The Music Publishers vs. Anthropic: What the Filings Allege, What One Court Decided, and the Per-Work Arithmetic
The Music Publishers vs. Anthropic: What the Filings Allege, What One Court Decided, and the Per-Work Arithmetic
Updated September 27, 2026. Every figure and date below was read on that date from filed court documents, the U.S. Copyright Office, the Federal Register and Anthropic's own published terms, listed at the end. Each claim is marked as alleged (what a party asserts in a filing) or decided (what a court has ruled). This is a summary of public documents by someone holding no law licence; it is not legal advice.
First correction: the record labels did not sue. The publishers did.
The complaint filed on August 28, 2026 is captioned Sony Music Publishing (US) LLC et al. v. Anthropic PBC et al., case 5:26-cv-09217, in the Northern District of California, San Jose Division. It runs 48 pages and names roughly 35 plaintiff entities — Sony Music Publishing, Warner Chappell Music, the EMI catalogue companies, Jobete, Famous Music, Hipgnosis, Warner-Tamerlane and others.
Every one of them is a music publisher. Publishers hold rights in the underlying musical composition: melody and lyrics. Record labels hold rights in the sound recording — the specific master. Sony Music Entertainment and Warner Music Group are not plaintiffs here. That distinction is not pedantry: it determines what the works-in-suit are, and therefore what gets multiplied when damages are calculated. A composition case counts songs. A recordings case would count masters.
Two individuals are also named as defendants alongside Anthropic PBC: Dario Amodei and Benjamin Mann.
There are three live cases, not one
Anyone following this from headlines alone will have missed two of them.
| Case | Number and court | Filed | Works in suit | Published posture |
|---|---|---|---|---|
| Concord Music Group v. Anthropic ("Concord I") | 5:24-cv-03811, N.D. Cal. (filed in M.D. Tenn.) | Oct 18, 2023; transferred Jun 24, 2024 | 500 compositions | preliminary injunction denied Mar 25, 2025; three counts dismissed with leave Mar 26, 2025; fact discovery closed Nov 12, 2025 |
| Concord Music Group v. Anthropic ("Concord II") | 5:26-cv-00880, N.D. Cal. | Jan 28, 2026 | not stated in the orders read | preservation order Jul 16, 2026; spoliation claims rejected |
| Sony Music Publishing v. Anthropic | 5:26-cv-09217, N.D. Cal. | Aug 28, 2026 | Exhibit A "hundreds or more"; Exhibit B "tens of thousands" | nothing decided |
Concord I is the one with a real paper trail, and two of its rulings matter. On March 25, 2025 the court denied the publishers' preliminary injunction, concluding they had not shown irreparable harm — and expressly did not resolve fair use. The next day it granted in part a motion to dismiss: contributory infringement, vicarious infringement and the DMCA section 1202 claim over copyright management information were dismissed with leave to amend. Direct infringement survived. Both are decided; neither is a verdict.
What is alleged in the new complaint, kept separate from what is proven
The August 2026 complaint pleads infringement on several theories — torrenting from pirate libraries, scraping, destructive scanning of purchased books, training itself, and outputs both verbatim and derivative — plus a claim under 17 U.S.C. 1202 for removal or alteration of copyright management information. It alleges that between two shadow libraries the defendants "illegally torrented over seven million pirated books", and it quotes the earlier record for the figure of 196,640 unauthorised copies from one dataset. It names around 23 songs by title. All of that is allegation. No ruling of any kind has issued in 5:26-cv-09217, and Anthropic has published no response to it.
The one fair-use ruling that exists, and the three-way split it produced
In Bartz v. Anthropic, 3:24-cv-05417, Judge William Alsup issued an order on fair use on June 23, 2025. It is one district court on one record, and it split the question three ways rather than answering it once. Its concluding paragraph grants summary judgment that the training use was fair use; grants that converting purchased print books to digital was fair use for a different reason; and denies summary judgment that the pirated library copies may be treated as training copies. On that last point the order states that every fair-use factor pointed against the defence.
The operative distinction, and the only one that has actually been decided anywhere: training on lawfully obtained copies was held to be fair use; assembling a permanent library out of pirated copies was not excused by it. Anyone summarising this as "courts have said AI training is legal" is quoting half a sentence.
The settlement that put a price on the other half
The pirated-library half never went to trial. It settled, and the final approval order of July 20, 2026 carries the numbers.
- Settlement fund: 1.5 billion dollars, non-reversionary, paid in four instalments of 300, 300, 450 and 450 million dollars.
- Works on the Works List: 482,460. Notice reached 482,374 (99.98%). Claims were filed for 440,490 works, a 91.3% claims rate — up from 54% four weeks earlier.
- Per-work award, in the court's own words: "about 3,000 dollars, less costs and fees. That is four times the statutory minimum."
- Attorneys' fees: 101,561,111 dollars, about 6.8% of the fund. Expenses 2,635,197.46 plus an 18,220,000 cost reserve. Service awards 15,000 dollars each to three class representatives.
- Anthropic must destroy the original torrented files within 30 days of final judgment. First payments are expected on or before November 15, 2026.
Two derived figures, mine rather than the court's: 1.5 billion divided by 482,460 works is about 3,109 dollars per listed work, which is where the court's "about 3,000" comes from. Net of the fees, expenses, reserve and service awards above, and if the remainder were divided only among the 440,490 claimed works, it comes to roughly 3,127 dollars each — the Plan of Allocation governs the actual distribution and I have not read it, so treat that as arithmetic, not as an entitlement.
The per-work arithmetic, done from the statute
This is why work counts are the only number in a copyright case that really matters. Under 17 U.S.C. 504(c), statutory damages are elected per work, not per act of copying: not less than 750 dollars and not more than 30,000 dollars per work; up to 150,000 dollars if infringement is found willful; as little as 200 dollars if the infringer proves innocence.
| Works | At the 750 floor | At the 30,000 ceiling | At the 150,000 willful ceiling |
|---|---|---|---|
| 500 (Concord I) | 375,000 | 15,000,000 | 75,000,000 |
| 10,000 (one reading of "tens of thousands") | 7,500,000 | 300,000,000 | 1,500,000,000 |
| 482,460 (the Bartz Works List) | 361,845,000 | 14,473,800,000 | 72,369,000,000 |
All dollars; the multiplication is mine, the per-work figures are the statute's. Read the middle row: the same ten thousand works span 7.5 million to 1.5 billion dollars depending on nothing but which subsection a court applies. And the bottom row shows why the 1.5 billion settlement was approved as generous — it sits at roughly 4.1 times the statutory floor for that work count, which is exactly the comparison the court drew.
Registration is the gate, and this is the part that applies to your own work
Here is where a case about somebody else's catalogue becomes a fifty-dollar decision about yours. 17 U.S.C. 411(a) bars filing a civil infringement action for a US work until registration or preregistration has been made. 17 U.S.C. 412 goes further: no statutory damages and no attorney's fees for infringement that began before the registration's effective date — unless registration is made within three months after first publication.
That is not theory. The certified class in Bartz was defined around exactly those conditions: books with an ISBN or ASIN, registered with the Copyright Office within five years of publication, and registered before being downloaded or within three months of publication. Works that missed the window were not in the class. They got nothing from the 1.5 billion dollars.
The Copyright Office's published fees: 45 dollars for a single application (one work, one author, same claimant, not made for hire), 65 dollars for the standard electronic application, 125 dollars on paper, 65 dollars for a group music-album registration. The Office publishes no numbered step list — its FAQ describes three elements sent together: a completed application, a non-returnable deposit copy, and the fee.
So the real-use conversion is blunt. Sixty-five dollars and a three-month calendar reminder is the entire difference between being able to elect 750-to-150,000 dollars per work and being left to prove actual damages. Against a 3,109-dollar-per-work outcome in the one case that paid out, that is the cheapest paperwork in this entire article.
Tip: Section 412 turns on dates, which means the useful habit is keeping a dated, unaltered copy of every finished piece of work from the day it is published. An external hard drive kept offline is the least glamorous and most useful part of that. (These are Amazon Associate links — we may earn a small commission on qualifying purchases.)
What the Copyright Office has published about AI-assisted work
The Office's study comes in three parts: Part 1, Digital Replicas, released July 31, 2024; Part 2, Copyrightability, released January 29, 2025; and Part 3, Generative AI Training, released May 9, 2025 as a pre-publication version whose preface states no substantive changes are expected.
Part 2 states the rules that bind an ordinary creator. Prompting alone is not enough: "prompts do not alone provide sufficient control." Using AI as a tool does not forfeit protection: "the use of AI tools to assist rather than stand in for human creativity does not affect the availability of copyright protection for the output." And a human keeps copyright in "their works of authorship that are perceptible in AI-generated outputs, as well as the creative selection, coordination, or arrangement of material in the outputs, or creative modifications of the outputs."
The registration paperwork follows from that. Under the March 16, 2023 Federal Register guidance (88 FR 16190), applicants have "a duty to disclose the inclusion of AI-generated content", must describe the human contribution in the Author Created field, and must explicitly exclude more-than-de-minimis AI content in the Limitation of Claim field. Anyone who already filed without doing so "should take steps to correct their information so that the registration remains effective."
What the vendor's own terms say, since that is the other document you are already bound by
Anthropic's Commercial Terms, effective June 17, 2025, state that the customer "retains all rights to its Inputs" and "owns its Outputs", and that Anthropic assigns its "right, title and interest (if any)" in outputs — that parenthetical is doing real work, because the Copyright Office's position is that there may be nothing to assign. The same section states "Anthropic may not train models on Customer Content from Services" (commercial terms only; consumer terms are a separate document). Section K.1 commits Anthropic to defend paying customers against third-party intellectual-property claims arising from their paid use, with K.3 excluding customer modifications and patent practice.
The Usage Policy, effective September 15, 2025, separately prohibits infringing third-party IP, prohibits submitting AI-assisted work "without proper permission or attribution", and prohibits using inputs or outputs to train another model without authorisation. The obligation that trips people up is the attribution one, and it is a policy term, not a legal one.
What discovery in these cases is actually made of
One number from Concord II is worth sitting with. On July 16, 2026 the court ordered Anthropic to preserve a deduplicated random sample of 500 million records — 250 million from before the January 28, 2026 filing and 250 million from the two months after — plus every record containing the terms song, lyric, verse or chorus. In Concord I, the sampling protocol was set at 95% confidence with an 11.3% margin of error over 5 million prompt-output pairs, half pre-suit and half post-suit. The publishers' complaint had rested on 4,659 such pairs.
Reviewed at one record per second without pause, 500 million records is about 15.8 years of work. That arithmetic is mine, and it is the whole reason statistical sampling exists in these cases — and the reason a margin of error of 11.3% ends up baked into a damages figure.
Where the government's own pages disagree
Three gaps, unresolved and worth knowing before you cite a date. The Copyright Office describes its Notice of Inquiry as published August 20, 2023; the Federal Register shows August 30, 2023 (88 FR 59942) — a ten-day gap between two government pages, with the Federal Register being the document of record. The comment deadlines differ too: the Federal Register gives October 18 and November 15, 2023, while the Office's page gives October 30 and December 6 — gaps of 12 and 21 days, probably an extension whose citation I could not locate. And in Bartz, the noticed final-approval hearing was April 23, 2026 while the approval order is signed July 20, 2026, 88 days later, with no document explaining the interval.
What is not in the record
The Sony/Warner complaint states no total work count anywhere in the paragraphs available — only "hundreds or more" in one exhibit and "tens of thousands" in another. There is no answer, no scheduling order and no trial date published in that case. Part 3 of the Copyright Office report contains the fair-use analysis, and its conclusions are not summarised here because I could not read those pages rather than because they do not exist. And nothing in any of this tells you how a different court, on a different record, would rule on the same question. The only decided fair-use holding is one district judge's, and it split three ways.
Four checks that follow from the documents
- Register finished work within three months of publishing it. 65 dollars, electronically. Section 412 is a calendar rule, and calendars are cheap to obey.
- Disclose AI assistance on the application. The Author Created field and the Limitation of Claim field, per the 2023 guidance. A registration that hides it may not stay effective.
- Read the terms you are actually under, not the ones you read about. Commercial and consumer terms are separate documents with different training provisions.
- Attribute AI-assisted work. The Usage Policy requires permission or attribution regardless of what copyright law would say.
None of this depends on how these three cases end. The pattern already visible in the record — that the lawful-acquisition question and the training question are separate, and that registration dates decide who is even in the room — is the durable part. It is the same discipline of reading the primary document rather than the summary that made an earlier court story on this site look different once the filings were opened, and that shows up again in what the federal employment projections actually say about entry-level work.
Documents every figure above was read from
- Bartz v. Anthropic, order on fair use, June 23 2025: govinfo.gov
- Bartz settlement documents, including the order granting final approval of July 20 2026 and the claims report: anthropiccopyrightsettlement.com
- Concord I orders (preliminary injunction denied; motion to dismiss granted in part; discovery and sampling orders): govinfo.gov
- Concord II preservation order, July 16 2026: govinfo.gov
- 17 U.S.C. 504(c) statutory damages, and 411(a) and 412 on registration: copyright.gov and copyright.gov
- Copyright and Artificial Intelligence, Parts 1 to 3: copyright.gov. Registration fee schedule: copyright.gov
- Copyright Registration Guidance for works containing AI-generated material, 88 FR 16190, March 16 2023: federalregister.gov
- Anthropic Commercial Terms of Service, effective June 17 2025: anthropic.com. Usage Policy, effective September 15 2025: anthropic.com
This article summarises filed court documents and published agency and company materials as of September 27, 2026. It is written by no one holding a law licence and is not legal advice for any situation. Allegations in a complaint are not findings. Registration requirements, fees and statutory figures carry their own effective dates; confirm any of them with the Copyright Office or a qualified attorney before acting on them.
Comments
Post a Comment