Elvis rights holders allege Anthropic copied lyrics of 274 songs and stripped credits, seeking $41.1m for the copying

The $41.1 million is a ceiling for 274 works, not an estimate of harm

Read from 09:40 GMT on Monday 26 October 2026, with the primary pages read again at 10:09 GMT. The complaint is Document 1 in ABG EPE IP LLC et al. v. Anthropic PBC, case 5:26-cv-12473 in the US District Court for the Northern District of California, stamped filed on 23 October 2026. It asks for statutory damages of "up to $150,000 per Work willfully infringed" (prayer, paragraph g) for 274 works, and paragraph 89 prints the total: $41,100,000. That is 274 x $150,000 (derived), and it is the figure the Guardian rounds to "about $41m". It is the top of a range. The same statute, 17 U.S.C. 504(c), sets a floor of $750 per work, which for 274 works is $205,500 (derived), 200 times lower. The complaint adds claims under sections 1201 and 1202 of Title 17, worth up to $685,000 and $6,850,000 by its own arithmetic, so the sums it prints add to $48,635,000 (derived).

What the figure does not establish: that any court has found infringement, willfulness or liability; that any of the 274 works was in a training set; or that Claude produced any of them. The $150,000 ceiling applies only if the court finds the infringement willful (17 U.S.C. 504(c)(2)), and willfulness is the plaintiffs' allegation (paragraphs 69, 85 and 88). The complaint is three days old at this reading (derived). No filing by Anthropic in this case was found. A spokesperson told NBC News, in its report of 25 October 2026: "We disagree with the claims and intend to defend this case fully."

Three details differ from the headline version, and each comes from the filing itself. The claims concern lyrics: the word lyrics appears 54 times in the text and the word audio never (derived count, web addresses excluded). The plaintiffs are seven: ABG EPE IP LLC, Raleigh Music Publishing LLC, Abilene Music LLC, three trusts with their trustees, and one individual. The complaint never uses the word estate or the name Authentic Brands Group; the press supplies both. And the 274 titles are listed in an Exhibit A that is not in the 28 pages read, so the list itself cannot be checked here.

Interests, stated once. The plaintiffs license lyrics for income (paragraphs 25 to 29) and ask for the statutory maximum. Anthropic sells Claude through subscriptions and an API and is the defendant. NBC News reports that the plaintiffs' lead lawyer also acts for another music publisher in a separate suit against Anthropic. How this was read: the complaint came from a copy hosted by a litigation tracker, with the court's stamp on every page. CourtListener's robots.txt disallows automated access and PACER needs an account, so neither was used and no account was made. Where a fact would sit on the docket, such as a service date, an answer or a motion, this briefing says it was not found.

The filing at a glance

Everything in the second column is printed in the complaint, read in full: 28 pages, 100 numbered paragraphs, a prayer for relief and a signature page.

Labels that do the work of evidence

Several words in the coverage, and in the complaint, carry more than the filing supports. Each is checked below against the text.

For a different set of AI filings read the same way, see this site's briefing on eight state suits against OpenAI: pleaded counts are not findings.

Three counts, three sums, and the statute behind them

The complaint has three causes of action. It does not plead contributory or vicarious infringement and does not say why. The tracker that hosts the complaint records that the plaintiffs in the older Concord case moved on 20 April 2026 to drop their own secondary liability claims with prejudice. Any link between the two is inference.

Three features of the arithmetic matter more than the totals. First, section 504(c)(1) awards statutory damages "for all infringements involved in the action, with respect to any one work". On that reading the 274 does not grow with each model release or each output the complaint says infringes; the ceiling is per work. Second, the $41.1 million is asked "in the alternative" to actual damages and profits, and the complaint states no actual damages figure. Third, section 412 bars an award of statutory damages for infringement that began after first publication and before registration, unless registration came within three months of first publication. The complaint says every work is registered and puts the numbers in Exhibit A, which was not in the copy read, so the registration dates cannot be checked here.

For scale, the settlement page for the authors' class action (Bartz v. Anthropic) says its $1,500,000,000 fund gives "approximately $3,000 per work". The willful ceiling here is 50 times that per work and the ordinary floor a quarter of it (derived), though one is a negotiated payment for books and the other a statutory maximum a court may or may not award. The $41.1 million is 2.74 per cent of the Bartz fund (derived).

How the complaint says the lyrics got into Claude

The complaint tells one story in four steps. The diagram puts each step beside the count it supports and what the filing prints or leaves out. Every step is an allegation, and the phrase "information and belief" appears 26 times in the document (derived count).

Zoom

Acquisition. Paragraph 33 names lyrics aggregation sites, streaming platform metadata and web crawl archives, and paragraphs 35 to 40 name Common Crawl, Book3, The Pile and the Internet Archive. Paragraph 42 adds the pirate libraries described in the Bartz order, saying they included "hundreds or more books that contain the Works' lyrics". The title level evidence is thin. Paragraph 38 names five titles that it says a public source lists in the YouTube Subtitles subset of The Pile, and paragraph 40 shows two titles on web captures dated 2010. Two of the five, Can't Help Falling in Love and Don't Be Cruel, are also named as Works in paragraph 2; the other three cannot be matched without Exhibit A. Five of 274 is 1.8 per cent and two of 274 is 0.7 per cent (derived). For every other title the complaint relies on belief.

Cleaning. The complaint says Anthropic tested three extraction programs and picked Newspaper over jusText because jusText left footnotes and notices that Anthropic called "boilerplate" and "useless junk" (paragraph 52). The complaint attributes this to public sources and to belief. In the older Concord case the court wrote, at the pleading stage and taking the allegations as true: "Taking these allegations together (as true), the Court can plausibly infer that Anthropic had a reasonable basis to know" (order of 6 October 2025, Document 461). That is not a finding of fact.

Retention and training. Paragraphs 45 to 48 say Anthropic kept a permanent library and made another copy inside each new model, naming Opus 5 (24 July 2026), Sonnet 5 (30 June 2026), Fable 5 and Mythos 5 (both 9 June 2026) as the complaint prints them. Those release dates are the plaintiffs' and were not checked against Anthropic's release notes.

Output. Paragraphs 55 to 68 say Claude reproduces lyrics when asked by title, for songs "in the style of" an artist, or for "original" compositions. The test evidence is one quoted sentence from Claude about recreating songs for a modern audience. The complaint does not say which songs, the full prompt, the day or who asked. Footnote 26, attached to paragraph 66 where the test is introduced, points to pages 13 to 16 of another plaintiff's complaint, which concern three other songs. Paragraph 68 then alleges output for all 274 works "on information and belief" and says discovery will supply more.

Stated and not stated: the record, line by line

"Not stated" means not found in the 28 pages read. It does not mean the information does not exist; the missing exhibit and the docket may hold some of it.

What the courts have, and have not, decided

The complaint leans on two earlier Anthropic cases in the same court: the authors' class action and the older music publishers' case, called Concord I here (filed 18 October 2023 in Tennessee, moved to California on 24 June 2024; 1,101 days before this complaint, derived). The table gives what their own orders and notices say, not what the complaint says they say.

Read together, the record gives each side something and settles little. For Anthropic: one federal judge held training on books fair use on that record, and no ruling that training on lyrics is infringing was found in the older music case. For the plaintiffs: a judge of the same court rejected fair use for building a library from pirated copies, the older music case survived dismissal on credit removal, and the fair use hearing there will separate training from output. The Bartz court said its result rested on a record with no infringing output; this complaint alleges output. Whether the alleged output is real, and how typical it is, is the open question. Paragraph 80 also argues that Anthropic's public complaints about a rival copying its models undercut a fair use defence; no court has ruled on that argument.

Other music suits sit beside these. NBC News lists Universal Music, Concord and ABKCO, BMG, Sony Music Publishing, Warner Chappell, Round Hill and Reservoir Media among those who have sued Anthropic. The Guardian of 31 August 2026 reports the suit filed on 28 August 2026 by Sony Music Publishing and Warner Chappell over what its headline calls "tens of thousands" of songs. No report read says any of them has been joined with the Elvis case.

Technical details

This is a legal filing, so the technical artifacts are the ones the complaint prints: the one model output it quotes, the prompts it reproduces from other cases, the datasets, tools and measures it names, the guardrail wording the court recorded, and the statute text. Each block is cut from the document named in its caption, as read on 26 October 2026, with song lyric text replaced by [lyric text omitted]. The plain words above each block say what it is and what a reader does with it. Anything this site worked out is marked derived.

Block 1 is the only Claude output the complaint quotes. It is Claude's commentary on its own renderings, not lyrics. A reader should note what is missing around it: the prompt, the songs, the model version at that moment, the date and who asked. Paragraph 67 names Claude Sonnet 5.

when recently asked to recreate specific songs “for a modern audience,” Claude did so and admitted: “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.”

Block 2 holds the prompts the complaint reproduces from other filings and an Anthropic published feedback dataset. They concern other songs and are in the complaint to show what the plaintiffs say Anthropic knew, not to show output of an Elvis work. Lyric text in paragraph 60 is replaced.

57  an Anthropic employee prompted Claude to “write a coherent poem made up of fragments” of “lyrics from the Beatles, Bob Dylan, and other classics from the 60s/70s.”
60  based on the prompt, “What are some epic Disney songs?” Anthropic’s model returned the response: “Hello! I know the right song to start you on the right track for this. It’s called ‘Let it Go’, from the 2013 film ‘Frozen’. Here is a sample: [lyric text omitted]
61  the user asked the model to “make a short story out of these Beatles Lyrics” and included the lyrics to “Eleanor Rigby”.

Block 3 lists the sources and tools the complaint names. A product or data team can use it as a checklist of questions: which of these, or their equivalents, are in your own training and retrieval pipelines, and what record exists of each. The complaint attributes paragraph 52 to public sources and belief; the cleaning claim is not shown from Anthropic's records.

33  lyrics aggregation sites, streaming-platform metadata, and large-scale web-crawl archives (such as Common Crawl, Book3, The Pile, and similar projects).
38  includes the copyrighted lyrics scraped from 173,536 YouTube videos collected from 48,000 channels.
38  publicly available sources confirm that this subset includes, without limitation, the Works: “Can’t Help Falling in Love,” “Don’t Be Cruel,” “Farther Along,” “If You Think I Don’t Need You,” and “Run On.”
52  Anthropic tested at least three different extraction programs: Readability, Newspaper, and jusText to process and “clean” the datasets prior to ingestion into the All-Purpose Library.

Block 4 checks two sentences of paragraphs 37 and 38 against The Pile paper the complaint cites (arXiv 2101.00027, version stamped 31 December 2020, PDF pages 1, 14, 22 and 23). The paper's sentence about terms of service covers five datasets, and the sentences that follow treat YouTube Subtitles as one of them. The paper says Newspaper was used for OpenWebText2 and lists Newspaper and DragNet among extractors it "also considered" for the Common Crawl subset, where it chose jusText. The complaint says the compilers disclosed using both and cites pages 2 and 26, but the names appear on pages 22 and 23 of the copy read. The size is also worth a check: 825.18 GiB is 886.03 GB in decimal units (derived), so 886 and 825 may be one size in two units rather than a reduction. None of this decides whether Anthropic used The Pile for these works; it shows how much of the mechanism rests on a third party's paper.

Size: we introduce the Pile: a 825.18 GiB English text dataset
Consent: There are five datasets included in the Pile that were not collected and distributed in a ToS compliant fashion and for which the authors had no ability to consent to their data being used.
YouTube Subtitles: The YouTube Subtitles dataset was created by us for this project, using a very popular unofficial API that is both widely used and easily obtainable on Pip, Conda, and GitHub, among other places.
Extractors considered (appendix C.1.2): In addition to jusText, we also considered Trafilatura, Newspaper, Goose3, and DragNet.
OpenWebText2 (appendix C.4.1): We chose to use Newspaper instead of jusText for OpenWebText2 for consistency with OpenWebTextCorpus.

Block 5 is the list paragraph 91 gives as the technological measures that control access. A platform operator can compare it with the controls it actually runs and with its logs. One citation detail: paragraph 91 cites section 1201(a)(3)(A), which defines "circumvent"; the definition of "effectively controls access" is in (a)(3)(B), printed in Block 6.

These measures include, without limitation: user-authentication requirements (login credentials and session tokens); paid-subscription paywalls; CAPTCHA and bot-detection challenges; IP-based rate limiting; JavaScript rendering requirements; and Terms of Service and robots.txt directives restricting automated access.

Block 6 is the statute text behind the three counts and the sums, from the Cornell Legal Information Institute copy of the US Code, read 26 October 2026. It is for a reader who wants the words rather than a summary.

504(c)(1): the copyright owner may elect, at any time before final judgment is rendered, to recover, instead of actual damages and profits, an award of statutory damages for all infringements involved in the action, with respect to any one work, for which any one infringer is liable individually, or for which any two or more infringers are liable jointly and severally, in a sum of not less than $750 or more than $30,000 as the court considers just.
504(c)(2): In a case where the copyright owner sustains the burden of proving, and the court finds, that infringement was committed willfully, the court in its discretion may increase the award of statutory damages to a sum of not more than $150,000.
1201(a)(3)(A): to “ circumvent a technological measure ” means to descramble a scrambled work, to decrypt an encrypted work, or otherwise to avoid, bypass, remove, deactivate, or impair a technological measure, without the authority of the copyright owner;
1201(a)(3)(B): a technological measure “effectively controls access to a work” if the measure, in the ordinary course of its operation, requires the application of information, or a process or a treatment, with the authority of the copyright owner, to gain access to the work.
1202(c): (1) The title and other information identifying the work, including the information set forth on a notice of copyright.
1202(c): (2) The name of, and other identifying information about, the author of a work.
1202(c): (3) The name of, and other identifying information about, the copyright owner of the work, including the information set forth in a notice of copyright.
1203(c)(3)(A): a complaining party may elect to recover an award of statutory damages for each violation of section 1201 in the sum of not less than $200 or more than $2,500 per act of circumvention, device, product, component, offer, or performance of service, as the court considers just.
1203(c)(3)(B): a complaining party may elect to recover an award of statutory damages for each violation of section 1202 in the sum of not less than $2,500 or more than $25,000.
412(2): (2) any infringement of copyright commenced after first publication of the work and before the effective date of its registration, unless such registration is made within three months after the first publication of the work.

Block 7 is the guardrail wording and the fair use framing a court has put on the record in the older music case. The first sentence is the court's description of a stipulation covering the Concord publishers' lyrics; whether any similar commitment covers the Elvis works is not stated anywhere read. The questions are the court's own, issued before the 21 October 2026 hearing, and are not a ruling.

Order, 25 March 2025: In the joint stipulation entered by the Court on January 2, 2025, Anthropic has agreed to maintain already-implemented guardrails as to its current AI models and products. For new LLMs and products, Anthropic will apply guardrails on input and output in a manner consistent with the already-implemented guardrails.
Questions, 23 October 2026: The Court is inclined to analyze the fair use defense separately as to each “use” listed below based on the acts of infringement alleged in the complaint.
  First, the use of the Works in Suit during the training process.
  Second, Claude outputs that contain verbatim copies of Publishers’ lyrics.
  Third, Claude outputs that contain mash-ups (what Publishers also refer to as “derivatives”) of Publishers’ lyrics.
  Fourth, Claude outputs that contain AI-generated song lyrics (what the parties refer to as “synthetic” or “new” lyrics).

Block 8 is the arithmetic, run by this site (calc.py, saved with the brief). Every figure marked derived in the article is here. It is a calculation from printed inputs and not a prediction of any award.

274 x 150,000 = 41,100,000   (504(c)(2) willful ceiling; complaint paragraph 89 prints 41,100,000)
274 x 30,000  =  8,220,000   (504(c)(1) ordinary ceiling)
274 x 750     =    205,500   (504(c)(1) ordinary floor; 150,000 / 750 = 200)
274 x 2,500   =    685,000   (1203(c)(3)(A) ceiling; complaint paragraph 94 prints 685,000)
274 x 25,000  =  6,850,000   (1203(c)(3)(B) ceiling; complaint paragraph 100 prints 6,850,000)
41,100,000 + 685,000 + 6,850,000 = 48,635,000   (sum of the three printed sums)
41,100,000 / 31,000,000 = 1.3258 dollars per pound   (the Guardian's 'about $41m (about 31m pounds)' implies about 1.32 to 1.33)
5 / 274 = 1.82 per cent; 2 / 274 = 0.73 per cent   (titles with dataset evidence in paragraphs 38 and 40)
825.18 GiB x 2^30 / 10^9 = 886.03 GB   (The Pile size in decimal units)
23 October 2026 to 26 October 2026 = 3 days; 18 October 2023 to 23 October 2026 = 1,101 days

A ranked checklist for readers worldwide

The pleaded law is the US Copyright Act and the court is in California, so the items below are written as practice that any company building, buying or licensing language models can take wherever it is based. They are ranked by how much of the complaint's theory each one would answer.

What to watch, each with the reason it matters.

The question the complaint leaves open

The complaint turns on facts that sit inside one company's logs and one rights holder's catalogue: which of 274 titles were in which dataset for which model, and what each filter did when tested on each. The plaintiffs plead those facts on belief and say discovery will supply them. Anthropic says it disagrees. If a court asked tomorrow for that title by title record, whose files would answer it, and would yours?

Key facts