The $1.5 billion Anthropic settlement paid only those rightsholders who could prove ownership of their work; every creator who could not stood outside the largest copyright class action settlement in history. On 20 July 2026, the court granted final approval and closed the case — drawing the sharpest line yet between creators who can prove what they own and creators who cannot.
For most working creators, the catalogue is the pension: the manuscript, the song masters, the photograph archive, the codebase built over years. Almost none of that work carries any record of ownership beyond the creator's own devices, held together by an assumption that authorship will be obvious if anyone ever asks. The Anthropic settlement has now demonstrated, at the scale of half a million works, precisely what that assumption costs.
This post explains what the settlement decided and what it deliberately left open, why its structure excluded most working creators despite the size of the fund, how the same proof problem plays out across the United States, the European Union, and the United Kingdom, and what establishing independent proof of ownership requires.
Anthropic never lost on AI training; it settled on how it acquired its training data. In June 2025, Judge William Alsup of the US District Court for the Northern District of California ruled on summary judgment that training large language models on lawfully acquired books was fair use. What survived that ruling was narrower and more dangerous: the claim that Anthropic had downloaded roughly 500,000 books from two pirate libraries, Library Genesis and Pirate Library Mirror. With a December 2025 trial approaching, Anthropic faced statutory damages under 17 U.S.C. § 504(c) of $750 to $30,000 per infringed work, rising to $150,000 where infringement is willful — across the works list, a range running from roughly $360 million to more than $70 billion. Anthropic agreed to pay $1.5 billion, the largest copyright class action settlement in history.
The final chapter closed this week. On 20 July 2026, Judge Araceli Martínez-Olguín granted final approval in a 23-page order, resolving the remaining objections, awarding class counsel approximately $101.5 million of the $187.5 million requested in fees, and requiring Anthropic to destroy the pirated source files. The settlement is not a verdict on whether AI training is lawful — Judge Alsup had already answered that question in Anthropic's favour. It is a verdict on unlawful sourcing, and a demonstration of what proven ownership is worth when sourcing is challenged. One further limit deserves attention: the class was certified only on the piracy claims, so the fair use ruling on training binds only the three named plaintiffs. For every other rightsholder, the lawfulness of training on their work remains formally undecided — which means the proof question this settlement exposed will be asked again.
The class was defined by proof, not by infringement. To be a class member, a rightsholder needed a book that Anthropic downloaded from Library Genesis in June 2021 or from Pirate Library Mirror in July 2022, carrying an ISBN or ASIN, and registered with the US Copyright Office within five years of first publication and before the download took place. The works list ran to 482,460 eligible titles. By the close of the claims window, 92.77 per cent of them had been claimed — a participation rate almost unheard of in class actions, where the typical claim rate sits around 10 per cent. The payout works out at roughly $3,000 per claimed work.
Read the class definition the other way and its meaning changes. Every eligibility requirement is a proof requirement. The ISBN places a work inside a publishing infrastructure that documents existence and date as a matter of routine. The registration fixes the work's content and ownership on a public record before the infringement. The class did not consist of the creators whose work was taken — it consisted of the creators whose ownership was already provable.
A photographer, illustrator, independent musician, software developer, or self-published writer whose work sat in the same pirate libraries held copyright no less valid. Their work was taken in the same way. But without a registration and without an ISBN, they had no practical route into the recovery, because they could not produce the one thing the process ran on: independent, dateable proof of what they created.
The settlement's edges make the same point from the other direction. Rightsholders who opted out — among them Chicken Soup for the Soul and journalist John Carreyrou — are pursuing their own individual actions, and the court's final order reminded non-class members that, precisely because the settlement excludes them, they remain free to bring their own claims. Every one of those future claims begins at the same threshold: proving what the claimant owned, in what form, and from what date.
The difference prior proof makes — the Anthropic settlement as a case study
With formal proof of ownership | Without formal proof of ownership | |
Work in the pirated dataset? | Class member. Eligible to claim roughly $3,000 per work. | Not a class member. No route into the recovery. |
Copyright valid? | Yes | Yes — but unprovable to the required standard |
Work taken in the same way? | Yes | Yes |
Practical outcome | Claimant | Bystander |
The registration requirement that shaped this class is American; the underlying proof burden is universal. One point needs stating precisely, because it protects creators from a false hope. In the United States, access to statutory damages and attorneys' fees under 17 U.S.C. § 412 requires timely registration with the US Copyright Office. No third-party timestamp, blockchain record, or private deposit substitutes for that registration. What independent proof of ownership establishes is different and broader: that a specific work existed, in a specific form, on a specific date, in the hands of a specific creator. In US proceedings that record supports authentication under Federal Rule of Evidence 901; it complements registration rather than replacing anything on the statutory side.
The European Union and the United Kingdom run on the opposite model, and the proof problem cuts the other way there. Under the Berne Convention, copyright arises automatically on creation, with no registry anywhere recording the right. That spares creators an administrative step and removes the safety net in the same stroke: when authorship or dating is disputed, no public record exists to point to, and the entire evidential burden falls on material the creator controls. Draft files, hard-drive dates, and self-addressed emails all carry the same weakness — the record's timing rests entirely on the word of the person it benefits.
Within that gap, the qualified electronic timestamp occupies a specific legal position. Under eIDAS Article 41, a qualified timestamp carries a legal presumption of accuracy in EU member states and cannot be denied legal effect solely because it is electronic. UK law, which retained the eIDAS framework after leaving the EU, treats qualified timestamps as strong electronic evidence. The UK adds one narrow statutory aid: under section 104 of the Copyright, Designs and Patents Act 1988, a person named as author on published copies is presumed to be the author until the contrary is proved. The presumption helps published authors and does nothing for the unpublished draft, the unreleased track, or the codebase that never carries a name on a title page. In a system with no registry, an independent record of that standing is not an upgrade to the creator's evidence — in most disputes it is the only independent evidence available.
Not every timestamp is evidence; the standard is independence, verifiability, and permanence. The oldest form of self-help — posting the manuscript to yourself, the so-called poor man's copyright — fails on the first requirement, and courts have given it no meaningful weight. Emailing a file to your own inbox or relying on file-system dates fails the same way: each record lives on systems the creator controls, and its timing is exactly what an opposing party will dispute. Free online timestamping tools fail more subtly. A bare hash on a blockchain proves that some data existed at some point, but without an accredited provider, a verifiable certificate, and a durable link between hash, file, and creator, the record's evidential weight in court remains limited.
Cryptographic prior proof at the qualified standard works in four layers:
Set that against the Anthropic fact pattern. A creator holding such a record for each work is no longer reduced to asserting authorship and hoping to be believed. If the work surfaces in a training dataset, a leak, or a competing product, the creator can demonstrate, from a source no opposing party controls, what the work contained and the date it existed — the same threshold the settlement administrator applied before releasing a single dollar. For the class members, publishing infrastructure supplied that evidence as a by-product of the trade. For everyone else, nothing supplies it by default.
The lasting significance of the settlement is not the fund but the filter. The pattern is already repeating: most new copyright actions against AI developers now lead with allegations of downloading from shadow libraries, taking the route this settlement mapped — past the unsettled fair use question and straight to unlawful sourcing, where proven ownership converts directly into recovery. When the next of those cases reaches a class definition, it will be drawn the way this one was: around the rightsholders whose ownership can be demonstrated to an administrator, on paper, at scale. Class definitions are not written generously. They are written around available proof.
For traditionally published authors, the publishing trade supplied that proof as a by-product — the ISBN on every edition, the registration filed on publication, the contract in the drawer. The 92.77 per cent claim rate, against a class action norm of around 10 per cent, is what a population looks like when its evidence infrastructure already exists. No equivalent infrastructure stands behind the photographer's archive, the independent release, the private repository, or the unpublished manuscript. For those creators, a provable record exists only where it was established deliberately, before anything made it urgent.
The settlement put a number on the difference — roughly $3,000 a work, multiplied across more than 447,000 claimed works. That is the dividing line this case drew, and nothing about the next case will move that line. The work decides nothing; the record decides everything.
NDA Breaches and the Evidence Problem: Proving What Was Disclosed Under an NDA
CPR 32.19: What UK Courts Require to Prove a Document Is Authentic
Bartz v. Anthropic Settlement: What Authors Need to Know — The Authors Guild
James Snell is the founder of Provlyn, a platform providing cryptographic prior proof of IP ownership. provlyn.com