Key takeaways
- Reservoir Media filed separate suits on 21 October 2026 in the Northern District of California against Suno and against Anthropic, which reports say include DMCA claims.
- These are allegations, Anthropic disputes the claims, and no court has ruled on any issue in these cases.
- Earlier cases suggest that how data was obtained and whether output harms a market carry great weight.
- Startups should audit data by source, avoid circumvention, preserve metadata, add output controls and consider licensing.
On 21 October 2026, music publisher Reservoir Media filed two lawsuits in the Northern District of California: one against AI music company Suno over 1,050 compositions, and one against Anthropic over the lyrics of 853 songs. Reports say that the complaints include claims under the Digital Millennium Copyright Act as well as ordinary copyright infringement.
We have kept the description of the cases to what is alleged and reported, and readers should obtain an independent legal view before relying on any characterisation of the claims. This guide summarises the allegations, what the defendants say, how earlier AI cases have come out, and practical steps for generative AI startups. US law points must be confirmed with US counsel.
What is alleged
According to press reports of the filings, Reservoir Media alleges that Suno used 1,050 of its compositions without authorisation in connection with training or operating its music generation service, and that Anthropic used the lyrics of 853 songs without permission. The complaints reportedly seek damages and other relief. Reported totals for claimed damages differ between sources, so we do not repeat a figure.
The reported claims go beyond straightforward copying. They are said to include claims under the DMCA, namely that the defendants circumvented access controls and removed or altered copyright management information, which is metadata such as titles, authorship and licensing information attached to a work. These are allegations in a complaint, not findings.
Reservoir Media’s suits are reported to be separate from other actions that music publishers have brought. Another publisher, Round Hill, has brought its own claims, and other music publishers have previously sued over lyrics. The details and status of those cases should be checked directly, as we did not verify them for this guide.
What the defendants say and what is not yet known
Reports say that Anthropic disputes the claims and intends to defend the case. We did not find a detailed response from Suno in the sources we reviewed. At this stage, no court has decided any issue in these two cases, and pleadings often change after motions to dismiss.
Questions the courts will have to address include whether training on lyrics or compositions is fair use, whether outputs reproduce protected material, whether the defendants obtained the works lawfully, and whether the DMCA theories can be sustained on the facts. The answers may differ between a music generation service that produces audio and a text model whose outputs may or may not reproduce lyrics.
It is also unknown how quickly the cases will move. Complaints in federal court typically take months to reach a ruling on a motion to dismiss and longer to reach summary judgment. A startup that is waiting for a clear rule from these cases will wait a long time.
A further point for startups: pleadings in these cases are public. The complaints, and the defendants’ answers or motions when they are filed, will show which legal theories each side thinks are strongest, how plaintiffs identify the works allegedly used, and how defendants describe their training and output systems. Reading the dockets directly will teach more than any summary, including this one, and is a reasonable use of an hour for any company in the space.
It is also worth distinguishing the two defendants’ products. A music generation service produces audio, and plaintiffs may argue that its outputs resemble recordings or compositions. A text model is asked about lyrics in conversation, and the question is whether it reproduces them and how it is trained. The same label, generative AI, covers technically and legally different products, and results in one case should not be assumed to carry over to the other.
How earlier AI cases came out
Earlier decisions give a mixed picture. In Bartz v. Anthropic, the court held that training on lawfully acquired books was fair use, while the claims about a library of pirated books went forward and ended in a class settlement of about 1.5 billion dollars, with final approval reported in July 2026. In Kadrey v. Meta, reports say the court ruled for Meta on the fair use claim on the record before it, and the case continued on other issues. In Thomson Reuters v. ROSS, the Third Circuit affirmed in September 2026 that using headnotes to train a non-generative tool was not fair use.
| Case | Reported outcome | Lesson |
|---|---|---|
| Bartz v. Anthropic | Training on lawfully acquired books held fair use; pirated library claims ended in a class settlement of about 1.5 billion dollars, final approval reported July 2026 | How data was obtained carries great weight |
| Kadrey v. Meta | Court ruled for Meta on the fair use claim on that record; the case continued on other issues | Results depend on the evidence of market harm |
| Thomson Reuters v. ROSS | Third Circuit held that using headnotes to train a non-generative tool was not fair use | Competing uses and licensing markets matter |
The pattern reported across these cases is that how the data was obtained, and whether the output harms a market for the original, carry great weight. Lawfully acquired material and transformative training have fared better than pirated sources and uses that compete with the original. Lyrics and compositions raise their own question, because songs are short, highly valuable and easy to reproduce in an output.
Because every case turns on its facts, and several are on appeal or unresolved, a startup should not treat any one ruling as a safe harbour. The sensible approach is to look at the factors courts have stressed and to apply them to your own data and product.
Why the DMCA claims matter to startups
Section 1201 of the DMCA bars circumventing technological measures that control access to copyrighted works, and Section 1202 bars removing or altering copyright management information in certain circumstances. These claims do not depend on proving that a use was not fair use, which is why plaintiffs add them. Statutory damages for them can be significant, and courts differ on how far they apply to AI training.
For a startup, the exposure comes from how data is collected. Scraping behind a login or paywall, bypassing rate limits or digital rights management, or stripping metadata from files during preprocessing can each be argued to engage these provisions. Cleaning data is normal engineering, but removing author and licence information should be a considered decision with a record of why.
Where a product outputs lyrics or melodies, there is also the ordinary risk of output that reproduces protected text or music. Output filters and logging are part of the answer, but they do not resolve claims about how the model was trained.
What generative AI startups should do, quick answers
Audit your training data by source. Separate licensed, public domain or openly licensed, and unlicensed material, and record how each was obtained. Stop using sources that involved circumvention or paywalls. Preserve copyright management information unless you have a documented reason not to. Add output controls for lyrics, melodies and other short, valuable works, and test them. Consider licences from publishers and collecting societies, since the market for these is developing. Review your terms of use and indemnities so that customers know what you promise about outputs. Check insurance for intellectual property claims. If you receive a demand letter, preserve documents, stop deletion routines that could destroy evidence, and take advice before replying. If you are named in a suit, deadlines to respond are short. Founders based in India with US customers should also consider Indian law, where the question of AI training and fair dealing is being tested in ANI v. OpenAI, where the Delhi High Court has so far declined an interim injunction on a prima facie view that favoured fair dealing, with an appeal pending. This is a fast-moving area, so review your approach every quarter.
Finally, remember the commercial side. Customers, distributors and investors read headlines about lawsuits and will ask what your position is. Prepare a short, accurate statement on your training data and output controls, reviewed by counsel, so that sales and communications teams do not improvise. Avoid commenting on pending cases beyond what counsel approves.
Related reading: Thomson Reuters v. Ross and legal tech startups, how US AI companies should structure data deals after Ross, copyright in legal headnotes and AI training, AI copyright in India vs the US, ANI v. OpenAI and AI training as fair dealing, AI-generated content and copyright in India and who owns the AI model, its weights and its datasets; also AI training data provenance audits after Bartz; also AI watermarks and trade mark infringement; also AI-made music and copyright in India and licensing Indian music for UK content.
Quick answers
What did Reservoir Media allege against Suno and Anthropic?
Press reports say it sued Suno over 1,050 compositions and Anthropic over the lyrics of 853 songs in the Northern District of California on 21 October 2026, including claims under the DMCA. These are allegations, and Anthropic disputes them.
Do these suits decide whether AI training is fair use?
No. No court has ruled in them. Earlier cases such as Bartz, Kadrey and Ross reached different results on different facts.
What should an AI startup do about training data now?
Know and record where each dataset came from, avoid circumvention and paywalls, preserve metadata, add output controls, and consider licensing, with advice from US counsel.
Do these lawsuits decide whether AI training is fair use?
No. No court has ruled in them. Earlier cases such as Bartz, Kadrey and Ross reached different results on different facts, and the claims here are allegations in complaints.
Why do DMCA claims matter?
Sections 1201 and 1202 concern circumventing access controls and removing copyright management information. They do not depend on proving a use was not fair use, so plaintiffs add them, and they can engage how data is collected and cleaned.
Final word
Halverton & Co. is an Indian law firm advising US companies, founders, investors and law firms on Indian law, including Indian-law and cross-border advice on AI training data, copyright and technology contracts for AI startups. We practise in Jharkhand, Maharashtra and before the Supreme Court of India, and work alongside US counsel, who advise on US law. Halverton & Co.: Where tech needs law! If you have a question about Indian law, write to us at office@halvertonandco.com, or get in touch.
This article reflects developments reported up to late October 2026. It is for general information only, is not legal advice, and does not create an attorney-client relationship. Halverton & Co. is an Indian law firm and does not advise on US law; US-law points should be confirmed with US counsel.
Sources
Related practice area
Intellectual Property
Trade marks, copyright, patents and designs. Searched, filed, licensed and enforced under Indian law.
View the practice areaComments
No account needed. Comments are reviewed before they appear, so yours will show here once approved. Please do not share confidential details or describe your own legal matter: replies are general discussion, not legal advice, and do not create an advocate-client relationship. For your own situation, write to office@halvertonandco.com.
