suit against Anthropic. It alleges infringement of more than 20,000 songs and seeks over $3 billion in statutory damages, one of the largest non-class-action copyright cases in US history. That sits on top of the original 2023 case over 499 works, still active as Concord Music Group v. Anthropic.
The specifics are ugly. The complaint says co-founder Benjamin Mann personally torrented works, and that CEO Dario Amodei discussed and authorized the conduct, with pirated files pulled from shadow libraries like Library Genesis and the Pirate Library Mirror. Anthropic doesn’t fully deny the lyrics were there. It admits at least one Claude model trained on a dataset containing lyrics to at least 100 of the publishers’ works.
The company added guardrails after the litigation started. Plaintiffs say they leak. The filings claim outputs still reproduce copyrighted lyrics and that the restrictions can be jailbroken.
The line the courts actually care about
Here’s the nuance most headlines skip. Training an AI on copyrighted work isn’t automatically illegal. The fight is about how the data was obtained.
We saw the template in Bartz v. Anthropic, a books case, not a music one. In June 2025, Judge William Alsup ruled that training on legally acquired books was “exceedingly transformative” fair use. Copying pirated shadow-library files was not, because it displaced demand for the works copy for copy. Anthropic later settled that case for $1.5 billion, roughly $3,000 per work, and agreed to destroy the pirated copies.
Now read the music case again. Same defendant. Same shadow libraries. Same transformative-versus-piracy line. In March 2026, the publishers asked the court for partial summary judgment, arguing the evidence against fair use is “overwhelming.” No public ruling had landed as of this writing, so treat any specific verdict date as anticipated, not confirmed.
This is the same authorship tension we’ve traced with AI video generation. It also echoes older arguments about whether large language models really understand the material they train on, or just recombine it.
Suno and Udio chose the opposite path: license at generation time
While Anthropic litigates, Suno and Udio decided to pay.
In December 2025, Warner Music Group and Suno struck a “first-of-its-kind” deal that also settled prior litigation. Suno agreed to eliminate its unlicensed models and replace them with licensed alternatives during 2026. Warner artists and songwriters got opt-in control over whether their voice, name, likeness, and compositions show up in AI output. Suno even acquired Songkick from Warner as part of the package.
Udio went further into the walled garden. UMG settled its suit in October 2025, and the two are building a jointly operated licensed platform slated for the first half of 2026.
Inside that garden, the rules bite. Outputs can’t be exported, downloaded, or uploaded to Spotify, Apple Music, or YouTube. Every song is fingerprinted and filtered, and UMG artists who opt in earn a per-output royalty scaled to how closely a track resembles the training data. It behaves like a SaaS product, not a distributor.
The momentum is real. Warner signed its own Udio-style deal in November 2025, the indie coalition Merlin followed in December, and Kobalt in January 2026. Consensus is not universal, though. Sony Music was still litigating against both Suno and Udio as of May 2026.
AI music copyright lawsuits: two roads, one destination
Step back and the shape gets clear.
Anthropic took the pirate-first road: train on whatever you can grab, then defend it in court. Suno and Udio took the license-first road: pay the rights holders, then wrap the output in restrictions. Different means, the same end. AI companies are converting an existential copyright threat into a manageable business cost.
Neither road is obviously “the ethical one,” and that matters.
Litigation drags the whole question into public view, with named executives and torrent logs. The licensing deals feel cleaner. But they quietly hand the labels a veto over what gets made, and they shrink what the creator walks away with. Sony’s holdout hints at a possible third road: neither settle nor scrape, just keep fighting until the doctrine hardens.
What it actually means for the person making the music
This is where mindoxai always lands: not on the companies, on you.
The licensing pivot is sold as empowerment. Look closer. Under Suno’s new licensed models, the free tier can’t download songs at all. You can play and share, but not save the file. Paid tiers get capped monthly downloads with pay-per-extra pricing. The music you “made” now sits behind a meter.
Then there’s Suno’s Spark incubator. It markets “full creative control and 100% of your commercial rights.” The actual terms tell a different story: mandatory exclusive use of Suno, broad promotional rights, and a “Good Vibes Only” non-disparagement clause where criticism can count as a material breach and get you removed. CEO Mikey Shulman didn’t help the mood when he said the traditional process of learning music is “not something that most people actually enjoy.”
The deepest problem isn’t the fine print. It’s authorship itself. Suno concedes it can’t guarantee copyright vests in its output, because the US Copyright Office still treats human authorship as the baseline. Type a prompt, get a song, and you may hold something legally hollow: a track you can’t clearly own and can’t freely move.
We’ve circled this drain before, asking whether AI creations even need new ownership rights and how AI is already reshaping creativity and choice. Music is just where the money finally forced the question.
Can training ever be “fair use” if the output copies the input?
Here’s the tension no settlement resolves.
Fair use rewards transformation. A model that learns from a song and produces something genuinely new has a real argument. But if the output can regurgitate the input almost verbatim, the exact thing the publishers say Anthropic’s guardrails fail to stop, the “transformative” defense gets thin fast. You can’t call it new if it hands back the original.
That’s why these cases reach past music. They test whether the whole build-first, license-later model of AI development survives contact with copyright law. The same regulatory pressure is reshaping AI governance broadly, and music is the loudest early test case.
So before you generate your next track, ask the uncomfortable question the industry hasn’t answered. If a model built on borrowed art decides what you’re allowed to own, in what sense is the song ever really yours? Watch the summary-judgment ruling in the Anthropic case. Whichever way it goes, it sets the price of every AI song that comes after.
