Here are the 30,000 songs Sony is suing Udio’s AI music generator over

Sony’s New Lawsuit Targets 30,000 Songs: What It Means for AI Developers

Sony Music Entertainment has filed a fresh lawsuit against the AI music generator Udio, alleging copyright infringement over more than 30,000 songs — a catalog spanning Elvis Presley’s “Hound Dog” to Beyoncé’s “Say My Name” and Harry Styles’ “As It Was.” Filed in a New York court, the complaint asserts that this list represents “only a small portion” of the works Udio allegedly infringed. For developers building with generative AI, this case crystallizes the highest-stakes question facing the field: what happens when training data can be forensically traced back to specific copyrighted works, and how do you defend against claims that scale to tens of thousands of individual instances?

The lawsuit is not an isolated event. It follows a 2024 action in which Sony, Universal Music Group, and Warner Records jointly sued Udio and the AI music generator Suno. Since then, the legal landscape has fractured: UMG and Warner have settled with Udio and now partner with the company. Sony, however, has chosen to escalate. That divergence creates a volatile environment for any developer working with AI-generated audio — and offers critical lessons for anyone building on top of large generative models.

The Legal Strategy: From a Blocked Amendment to a Fresh Lawsuit

Sony originally attempted to add the 30,000-plus songs to its existing 2024 case. After gaining access to Udio’s training data through discovery, the company used an “audio fingerprinting” technique to identify specific recordings Udio allegedly “copied and ingested into its generative AI models.” But a judge rejected Sony’s motion to amend the complaint, leaving the original suit limited to 333 works. Sony then refiled those 30,000 songs as a brand-new lawsuit — a tactical move that allows it to press separate claims while preserving its original action.

For developers, this procedural detail matters. It shows that copyright holders are willing to litigate in waves, and that a single model’s training corpus can generate multiple lawsuits. The scale is staggering: even if Sony ultimately prevails on only a fraction of the claims, the potential statutory damages — up to $150,000 per work — could reach into the billions. More importantly, the strategy signals that rights holders will use every procedural avenue to expand the scope of their cases, not just the number of works but the legal theories supporting them.

Sony alleges that Udio “admitted that its generative AI models were ‘constructed by showing the program a vast amount of different kinds of sound recordings,’” including from YouTube. That admission, if proven, could be leveraged to argue that Udio knew or should have known its training data contained copyrighted material. Developers should note that any public statement about training data — even a general one — can become evidence in litigation.

Audio Fingerprinting and the Challenge of Proving Infringement

The technical heart of Sony’s case is audio fingerprinting, the same technology used by services like Shazam to identify songs from short clips. By applying this technique to Udio’s training data, Sony claims it was able to match specific recordings — not just song titles or compositions — to copyrighted works. This matters because copyright in sound recordings is distinct from copyright in musical compositions; a lawsuit based on recording fingerprints targets the actual audio data the model ingested, not just the underlying lyrics or melody.

For developers, this establishes a new evidentiary baseline. Early generative AI lawsuits often relied on circumstantial evidence — outputs that sounded similar to known works. Audio fingerprinting represents a more direct method: prove that a specific file ended up in the training set. That raises immediate practical questions:

• Data provenance: If you use a third-party dataset or scrape audio from the web, can you account for every file? Audio fingerprinting can detect matches even if files are renamed, reformatted, or partially altered.
• Training pipeline transparency: The clearer your training data ingestion record, the easier it is to either disprove infringement claims or negotiate a license. Obscuring the origin of data is now a legal liability.
• Model architecture choices: Some techniques, like training on compressed or heavily transformed audio, might evade fingerprinting — but they also degrade output quality. Trade-offs are sharp.

This case suggests that the era of “just scrape and train” is ending for audio as it already has for text and images. Developers building music-generation tools need to invest in data auditing infrastructure now, not after a subpoena arrives.

The Industry Splits: Settlements vs. Litigation

Perhaps the most striking signal for the developer community is the industry’s split response. Universal Music Group and Warner Records have settled with Udio and are now partnering with the company, “as the music industry embraces AI.” Sony, by contrast, is doubling down on litigation. This divergence means there is no single legal standard for AI training on music — it depends on which rights holders control the works and what their business strategy is.

For a startup building an AI music generator, this creates a fragmented licensing landscape. You might reach a deal with one major label, only to face a lawsuit from another over an overlapping set of songs. The Sony case also implies that rights holders will monitor AI services continuously. The company says it could expand the new lawsuit further, meaning even after settling, developers could be hit with fresh claims based on newly discovered training data.

Developers should consider engaging proactively with rights holders on data licensing, but also recognize that some may prefer litigation as a negotiating tactic. As the landscape matures, services that offer clear audit trails and licensing frameworks will have a competitive advantage. For those evaluating the cost of compliance, tools that compare API pricing and provider terms — such as our LLM API pricing reference — can help model the financial implications of using licensed data sources versus scraping.

Financial Stakes: Up to $150,000 Per Work

Sony is asking the court to block Udio from further infringement and to award damages of up to $150,000 per work. With more than 30,000 works at issue, the maximum theoretical damages exceed $4.5 billion — though actual awards are rarely that high. Still, the request underscores the financial risk for AI companies that rely on unlicensed training data. Even a fraction of that sum could bankrupt a startup, and the threat alone can deter investors and partners.

This risk calculus should inform how developers budget for legal and licensing costs. Building an AI music generator involves not just model development but an ongoing operational expense for data compliance. Developers can use a resource like our LLM API cost calculator to estimate the cost of alternative approaches, such as using licensed audio from stock music libraries or partnering with labels — though those costs may be significantly higher than self-training, they remove the existential legal risk.

It is also worth noting that the $150,000 figure applies to each individual sound recording. A single song can have multiple rights holders: the recording label, the publisher, the performing artist, and the songwriter. A developer could face separate claims from each. The Sony case deals only with sound recordings owned by the label, but similar lawsuits over composition rights are likely.

What Developers Should Watch Next

The Sony-Udio case is far from resolved. Key developments to monitor include:

• Discovery rulings: Will Udio be forced to reveal its full training data? If so, further fingerprinting could lead to even more works being added.
• Fair use arguments: Udio may argue that its use of copyrighted recordings for training constitutes fair use, a defense that has fared poorly in some AI copyright cases but remains an open question for music.
• Legislative response: The fractured industry response — some labels settling, others suing — could pressure lawmakers to create clearer rules for AI training on copyrighted audio.
• Technical countermeasures: Expect AI companies to develop obfuscation techniques that make fingerprinting harder, though those may invite additional legal scrutiny.

For any developer building with generative AI, the lesson is clear: your training data is a legal record. The same technology used to identify songs can be used to reconstruct your pipeline. Investing in data provenance, licensing agreements, and transparent documentation is no longer optional — it is core to sustainable product development. The Sony lawsuit puts audio in the spotlight, but the same principles apply across modalities. The community should treat this case as a blueprint for what every sector will face.

Source: The Verge. Details as reported; verify specifics at the source.