Labels v. Suno

What the labels allege

UMG, Sony and — until their late-2025 settlement — Warner sued Suno in June 2024, alleging its text-to-music generator was trained by copying decades of copyrighted recordings at scale. The labels seek statutory damages of up to $150,000 per infringed work. The suit was filed the same day as a companion case against Udio in New York, part of a coordinated RIAA litigation strategy against the leading AI music generators.

By 2024, Suno had emerged as one of the most prominent AI music startups, drawing rapid consumer adoption for its ability to generate full songs, vocals included, from a text prompt — which is precisely what made the scale of alleged copying central to the labels’ case rather than a peripheral detail.

Suno’s answer: an unusual admission

Suno’s August 2024 answer did not dispute the scope of copying the labels alleged. Instead, it stated that its training data included essentially all reasonable-quality music files that could be obtained on the open internet, and pivoted straight to a fair-use defense rather than contesting what was actually used to train the model.

That framing shifted the entire fight onto fair use from the outset — there is little factual dispute left over what Suno trained on; the case turns on whether that scale of copying is lawful.

“Suno’s training data included essentially all reasonable-quality music files that could be obtained on the open internet.”— Suno, Inc. Answer and Affirmative Defenses, No. 1:24-cv-11611 (D. Mass. Aug. 2024)

Discovery by fingerprint: from 560 works to 61,026

Using audio-fingerprinting analysis of Suno’s training data, the labels told the court the training set contained millions of copyrighted recordings. In May 2026, UMG and Sony moved to expand the complaint from roughly 560 specifically identified works to 61,026; public reporting does not show a ruling on that motion yet.

The fingerprinting exercise matters beyond this case: it shows that even a training set described in vague terms — “essentially all” of the open internet’s music — can be forensically matched against specific commercial recordings once litigation reaches discovery, changing the practical calculus for defendants going forward. A dataset’s opacity is a starting assumption, not a durable protection.

Warner’s exit vs. UMG and Sony’s continued fight

Warner Music Group settled and exited the case in November 2025 through a settlement and licensing partnership with Suno; terms were undisclosed, though reporting noted Suno’s acquisition of Songkick from Warner as part of the deal — a rare instance of AI-training litigation converting directly into a commercial licensing relationship.

UMG and Sony continue to litigate. A June 30, 2026 scheduling order reportedly reset dispositive-motion briefing to April 2027, meaning the fair-use showdown many expected in mid-2026 did not happen, and the case’s timeline extended by roughly nine months. That divergence — one major label licensing while two keep litigating — leaves buyers and competitors with two very different reference points for how this kind of dispute can resolve.

Why it matters for training data

This is the leading US test of whether training a music model on scraped commercial recordings is fair use, and it shows how discovery actually works in practice: audio fingerprinting matched training data to specific recordings, so neither a supplier nor a buyer of a training dataset can assume opacity survives litigation.

Warner’s exit shows one path — litigation converting into a paid license. UMG and Sony’s continued fight keeps per-work statutory exposure live at a scale approaching 61,000 works. For anyone holding or sourcing audio training data, the open question is whether provenance records exist before a subpoena or fingerprinting analysis forces the issue — and Suno’s continued fundraising through the litigation, including a reported $400 million round in June 2026, signals that investors are pricing this exposure rather than waiting for it to resolve.