Elvis Suit May Shake Up IP Risks For AI Developers
Elvis may have left the building but, according to a new lawsuit, his songs never left Anthropic's servers.
Filed in the U.S. District Court for the Northern District of California on Oct. 23, ABG EPE IP LLC v. Anthropic PBC was brought by the owners of 274 musical compositions recorded by Elvis Presley against Anthropic, the developer of the Claude family of large language models, for copyright infringement. The works include some of the most recognizable songs in American music: "All Shook Up," "Can't Help Falling in Love," "Don't Be Cruel," "A Little Less Conversation" and "Return To Sender."
With copyright lawsuits against artificial intelligence developers now commonplace, why does this one matter? In short, because it is one of the most high-profile actions brought on behalf of a single artist's catalog, and because it reveals, with unusual clarity, where copyright owners believe AI developers are most vulnerable.
It is in this vulnerability that many of the underlying questions remain. What can be used to train AI? When can it be used? In what manner can it be used? Courts have shown some willingness to treat training as fair use. But the plaintiffs in this case seem to be betting that the way Anthropic gathered, cleaned and stored the lyrics will not hold up as well.
The Case at a Glance
The plaintiffs include the owners of much of Presley's intellectual property, joined by Raleigh Music Publishing LLC, Abilene Music LLC, trusts associated with songwriter Luigi Creatore and the Aberbach family, and Jason Simeon Aberbach. The complaint describes them collectively as the Elvis plaintiffs.
The complaint pleads three causes of action: direct copyright infringement, circumvention of protection measures, and removal or alteration of copyright management information.
The plaintiffs also seek a permanent injunction, an accounting of training data, and an order directing impoundment or destruction of unauthorized copies, including copies "retained in training datasets, internal repositories, model weights, and server infrastructure."
That last request deserves attention. A court order reaching model weights would be a remarkable remedy. In a press release, Anthropic has stated that it disagrees with the claims and intends to "defend this case fully."
Three Theories of Liability
The Central Library
The plaintiffs allege that Anthropic collected lyrics from lyrics aggregation sites, web crawl archives such as Common Crawl and third-party datasets such as The Pile, and then placed them in "a permanent, general-purpose central library intended to be retained indefinitely." According to the complaint, that retention "constitutes standalone direct infringement, independent of any downstream training use."
This framing is deliberate. In Bartz v. Anthropic last year, the Northern District of California found that training on lawfully acquired books was "quintessentially transformative" and a fair use, but that building and keeping a library of pirated works was not.[1] Anthropic later agreed to a $1.5 billion settlement in that case, which the court approved on July 20 this year. The Elvis plaintiffs quote Bartz for the proposition that "[t]here is no carveout ... from the Copyright Act for AI companies," and lean on its findings regarding Library Genesis and the Pirate Library Mirror.
In other words, the plaintiffs are not asking the court to decide whether training is fair use in the abstract. They are asking the court to look at where the data came from and how long it was kept.
Stripping the Credits
The plaintiffs license their lyrics to sites such as Musixmatch and LyricFind under licenses that require display of the song title, songwriters and other identifying information, which is the copyright management information, or CMI. They allege that Anthropic's data cleaning pipeline removed that information, and that Anthropic tested several text extraction tools before choosing one "because of its superior ability to remove CMI."
Under the U.S. Court of Appeals for the Ninth Circuit's 2018 decision in Stevens v. CoreLogic, Section 1202 of the Digital Millennium Copyright Act claims require double scienter: The defendant must have removed CMI knowing, or having reasonable grounds to know, that doing so would induce, enable, facilitate or conceal infringement.
Satisfying that standard can be a challenge for plaintiffs. Here, the complaint points to the Northern District of California's ruling in Concord Music Group Inc. v. Anthropic last October, holding similar allegations sufficient at the pleading stage. If the tool selection allegations survive discovery, they could supply the kind of intent evidence that Section 1202 claims usually lack.
Outputs and Guardrails
The plaintiffs also target what Claude produces. They allege that the output filters Anthropic adopted after the first music publisher suit in 2023 "cover only a narrow subset of commercially prominent songs and are readily circumvented by rephrasing the prompt."
Most strikingly, the complaint quotes a Claude response, given after the model was asked to rewrite songs for a modern audience: "These renders keep landing very close to the originals: same structure, same hooks reused with light rewording, and full sets of verses/choruses/bridges. That's edging past 'inspired by' into reproducing the copyrighted song, even when the words are swapped out."
The plaintiffs characterize this as an admission. Whether a model's self-assessment carries any evidentiary weight is an open and interesting question. Still, the allegation illustrates a broader point: Presuit output testing is becoming a standard feature of these complaints.
The plaintiffs further allege that implementing guardrails after being caught, while knowing they were ineffective, demonstrates willfulness. That argument matters because, under the Copyright Act, willfulness is what raises the statutory damages ceiling to $150,000 per work.
What This Case Is Not About
Notably, this is a case about compositions and lyrics. It is not about Elvis' voice, his sound recordings or his likeness.
That is a practical consequence of the technology. Claude generates text, not audio. In Round Hill Music LP v. Anthropic, filed in the Northern District of California on Aug. 17, the Anthropic action was, according to press reports, similarly limited to lyrics and compositions.
There is some irony here. Tennessee's Ensuring Likeness, Voice and Image Security, or ELVIS, Act, enacted in 2024, treats a person's voice as a property right and provides civil and criminal remedies for unauthorized digital replicas. Yet this Elvis-related AI suit does not invoke it. The complaint asserts only federal copyright and Digital Millennium Copyright Act claims. For a text-based model, copyright is simply the better fit. The voice and likeness battles will likely be fought against audio and video generators.
A Crowded Docket
The Elvis suit joins a growing list of music cases against Anthropic. The complaint itself notes that at least six suits relating to musical compositions are pending in the Northern District of California.
These include:
- Concord Music Group v. Anthropic, filed in 2023 by Concord, Universal Music Publishing Group and ABKCO over song lyrics, and a second 2026 action by the same publishers involving more than 20,000 works;
- BMG Rights Management (US) LLC v. Anthropic, filed in March; and
- Suits by Round Hill, filed in August; Sony Music Publishing and Warner Chappell Music, also filed in August; and Reservoir Media, filed in October.
Each new complaint asks substantially the same legal questions with a different catalog attached. Fair use for AI training remains unresolved across the country.
In Kadrey v. Meta Platforms, the Northern District of California last year granted summary judgment to Meta for lack of market harm evidence.
In Thomson Reuters v. Ross Intelligence, the U.S. District Court for the District of Delaware also found last year that copying Westlaw headnotes to train a competing tool was not a fair use, a conclusion the U.S. Court of Appeals for the Third Circuit upheld on interlocutory appeal on Sept. 29. The outcomes remain highly fact-dependent.
Takeaways
The Elvis complaint reads less like a test case and more like a playbook. Each of its theories turns on documents, contracts and engineering choices that exist long before any lawsuit is filed. That makes the lessons practical for everyone in the AI supply chain.
For Rights Holders
Register your works — and do it early.
Statutory damages drive the headline numbers in these cases, and they generally are available only for works registered before the infringement began or within three months of first publication. Without them, a rights holder must prove actual damages or the defendant's profits, which is difficult when the alleged use is buried inside a model. The Elvis plaintiffs list a registration number for every work in Exhibit A to the complaint, which positions them to seek the maximum statutory award per work. Catalog owners should audit registrations now, including for older works and for any lyrics that were never separately registered.
Treat your licenses as litigation assets.
The Section 1201 and Section 1202 claims here exist largely because the plaintiffs' licenses with lyrics platforms require display of titles, songwriter credits and other identifying information, and because those platforms use access controls. Rights holders should confirm that their licensing agreements require licensees to display CMI, maintain technical protection measures and prohibit scraping, and that the licensor has audit rights to verify compliance. Those terms cost little to negotiate and can add independent statutory claims if the content is later taken.
Keep licensing on the table.
Litigation is not the only path to compensation. The complaint itself identifies AI companies and collective rights organizations that already pay rights holders for training data, and that market evidence also strengthens the fourth fair use factor in litigation. A documented willingness to license, along with a clear price, can make it harder for a developer to argue that no market for training licenses exists.
For AI Developers
Know where your data came from.
After Bartz, the source of training data may matter as much as how it is used. Developers should maintain records showing, for each dataset, its origin, the terms under which it was obtained and who approved its use. Third-party compilations such as web-crawl archives deserve particular scrutiny, because a developer may inherit the provenance problems of whoever assembled them.
Rethink indefinite retention.
The plaintiffs frame Anthropic's central library as infringement in its own right, separate from any training. A general purpose repository of copyrighted material kept with no end date is now a primary litigation target. Retention policies that limit how long raw copies are kept, and for what purpose, can narrow this exposure, although any changes must be weighed against litigation hold obligations once a dispute is reasonably anticipated.
Be deliberate about metadata.
Data cleaning is a routine engineering step, but the complaint shows how it can be recast as intentional CMI removal, particularly where internal records suggest a tool was chosen because it stripped credits most effectively. Where feasible, pipelines should preserve authorship and ownership information, or at least store it separately, and engineering teams should document the technical reasons behind tool selection.
Test guardrails the way a plaintiff would.
Plaintiffs now routinely run presuit prompts designed to elicit protected content, including requests to rephrase or modernize songs. Output filters should be tested against deep catalog material and creative work-arounds, not just the most popular titles. Developers should also be prepared to explain when and why guardrails were adopted, since the complaint argues that late or ineffective filters support a finding of willfulness.
For Businesses Using Generative AI
Review before you publish.
AI-generated marketing copy, scripts and other content can reproduce lyrics or other protected expression, sometimes with only light rewording. Businesses should adopt review procedures for AI-generated text intended for public use and train employees not to prompt tools to recreate or update existing works.
Read the indemnity, including the exclusions.
Vendor indemnities for AI output are often narrower than they appear. The complaint notes that Anthropic's indemnity for Claude users excludes users who knew or should have known their queries would produce infringing output. A prompt asking a model to rewrite a famous song could fall squarely within that exclusion. Procurement and legal teams should review indemnity scope, exclusions and caps before deploying generative AI tools, and should align internal use policies with those terms.
Conclusion
The Elvis suit does not introduce a new legal theory. What it introduces is a name that every juror will recognize, attached to a carefully constructed complaint that targets data acquisition, retention, metadata removal and outputs rather than the act of training alone.
The case is at its earliest stage. None of the allegations have been tested, and Anthropic has not yet responded in court. But the strategy is clear. Copyright owners are increasingly betting that even if training survives a fair use challenge, the way the data was gathered and kept will not.
Those developing, licensing or deploying AI should monitor this case and its companions closely as the law continues to develop.
Reprinted with permission from Law360 © 2026 Portfolio Media. Further duplication without permission is prohibited. All rights reserved.

