Anthropic Chose Its HTML Parser in 2021. That's Now Count IV.

TL;DR

Thirty-five music publishing entities led by Sony Music Publishing and Warner Chappell sued Anthropic, Dario Amodei, and Benjamin Mann on Friday in the Northern District of California. Case 5:26-cv-09217, 48 pages, four counts, jury demanded, "tens of thousands" of compositions at issue. The part that should make every scraper operator sit up straight is Count IV, which argues that choosing an HTML boilerplate extractor because it removes copyright footers is itself a federal violation, at up to $25,000 an instance.

Who filed, and what they are asking for

The caption runs 35 plaintiff entities. Sony Music Publishing (US) LLC brings its EMI, Colgems, Screen Gems, Famous Music, Jobete and Hipgnosis catalogs. Warner Chappell Music, Inc. brings Unichappell, Warner-Tamerlane, Cotillion, Intersong and Walden. Universal Music Publishing is already in the same courthouse alongside Concord and ABKCO, so the publishing arms of all three majors are now litigating against Claude at once.

The four counts split the defendants in a way worth reading twice:

• Count I, direct infringement by torrenting, against Anthropic, Amodei, and Mann.
• Count II, contributory infringement by torrenting, against Amodei and Mann personally.
• Count III, direct infringement, against Anthropic.
• Count IV, removal or alteration of copyright management information under 17 U.S.C. 1202, against Anthropic.

The prayer for relief asks for up to $150,000 per work willfully infringed under Section 504(c), up to $25,000 per CMI violation, a permanent injunction, destruction of every infringing copy under court supervision, and an accounting of Claude's training data, training methods, and "known capabilities." Hold that last one. It matters more than the money.

Count IV is an argument about your scraping stack

Most of the complaint retreads ground the authors' class action already covered: BitTorrent, shadow libraries, roughly seven million pirated books, the $1.5 billion settlement. Count IV is newer territory, and it is uncomfortably specific about engineering.

Every web-scraping pipeline runs a boilerplate extractor, the library that takes raw HTML and returns article text without the nav bar, the cookie banner, and the footer. The complaint alleges that in May 2021 Mann, Jared Kaplan, and other senior staff weighed three of them: jusText, Readability, and Newspaper. In June 2021, per internal messages produced in the earlier Concord litigation and quoted in this filing, they dropped jusText because it left too much "useless junk," including a copyright owner name and a "(c) 2019" notice in a page footer. Newspaper, the complaint says, was hailed as "a significant improvement" because it reliably removed footers, copyright owner names, and copyright notices.

Section 1202 does not ask whether stripping the footer was the goal. It asks whether CMI was removed intentionally, by someone with reason to know it would conceal infringement. The publishers' move is to collapse those two things into one act: the string identifying the owner got deleted, the lyric that pays the owner got kept, and the complaint says people in the room noticed the difference and picked accordingly. It is the bouncer who waves the band inside and quietly wipes their names off the marquee, then insists he was only tidying up the sign.

Whether that survives a motion to dismiss is genuinely open. Courts have been unfriendly to Section 1202 claims where CMI loss is an incidental byproduct of ordinary text processing, and "we picked the parser with cleaner output" is a plausible answer. But it is now a question a data engineer may have to answer under oath, and the answer lives in a chat log from 2021.

Your published dataset is exhibit material

Paragraph 126 is the one that should ruin an ML researcher's weekend. To argue Anthropic actively trained Claude to emit lyrics, the publishers do not cite a leak or a whistleblower. They cite Anthropic's own open release: the footnote points at the HH-RLHF dataset card on Hugging Face, alongside the appendix of Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback, published April 2022.

From those public transcripts the complaint pulls a request for the first verse of "I am The Walrus," a request to rewrite Eminem lyrics "into Beyonce's style," and an exchange where the model volunteers that it has "an extensive database of songs and their lyrics" before producing "Hallelujah." Then the sting: the complaint alleges the human reviewer chose the output with verbatim lyrics and rejected the one with inaccurate lyrics.

If that reading holds, the preference signal did not merely tolerate memorization. It graded memorization for accuracy. Four years later that click is paragraph 126 of a multibillion-dollar complaint, sourced from a public download link.

Publishing your alignment data is good science. It is also permanent, timestamped, and admissible.

The guardrail defense, per the complaint

Anthropic added output filters for copyrighted text after the earlier suits. The publishers concede that and then argue it changes nothing, because the guardrails are "easily circumventable by simply 're-prompting' Claude." They cite a January 2026 paper, Extracting books from production language models, for the proposition that shipped models still return large spans of copyrighted text.

That is the structural weakness of output filtering as a legal position. A filter is a claim about what your model will not say. A plaintiff only needs one transcript where it did.

The math, and the timing

The complaint puts Anthropic's balance sheet on the record deliberately, citing the company's own announcement of a $65 billion Series H at a $965 billion post-money valuation in May, and a Forbes report that investors expect an October IPO near $2 trillion. The argument is explicit: $1.5 billion was "just the cost of doing business."

Which makes the accounting demand the sharpest instrument in the filing. Weeks out from a reported record IPO, a plaintiff group is asking a federal court to order Anthropic to identify every copyrighted work in its training corpus and disclose how each was collected, copied, processed, and encoded. Statutory damages are a number you can reserve against. A court-supervised training-data inventory is a disclosure event.

Three days earlier, the same parents wrote a check

On August 25, Sony Music Group, Universal Music Group, and Warner Music Group all took equity in Stability AI's $76 million Series B, alongside Electronic Arts, AMD Ventures, and Pacific Alliance Ventures. Three days later the publishing arms of two of those same parents filed this complaint. That is not incoherence, it is a pincer: buy into the model companies that license, sue the ones that scraped. The majors are drawing a line and charging admission on both sides of it.

Caveats

• Everything above is an allegation. Anthropic has not answered, and told TechCrunch it disagrees with the claims and intends to "defend ourselves robustly in court."
• Naming the founders is not new. The January 2026 Concord filing did the same, and Amodei has already moved to drop the direct-infringement claim against him personally in that case.
• The internal quotes are reproduced from filings in the earlier Bartz and Concord dockets. They are quoted in this complaint, not independently obtained by it.
• The Exhibit A and Exhibit B work lists are described as non-exhaustive and exemplary, so "tens of thousands" is a floor the publishers set, not a court finding.

What to do if you train on scraped text

• Treat extractor selection as a documented engineering decision. A benchmark showing you picked the library with the best text quality reads very differently from a chat message celebrating that it removes copyright notices.
• Keep CMI in a sidecar. Strip boilerplate for training if you must, but retain source URL, owner, and notice strings in metadata. Deleting attribution and being unable to reconstruct it is exactly the fact pattern Section 1202 was written for.
• Assume every public artifact is discoverable. Your preference dataset, model card, eval transcripts, and appendix examples are all permanently citable. Review them the way you would review a filing.
• Do not treat output filters as a shield. They reduce harm and they will not stop a plaintiff who owns a re-prompt transcript. Removing the material from the corpus is the only argument with a floor under it.

Key Takeaways

• Sony Music Publishing, Warner Chappell, and 33 affiliated entities sued Anthropic, Dario Amodei, and Benjamin Mann on August 28 in N.D. Cal., case 5:26-cv-09217, over "tens of thousands" of compositions.
• Four counts: direct and contributory infringement by torrenting (the founders named on both), direct infringement, and CMI removal under 17 U.S.C. 1202.
• Count IV rests on an engineering choice: dropping jusText for Newspaper in June 2021 because Newspaper reliably removed copyright footers. Statutory ask is up to $25,000 per violation.
• The complaint sources its "Anthropic trained Claude to emit lyrics" argument from Anthropic's own public HH-RLHF dataset and its April 2022 RLHF paper appendix.
• Publishers concede the output guardrails exist and argue re-prompting defeats them, citing a January 2026 extraction paper.
• The relief sought includes a court-supervised accounting of Claude's training data, filed weeks before a reported October IPO near $2 trillion.

Sources: Complaint, Sony Music Publishing (US) LLC v. Anthropic PBC, No. 5:26-cv-09217 (N.D. Cal. Aug. 28, 2026), Music Business Worldwide, TechCrunch, Engadget, Round Hill v. Anthropic complaint, TechCrunch on the $3B Concord filing, Anthropic Series H announcement, Tech.eu on the Stability AI round