Round Hill wants Suno's and Anthropic's model weights impounded, not just the training files
Round Hill Music is asking a California court to reach inside the trained models. Both complaints the publisher filed on August 17, one against Suno and Bright Data, one against Anthropic, ask for impoundment or destruction of every unauthorised copy under 17 U.S.C. Section 503, and define those copies to cover “copies retained in training datasets, internal repositories, model weights, and server infrastructure,” Music Business Worldwide reported after reading both filings.
The Suno complaint runs 24 pages, the Anthropic one 26. Each also demands a full accounting of training data, scraping activity, and every dataset touching the works.
What impounding model weights would mean for Suno and Anthropic
Every AI music case so far has argued over the input side: the files, the scraping, the datasets. Round Hill is aiming at the output of the training run.
The contrast the filings invite is Anthropic’s own settlement. Its $1.5 billion book-author settlement in September 2025 committed the company to destroying the files it torrented from Library Genesis and Pirate Library Mirror, along with copies originating from them, and Anthropic certified no commercially released model had trained on those datasets. The files went. The models stayed.
Both complaints lean on Bartz v. Anthropic in the same district, quoting Judge William Alsup’s June 2025 line: “There is no carveout, however, from the Copyright Act for AI companies.” The Suno filing applies that ruling to a company which was never a party to it, alleging Suno “has engaged in the exact same conduct” by retaining unlicensed copies indefinitely. The Anthropic filing goes further and treats each model release as a fresh act of copying, naming Fable 5 and Mythos 5 from June 9, 2026, Sonnet 5 from June 30, and Opus 5 from July 24.
Why Round Hill’s DMCA claim turns on which extraction tool Anthropic picked
The second claim never touches fair use. Round Hill alleges Anthropic tested at least 3 text-extraction programs before ingesting data: Readability, Newspaper and jusText. It says jusText was ruled out because it left copyright notices and ownership details intact, treated internally as “boilerplate” and “useless junk,” and Newspaper was chosen for its ability to strip them.
The same argument lands on Bright Data, whose Scraping Browser renders pages without headers and footers, the exact places copyright management information sits.
Making the deliberate choice to exclude headers and footers is identical to making the deliberate choice to remove CMI.
Copyright management information is the credit block attached to a file: who wrote it, who owns it, what licence applies. Removing it carries its own penalties under DMCA Section 1203, up to $2,500 for each act of circumvention and up to $25,000 for each removal, stacked on top of the per-work damages. Round Hill cites Stevens v. CoreLogic for the standard it has to meet and an October 2025 Concord ruling for the proposition that similar allegations survive at the pleading stage.
Round Hill’s market argument names GEMA’s PLAI dataset, ElevenLabs, Musical AI, Symphonic, Soundverse, GCX/Rightsify and Troveo as proof a licensing market exists. PLAI launched on July 23, under a month before the filings, and GEMA says it is built for tools helping creators make music, with generative AI licensing the subject of its own Suno case. None of Round Hill’s allegations has been tested in court.
Frequently asked questions
What does Round Hill mean by impounding Suno's and Anthropic's model weights?
Model weights are the trained parameters inside a working AI system, the product of the training run rather than the source files fed into it. Round Hill's prayers for relief cite 17 U.S.C. Section 503 and define the copies to be impounded or destroyed as including training datasets, internal repositories, model weights and server infrastructure, which reaches the deployed models themselves.
How is the $75 million claim against Bright Data calculated?
It is 500 works multiplied by the $150,000 statutory maximum for wilful infringement, applied to the contributory infringement count against Bright Data. It is the only fixed arithmetic in either complaint, since the roughly $1 billion figure Round Hill cites depends on amending its exhibits to cover thousands more works.
How does Round Hill's demand differ from Anthropic's $1.5 billion settlement with book authors?
The book authors' September 2025 settlement required Anthropic to destroy the original files it torrented from Library Genesis and Pirate Library Mirror plus copies originating from them, and Anthropic certified that no commercially released model was trained on those datasets. Round Hill is asking for the trained models to be covered as well, which no music case has previously sought.
How many works are listed in Round Hill's Exhibit A against Suno and Anthropic?
Exhibit A in each complaint lists 500 musical compositions rather than sound recordings. Round Hill asserts rights in 16,873 recordings and describes the exhibit as a bellwether it intends to amend to cover both compositions and recordings.
About the author
Covering AI music tools, industry shifts, and news for music creators and professionals. Twice-weekly newsletter at aimusicpreneur.com.
