On December 22, 2025, investigative journalist John Carreyrou — the reporter whose book Bad Blood exposed Theranos — and five fellow authors filed a federal copyright lawsuit that swept in nearly every major name in artificial intelligence. The lead-named defendant is Anthropic, but the complaint reached OpenAI, Google, Meta, xAI, and Perplexity as well. The case is Carreyrou v. Anthropic PBC, No. 3:25-cv-10897, filed in the Northern District of California and assigned to Judge P. Casey Pitts — and with 28 press articles tracked since filing, it is one of the more heavily covered AI-copyright filings on the IP Feed.
It is also a case where the story on the docket has drifted well away from the story in the headlines.
What the lawsuit says
The plaintiffs are working authors and journalists — Carreyrou joined by five others — who allege that the defendant AI companies copied their published books without a license and used them to train large language models. The cause of action is straightforward on paper: copyright infringement under 17 U.S.C. § 501. The theory is the one now familiar across the AI-training docket — that ingesting a copyrighted book into a training corpus is a reproduction the copyright owner never authorized, and that "we learned from it" is not the same as "we were licensed to copy it."
What made this filing newsworthy was less the legal theory than the timing and the ambition. It landed after Anthropic's landmark class-action settlement over training data — the roughly $1.5 billion deal that dominated AI-copyright headlines in 2025 — and much of the early coverage framed Carreyrou's group as authors who looked at that settlement and decided to go their own way. Publishers Weekly's headline said it plainly: "Authors File New Lawsuit Against AI Companies Seeking More Money." Rather than accept a class recovery, these plaintiffs filed direct claims and, per later reporting, attached eight-figure damages figures to them.
The context worth understanding: why "training data" is the whole ballgame
The AI-copyright wave now moving through the federal courts turns on a single unsettled question: is copying a work to train a model an infringement, or is it fair use? Defendants argue that training is a transformative, non-expressive use — the model learns statistical patterns, it does not republish the book. Plaintiffs argue that the copying happens at the input stage, before any "learning," and that where the books came from matters enormously: a licensed corpus is one thing, a library of pirated titles is another.
That input-side focus is why sourcing keeps surfacing in the coverage. One tracked headline — "Anthropic Faces a New $75 Million Lawsuit for Pirating Books to Train Claude AI" — captures the framing plaintiffs want: not an abstract fair-use debate, but an allegation about where the training text was obtained. Whether that framing survives contact with the evidentiary record is exactly what these cases are built to test.
Coverage vs. reality: the press is leaning toward the defendants
Here is where the IP Feed's news meter tells a story the day-one headlines don't.
Across the classified coverage, this case reads defendant-leaning with solid confidence — 15 neutral pieces, 8 that lean defendant, and just 4 that lean plaintiff. That is a notable posture for a suit fronted by a household-name journalist with a sympathetic "creators vs. the machines" narrative.
The shape of the coverage explains the tilt. The filing produced a burst of plaintiff-friendly, personality-driven stories on December 23 — "Theranos Exposer John Carreyrou Leads Authors In Major AI Copyright Lawsuit," "John Carreyrou Exposes a New Fraud: The Theft Behind Your Chatbot." Then the legal trade press took over, and the tone changed. As the procedural rulings landed, Bloomberg Law and Law360 ran headlines like "Anthropic, Nvidia Sway Judge to Split Authors' AI Copyright Suit" and "Anthropic Says Authors' New AI Copyright Claims Fall Flat." The emotional story peaked at filing; the skeptical, docket-driven story is the one that has carried since.
One caution the meter is honest about: it has not logged a merits ruling on this docket. The "defendant lean" here is a read on news sentiment, not on any court decision about infringement or fair use. Those questions remain wide open.
Where the case actually stands: one complaint became many
The most important development is one the celebrity framing tends to skip: the single blockbuster complaint did not stay single.
Faced with a suit naming a half-dozen unrelated AI companies in one action, the defendants moved to sever — and the court agreed to break the case apart, defendant by defendant. The docket shows the split happening in real time:
- The OpenAI defendants were severed into their own case, opened as No. 5:26-cv-2730-PCP on March 27, 2026.
- Meta Platforms was severed next, opened as No. 5:26-cv-03725-PCP on April 29, 2026.
That is the event Bloomberg Law captured with "Anthropic, Nvidia Sway Judge to Split Authors' AI Copyright Suit," and Mealey's with "Judges Sever Copyright Claims Naming Top AI Companies." Severance is a quieter headline than "Theranos whistleblower sues Big AI," but it is arguably the more consequential one: each defendant now defends on its own facts, its own sourcing, and its own timeline, rather than being swept along in a single mega-case. By July 2026, the coverage had moved on to the defense's substantive attack — Law360's "Anthropic Says Authors' New AI Copyright Claims Fall Flat" — signaling the fight had reached the sufficiency-of-the-claims stage.
Why it matters
For any writer, filmmaker, or rights-holder watching the AI-training fight, Carreyrou v. Anthropic is a useful case to track for two reasons.
First, it is a live test of the opt-out instinct. These plaintiffs had a class settlement available and chose individual litigation to seek more. Whether that pays off — or whether a severed, one-defendant-at-a-time grind is slower and harder than a class recovery — is a strategic lesson the whole creative community is about to learn in public.
Second, it is a reminder that coverage volume is not case strength. A famous name and a clean narrative generated a wave of favorable stories on filing day; the docket since has been a series of defense-side procedural wins. That gap — between what the headlines imply and what the record shows — is exactly the thing the IP Feed exists to surface.
We'll keep tracking Carreyrou v. Anthropic and its severed offshoots as the courts reach the fair-use questions at the center of every AI-training suit. Follow the case on the IP Feed to get the docket movement — and the coverage-vs-reality read — as it develops.
The IP Feed tracks intellectual-property litigation as it's filed, pairing the public docket with a news meter that flags where press coverage runs ahead of — or behind — the record. Nothing here is legal advice.