A Music Publisher Turned Suno's Licensing Deals Into Evidence

Reservoir Media filed two copyright suits on October 21, one against Suno and one against Anthropic, in the Northern District of California. Together they plead about $337 million.

The Suno complaint is the one the music industry will read closely, and it contains a line that is going to be quoted for a while: Suno's model, the filing says, was "infringe now, make billions by being the first mover, and try to settle on the cheap later."

That's rhetoric. The legal argument underneath it is much more interesting, and it applies well beyond music.

The arithmetic, since it's precise

Reservoir, its UK subsidiary, and Shapiro, Bernstein & Co. claim Suno copied 1,050 compositions to train its models. The named works include Joni Mitchell's "Both Sides Now," Johnny Cash's "Ring of Fire," Madonna's "Papa Don't Preach," Evanescence's "Bring Me to Life" and Miles Davis's "So What."

The money breaks into three counts. Statutory damages at the maximum of $150,000 per willfully infringed work come to $157.5 million. Two Digital Millennium Copyright Act counts follow: $2.625 million for allegedly circumventing access controls, and $26.25 million for allegedly removing copyright management information. Total pleaded: just over $186 million.

The Anthropic complaint, filed the same day, runs the same playbook over 853 works for about $151.4 million.

Both ask for an injunction and for something more consequential than money: the destruction of unauthorized copies, "including copies retained in training datasets, internal repositories, model weights, and server infrastructure."

Read that list again. Model weights are on it. A plaintiff asking a court to destroy weights is asking for the product itself, not a settlement discount.

The clever part: your own deals are the evidence

Most training-data suits fight over fair use in the abstract. Reservoir sidesteps that by pointing at Suno's licensing agreements with Warner Music Group, BMG and Believe.

The argument is elegant. If Suno signed deals to license recordings for training, then Suno has, in the complaint's words, "all but admitted that there is a functioning market for licensing music and sound recordings for training data." And if a market exists, the fourth fair-use factor — the effect on the market for the work — tilts toward the plaintiff before the argument even starts.

You can't simultaneously claim that training on copyrighted work is a transformative use outside the market and pay three major rights holders for the privilege. One of those positions has to give.

This is the argument every AI company with a mixed licensing record should be worried about, because almost all of them have one. Sign a deal with a stock library, then scrape the open web, and you've handed plaintiffs a confession with a receipt attached.

The leaked code is doing a lot of work in this complaint

The factual spine comes from Suno source code that a hacker obtained and 404 Media reported on in July. Suno's response at the time was that the incident mostly involved "outdated source code that is no longer in use."

Reservoir's complaint reads that code as a pipeline description. It alleges Suno bought proxy services from Bright Data to scrape from YouTube Music, Deezer and Genius, and cites specific counts from the workflows: 2,057,486 artist records saved by a YouTube Music job, 6,980,653 "successful song-page processing records" from a Genius job.

It also alleges a detail that matters for the fair-use fight: the code pairs a cover recording with the original recording of the same song when building a training instance. Training on covers as well as originals weakens any argument that the model learned style rather than specific performances, because a cover and its original share a composition and almost nothing else.

Whether that reading holds up is for a court. But it's the kind of specific, checkable allegation that survives a motion to dismiss far more easily than "the model was trained on the internet."

The guardrail section is the best part

Reservoir argues Suno's artist-imitation guardrails are trivially bypassable, and gives examples that are almost funny.

A prompt containing "Reba McEntire" is rejected. A prompt for a female-vocal country song "in the style of Reba McIntyre" — misspelled — is accepted.

With v6, the complaint says, Suno takes prompts that name artists and redirects them "toward descriptive musical characteristics." A prompt for "a Miles Davis like Blues track with horns and a quartet" reportedly produced two instrumentals, both titled "Blue Note Quartet," which Reservoir characterizes as unlicensed derivative works of Davis's music. "So What" is on Reservoir's list.

Every generative platform has a version of this problem, and it isn't limited to music. Blocking a name is a string filter. Blocking a style requires understanding what the style is, which is the thing the model was trained to do in the first place. You can't solve it at the prompt layer, and pretending otherwise is what turns a technical limitation into a litigation exhibit.

Why this matters if you work with images

Swap the nouns and the complaint still reads.

Image models have the same mixed licensing history: deals with stock libraries here, broad web scraping there. Image platforms have the same guardrail theater: block a living artist's name, allow "in the style of [misspelled name]". And image plaintiffs now have the same template — statutory damages per work, DMCA counts on top, and a demand for the weights.

The fair-use citations in this complaint are worth noting for the same reason. It relies on Bartz v. Anthropic, the Northern District of California's June 2025 ruling that maintaining a permanent library of pirated copies wasn't fair use, and on the Third Circuit's September 29 decision in Thomson Reuters v. ROSS Intelligence. Both are precedents that image defendants will face too.

The other thing to watch is the pile-up. Suno is already defending against Universal Music, Round Hill, GEMA in Munich, Denmark's KODA, Canada's SOCAN and a group of independent artists. A defendant fighting eight fronts can't settle cheaply on any one of them, which is exactly the scenario the "settle later" business model didn't plan for.

The necessary caveat

These are allegations in a complaint. No court has found anything. Suno hadn't publicly responded when the filing was reported, and "the code shows X" is a plaintiff's characterization of a document until a judge says otherwise.

What is not in dispute is the shape of the strategy. Rights holders have stopped arguing about whether AI training is good for culture and started building cases around paper trails — licensing deals, leaked repositories, guardrail workarounds. That's a much harder fight to win on vibes.

For anyone building on generative models, the practical lesson is unglamorous and immediate: know what your vendor licensed and what it scraped, because if there's ever a lawsuit, the difference between those two things is the entire case.