The latest wave of AI copyright cases does not turn on whether AI may learn from books — courts have started ruling on that. It turns on where the training data came from, and whether the creators whose work was used can prove they made it. Most cannot.
On 5 May 2026, five of the world's largest publishers — Elsevier, Cengage, Hachette, Macmillan, and McGraw Hill — and the bestselling novelist Scott Turow walked into a Manhattan federal court and sued Meta and Mark Zuckerberg personally. The complaint alleges that Meta torrented millions of copyrighted books and journal articles from pirate sites to train its Llama AI models, and that it stripped the copyright information from those works to conceal the source. The number at the centre of the complaint is 267 terabytes — many times the size of the Library of Congress's print collection.
But the scale is not what makes this case matter to you. What matters is the question the case is built around. The first wave of AI copyright lawsuits asked whether training on copyrighted work was legal at all. This one asks something narrower and far more consequential for every individual creator: where did the data come from, who created it, and can they prove it? That is a provenance question — and provenance is something most creators have no way to establish.
That gap is the one Provlyn exists to close. Provlyn gives any creator an independent, verifiable record of what they created and when — the same kind of evidence that separates the claimants in the Elsevier case from the millions of creators whose work was also in that dataset but who have no route into the recovery.
This post explains why the legal battleground has shifted to provenance, why that shift leaves most creators dangerously exposed, and how cryptographic prior proof of ownership gives any creator the same kind of provenance evidence that a major publisher takes for granted.
The first wave of AI copyright litigation failed to settle the training question — and in failing, revealed that the real cases will be won on provenance and evidence. The litigation began in 2023 framed around one question: is training a large language model on copyrighted material an infringement of the original work? The courts have started to answer that question — but not cleanly, and not in a way that settles much.
In June 2025, two judges in the same federal district reached different conclusions within days of each other. In Bartz v Anthropic, Judge William Alsup ruled that training on lawfully acquired books was fair use, but that downloading copies from pirate sites was a separate act that was not protected — and that case settled for $1.5 billion. In Kadrey v Meta, Judge Vince Chhabria ruled that Meta's training was fair use even though the books came from shadow libraries, treating the downloading and training as a single integrated process. He added, pointedly, that the authors lost not because they were wrong but because they made the wrong arguments and failed to develop a record on market harm.
That is the real state of the law: unsettled, fact-specific, and turning increasingly on two things — how the data was acquired, and whether the rights holder can evidence harm. The training question is not closed. It has simply revealed that the cases will be won or lost on provenance and evidence, not on grand principles about whether AI may learn from books.
The Elsevier complaint is built precisely for that environment. The plaintiffs are not relying on the argument that Meta was wrong to train Llama on books. They are arguing that Meta acquired the books unlawfully and concealed their origin — a provenance-and-evidence case, designed to succeed where the Kadrey plaintiffs fell short.
The complaint alleges that Meta considered licensing deals with major publishers in 2023 but decided against them — and that the decision was escalated to Mark Zuckerberg personally. A Meta employee is quoted in the complaint as saying: 'If we license once [sic] single book, we won't be able to lean into the fair use strategy.'
Whether or not that quote survives evidentiary challenge, the theory it illustrates is now public, attributed, and central to one of the largest copyright cases ever filed. The plaintiffs are seeking statutory damages, a permanent injunction, and an order requiring the destruction of the models trained on the disputed material. The allegation, in short, is that the choice to source from pirated rather than licensed material was deliberate — and that the provenance of the training data is the whole case.
The Elsevier case is being brought by parties who can prove ownership. The publisher plaintiffs hold tens of thousands of registered copyrights with International Standard Book Numbers. The proposed class extends to all legal or beneficial owners of registered copyrights for any book possessing an ISBN or journal article possessing a DOI or ISSN.
Read that class definition the second way it can be read. It includes everyone with formal proof of ownership. It excludes, in any practical sense, everyone without it. A photographer, illustrator, independent musician, software developer, course creator, or self-published writer whose work was caught in the same 267 terabytes — but who has no ISBN, DOI, or ISSN — is not a class member. Their copyright is no less valid. Their work is alleged to have been used in the same way. But they have no practical route into the recovery.
The provenance gap — who can claim and who cannot
With formal proof of creation | Without formal proof of creation | |
Work in the 267TB dataset? | Class member. Can pursue statutory damages. | Not a class member. No route into the litigation. |
Copyright valid? | Yes | Yes — but unprovable at the required standard |
Position in the case | Claimant | Bystander |
Practical outcome | Damages, injunction, or licensing position | Nothing |
This is the pattern that will repeat in every case to come. The creators who can prove what they made, when they made it, and that it pre-existed the alleged use are the ones who recover damages, secure injunctions, and negotiate licensing terms. The creators who cannot are bystanders to a process that runs entirely on evidence they do not hold.
Cryptographic prior proof of ownership gives any creator the same kind of provenance evidence that ISBN-level publishing infrastructure gives a book — without a publisher, a registration process, or any administrative overhead. When you deposit a file with Provlyn, you create an independent, verifiable record of exactly what that file contained and when it existed — impossible to alter after the fact, and verifiable by any court or counterparty without relying on your word.
One important distinction: accessing statutory damages in the United States requires registration with the Copyright Office within a defined window — Provlyn does not substitute for that. What prior proof provides is the independent evidential foundation every ownership claim rests on — the record of what existed, in what form, and when, established by a third party before any dispute begins.
What Provlyn does at deposit — in order:
Place that against the Elsevier fact pattern. A creator with a Provlyn vault certificate has an independent, verifiable record of exactly what they created and when. If their work later appears in a training dataset, a leaked file, or a competing product, they are not reduced to asserting authorship and hoping to be believed. They can prove it, to the evidentiary standard the entire litigation turns on. They have the evidential foundation to bring a claim — rather than being a bystander to a process built on evidence they do not hold.
Provlyn's vault certificates carry a legal presumption of accuracy under Article 41 of the EU eIDAS Regulation: a qualified electronic timestamp is presumed accurate as to the date and time it indicates and as to the integrity of the data it relates to, and that presumption is binding across all EU member states and already established in statute.
In the United States — where the Elsevier case sits — there is no equivalent federal statute creating the same presumption. But under Federal Rule of Evidence 901, electronic records are admissible where their integrity can be shown through a documented, reproducible process and cryptographic verification. A SHA-256-hashed, RFC 3161-timestamped, Bitcoin-anchored deposit issued by an accredited Trust Service Provider is precisely that kind of record. The argument against admitting it would be procedural, not substantive — and procedural authentication is a solvable problem. As US courts work through the provenance questions these cases raise, it is difficult to construct a coherent argument that a mathematically verifiable, independently anchored timestamp is not evidence of when something existed.
Elsevier v Meta will take years to resolve. The settlement, if it comes, may be larger than Anthropic's or smaller. The precedent may shift. What will not change is the structural fact underneath it: AI systems have ingested enormous volumes of creative work from sources of contested provenance, and the legal consequences flow to the rights holders who can prove ownership.
For every creator producing original work today — fiction, music, photography, design, software, research — the question is no longer whether their work might end up in someone else's dataset. It is whether, when it does, they will have the evidence to do anything about it. Prior proof is, by definition, prior. The time to establish it is before it is needed.
Establish prior proof of ownership before your work enters someone else's dataset
Provlyn provides cryptographic prior proof of ownership for any digital artefact. Each deposit is hashed with SHA-256, timestamped by an accredited Trust Service Provider under RFC 3161, qualified under eIDAS Article 41, and anchored on the Bitcoin blockchain. The vault certificate carries a legal presumption of accuracy under the EU eIDAS Regulation, is admissible evidence in UK and EU jurisdictions, and remains valid permanently, independently of Provlyn's continued operation. Start your free 7-day trial at www.provlyn.com. No credit card required.
Related Reading
Holland & Knight: Major Publishers Challenge AI Training Practices in Landmark Copyright Suit Against Meta — case analysis
CourtListener: Elsevier Inc. v. Meta Platforms, Inc., 1:26-cv-03689 — case docket
A $1.5 Billion Settlement — and Why Most Creators Couldn't Have Claimed a Penny — Provlyn — the Anthropic settlement that preceded this case
Blockchain-anchored timestamps and court-admissible certificates. Prove it. Permanently.
Get started free