The seemingly unstoppable growth and reach of generative AI has been hitting at least one hard wall in the form of American copyright law. Copyright holders, which include small artists but also some powerful industries, are looking for ways to enforce their rights to the materials that AI is trained on. When they do, they often seem to have two choices: strike a licensing deal or sue. Those approaches may have more in common than it seems. In the most affected industries — news, entertainment, literature, and visual arts — individuals and organizations have been trying both. In the media, The New York Times and prominent newspapers have taken legal action against OpenAI for copyright infringement — while others, including POLITICO parent Axel Springer, have struck licensing deals. Earlier this month, a group representing the major record labels got behind a single aggressive strategy, filing lawsuits for copyright infringement on a “massive scale” against two AI startups, Suno and Udio, whose technology can churn out original songs with just a little text prompting. The official voice of the industry, the Recording Industry Association of America, harked back to the last big threat to recorded music — the rise of online music streaming. “Winners of the streaming era worked cooperatively with artists and rightsholders to properly license music,” the RIAA said in a statement. “The losers did exactly what Suno and Udio are doing now.” It’s strong rhetoric, given how large and well-funded the AI tide feels right now. Is there a serious chance that major AI players could be losers if they don’t play ball? Matthew Sag, an Emory University law professor and authority on these issues, says that while the recording association’s winners-and-losers talk is mostly rhetorical flourish, there is a kernel of truth to it. If lawsuits are actually a tactic to force the negotiation of licensing deals, the music industry has a strong hand here. “Given the concentration of power in the music industry and the fact that musical content is relatively standardized, licensing might be more realistic for training music generation models than it is for other kinds of models,” he said. The labels were set off because the startups did not want to license the music they trained their models on, with the complaint citing an investor who implied Suno knowingly wanted to make its product without “the constraints” of seeking permission from rightsholders. Udio did not answer a question from POLITICO about whether its model had trained on copyrighted data, instead referring DFD to a company blog that said it is not interested in reproducing that content or artists’ voices. Suno did not respond when asked. The lawsuits aren’t trying to claim the generated tracks sound similar enough to count as copyright infringement. They’re based on the premise that training on copyrighted work requires permission. The association is suing the startups for three remedies: an admission that their services broke copyright rules by using the recordings; a stop to further infringement; and damages for past harms. Moving forward, the suit could force AI music generators to pay if they wanted to train on copyrighted music. While other industries may be closely watching, perhaps even admiring, the record labels’ aggressive stance, it’s unlikely that they will be able to replicate their model for several reasons. The recording industry is structured in a way that doesn’t apply to other businesses, according to legal experts. Music rights are concentrated among relatively few big holders, so the three biggest record labels and plaintiffs in the cases wield outsized influence in deciding whether to cut licensing deals. And they hold considerably more leverage in negotiating the terms. The fragmented and decentralized nature of the media industry, to take a counterexample, would make the same kind of unified action much harder. As MIT Technology Review’s James O'Donnell points out, though AI startups could train on music entirely from the public domain, that is far more limited for music, and military marches and other royalty-free songs are hardly what people want to listen to. The other big difference is that the music industry has fought this war before — and won. During the era of Napster, the world went from listeners paying for compact discs to having as much free music as they could download at their fingertips. “It became this existential threat to the way the music industry makes music,” said IP lawyer Louis Tompros. “They realized they had to both use the legal system to shut down [Napster] … and at the same time, provide some other avenues to get customers what they really wanted.” Authorized streaming became the music industry's response, giving rise to services like Spotify and before that, individual music purchases through platforms like the Apple iTunes Store. Tompros isn’t sure the industry will succeed this time by sticking to its old playbook. AI companies aren’t distributing exact copies of copyrighted materials, the way Napster did. The lawsuits are rife with examples attempting to recreate replicas of existing songs, but the end results are technically new works — which is a space that copyright law tries to carve out an exception for. It’s an illustration of how AI really is a different challenge to contend with, even for powerful and legally sophisticated industries. “I’m not sure it quite works as well here because we're not dealing with copies," Tompros. "We're dealing with something new coming along, and that something new may be an economic threat to the industry, but it may not be the kind of threat that they can so easily solve with copyright law."
| 
