A national newspaper's parent company has sued an AI company over how its journalism was used to train a model. That is the entire verifiable payload. No named defendant beyond the anonymized OpenAI. No enumerated works. No damages figure. No jurisdiction. No cited statute in the aggregated brief that reached me through a crypto media outlet—which, itself, tells you something about the state of the coverage.
I have read enough complaints to know what a thin filing means. It is not laziness. It is positioning. The plaintiff is not trying to win on the facts of a single article. It is trying to establish standing for a category: the category of "our corpus was ingested without a license." That is a provenance claim wearing copyright language.
Code does not lie, but it often omits the context. The context here is the training pipeline itself—what data went in, where it came from, under what license, and what the model retained. None of that appears in the brief. All of it will surface in discovery. And discovery is where this case, and the dozen like it, will actually be decided.
I want to examine this not as a legal commentator but as someone who has spent years building zero-knowledge systems for precisely the kind of provenance verification this lawsuit proves the industry never implemented. The reflex, watching from inside the crypto stack, is to say "blockchain fixes this." That answer is wrong in a specific and instructive way. I will show you why.
The Mechanics Nobody Bothers to Explain
Start with the statute, because the industry's habit of skipping it produces bad takes. The action sits under the U.S. Copyright Act, 17 U.S.C. Four sections matter. §106 grants the holder exclusive rights—reproduction, derivative works, public display. §107 defines fair use as an affirmative defense across four factors. §501 defines infringement. §504 sets remedies, including statutory damages, which for willful infringement can reach tens of thousands of dollars per work.
That last number is the leverage. A newspaper archive is not one work. It is tens of thousands. If statutory damages attach per work, nominal exposure scales multiplicatively, not additively. This is why media plaintiffs almost never plead actual damages—they are hard to prove, and the statutory track is far more potent. The arithmetic alone makes the case worth filing even when the merits are uncertain.
For the AI defendant, the merits turn almost entirely on §107's first and fourth factors: the purpose of the use, and the effect on the market for the original. The first factor used to be a comfortable place for technology defendants. Authors Guild v. Google in 2015 blessed mass digitization as transformative because the output was a search index, not a readable copy. That precedent has been narrowing. Andy Warhol Foundation v. Goldsmith in 2023 tightened the transformative-use test, requiring a further purpose or different character that is more than incidental.
Training a model on news to produce a model that can produce news is not obviously "different character." The purpose of ingestion is arguably the same commercial function the original served: informing readers. A defendant leaning on the Google Books lineage is leaning on a precedent the Supreme Court has already narrowed. I have watched teams make exactly this error in DeFi—citing a pre-exploit audit as proof of safety after the threat model changed. The document did not change. The world did.
The third factor—amount used—is worse for the defendant. Training ingests the full corpus, not fragments. Google Books won partly because the display was fragmentary. A trained model does not display fragments; it internalizes the whole and can emit reconstructions. That is a materially different posture, and it is where the fair-use defense is thinnest.
Output-side risk is the frontier nobody has priced. If a model can emit text substantially similar to a source article, that is a derivative-work claim, which is a separate violation from the ingestion itself. But if the model only learns abstractions and does not memorize, the traditional infringement framework struggles to attach. I have run memorization tests on models as a diligence exercise; the results are uncomfortable for anyone who assumed training was a one-way absorption. Whether the model retained the work is an empirical question, and empirical questions are exactly what discovery answers—which loops back to why the discovery dilemma matters.
This Is a Cluster, Not a Case
The single most important context for anyone pricing this: it is not isolated. Media, authors, image libraries, and music rights holders have filed parallel actions against AI developers. A major newspaper has sued over training data. A stock-image company has sued a generative-image developer. Author associations have filed class actions. Record labels have sued music-generation platforms.
The cluster matters because these cases share fact discovery and legal argument. A finding against one defendant becomes ammunition against the next. Risk is contagious. A single unfavorable holding on the third or fourth fair-use factor propagates across every pending and future suit, which is precisely why defendants fight so hard to keep these cases from reaching summary judgment—and why courts, wary of resolving novel questions prematurely, tend to let them proceed into discovery and trial.
The regulator here is not a regulator. The U.S. Copyright Office registers and advises; it does not enforce. Its series of AI policy reports carries no legal force, but courts cite it, which gives it de facto influence. The FTC could theoretically reach AI data practices under its unfairness authority, but the main event is civil litigation, not administrative action. This is an important structural point: there is no administrative window guidance, no sandbox, no compliance safe harbor. Copyright liability cannot be waived by a regulatory pilot. That is the fundamental difference from fintech, and it means AI developers cannot route around this through an innovation exemption.
Where the Crypto Analogy Actually Holds
Now the part my colleagues get wrong.
The instinct is to propose on-chain attestation of training data: publish a Merkle root of your corpus, timestamp it, prove provenance, done. I have built these systems. They are elegant. They also solve the wrong layer.
An on-chain attestation proves a specific dataset existed at a specific time. It proves integrity. It does not prove the right to use that dataset. A cryptographic commitment to a corpus you scraped without a license is a notarized confession, not a defense. Code does not lie, but it often omits the context—and the context here is legal title, which no hash can establish.
This is the distinction the crypto compliance narrative keeps eliding: provenance (where did this come from) versus entitlement (were you allowed to take it). Blockchain is excellent at the first and structurally incapable of the second, because entitlement is a social and legal construct, not a mathematical one. You cannot prove a license you never signed by hashing the content you ingested.
Where the analogy holds is narrower and more useful. The value of cryptographic provenance in this dispute is evidentiary, not exculpatory. If a defendant can show, with cryptographic certainty, exactly which works entered the pipeline and when, it can bound the class of plaintiffs, price a settlement precisely, and execute a targeted removal or retraining without nuking the entire dataset. That is a damage-control instrument, not a shield.
I have seen this exact pattern in cross-chain bridge security. An on-chain proof that a message was delivered does not tell you whether the signer was authorized. You need both the cryptographic proof and the governance proof. Most bridge exploits live in the gap between them. This lawsuit lives in the same gap, translated to copyright.
The Discovery Problem Is a Zero-Knowledge Problem
Here is the contrarian angle, and it is the one I would stake my reputation on.
The conventional read is "AI company versus media company over fair use." Correct, and shallow. The deeper structure is that discovery—the compelled disclosure of training data, internal communications, and model internals—is itself the weapon, and it is asymmetric against the defendant in a way no fair-use doctrine can fix.
To mount a fair-use defense, a defendant must prove the purpose and character of its use, the nature of the ingested works, the amount taken, and the market effect. Every one of those proofs requires disclosing the training pipeline. But the pipeline is also the defendant's core trade secret. So the defendant faces a dilemma: disclose to defend, and hand competitors a roadmap; or withhold, and lose the defense by default.
This is precisely the problem zero-knowledge proofs were invented to dissolve. A defendant could, in principle, prove properties of its training set—"this corpus contains no more than X% of any single publisher's output," "this work was excluded from training run Y"—without revealing the corpus itself. The proof is verifiable; the underlying data stays private. In 2024, working on a ZK-rollup, I identified a gas inefficiency in the proof-verification circuit and proposed a constraint-system optimization that cut verification cost by 15%. The same class of technique applies here. The mathematics to do this exists. It is not deployed, because no AI lab has treated training-data compliance as a cryptographic engineering problem.
Code does not lie, but it often omits the context. The context the AI industry omitted is that verifiability and disclosure are not the same thing—and the industry chose neither.
There is a second asymmetry, in the burden of proof. Fair use is an affirmative defense. The defendant carries the burden. So the AI company must affirmatively prove the negative—that its use was transformative, that the market was unharmed—using evidence only it possesses and would rather not produce. Information asymmetry compounds legal asymmetry. This is not a fight the defendant can win on doctrine alone.
The Comparative Compliance Puzzle
The same training run can be lawful in one jurisdiction and infringing in another, which turns compliance into a jigsaw no single team can assemble.
The EU's Digital Single Market Directive contains a text-and-data-mining exception, but rights holders can reserve their rights and opt out. China's Copyright Law uses an enumerated fair-use list with no dedicated TDM provision, leaving training in a gray zone. The UK has debated expanding its TDM exception for years without resolving it. The U.S. resolves the question through litigation, not legislation.
Now layer on the EU AI Act, which imposes training-data disclosure obligations on general-purpose models. That obligation is a regulatory artifact, but it doubles as an evidence-gathering tool: what the AI Act forces a lab to disclose in Europe can be cited by plaintiffs in America. Regulatory transparency in one jurisdiction becomes litigation ammunition in another. Any team training on cross-border data is running a compliance map, not a compliance checklist, and the map has holes.
Then there is the privacy overlay. News reports contain identifiable personal information. A training corpus of journalism is simultaneously a copyright object and a personal-data carrier. That stacks copyright exposure on top of GDPR exposure, and the two do not share a defense.
Comparative divergence is not a theoretical inconvenience. It is a design constraint. A team that trains once and deploys globally must satisfy the strictest applicable regime, which is rarely the one it read first.
The Bear Market Lens: Survival, Not Gains
We are in a bear market, and that reframes everything. In a bull market, this lawsuit is a headline that pumps a compliance-token narrative for a week. In a bear market, it is a solvency question for every project whose business model depends on free data.
Ask the survival question: which protocols and AI-adjacent crypto projects are structurally exposed because their value accrues from unlicensed or unverifiable inputs? The ones that marketed "decentralized data" while quietly scraping the open web have the same shape as a lending protocol that marketed "sustainable yield" while relying on a manipulable oracle. I wrote about exactly that in 2020—delayed price feeds producing undercollateralization. The mechanism is identical: a system whose safety depends on an input it never verified.
Watch LP behavior the way I watched lending pools in August 2020. The projects with real licensing behind their data will hold. The ones with a Merkle root and a promise will not. The distinction is not technical sophistication; it is entitlement, and entitlement is what this litigation is really about.
What the Licensed Data Market Actually Looks Like
The endgame of this litigation cluster is not a ruling. It is a price.
Media plaintiffs are not primarily seeking verdicts; they are seeking leverage to sign licensing agreements. The pattern is already visible: AI labs signing content deals with major publishers, sometimes with retroactive carve-outs that waive claims over prior training. That carve-out is the hidden center of gravity. It decides whether the lawsuit is a termination event or a negotiating chip.
If I were specifying the infrastructure for this market—and I have designed privacy-preserving compliance layers for institutional DeFi, so this is not hypothetical for me—I would build it on three primitives. First, a licensed-corpus registry with cryptographic commitments, so entitlement is recorded at ingestion, not asserted after the fact. Second, a zero-knowledge attestation layer that lets a model prove it trained only on licensed data without revealing the data or the model. Third, a settlement layer where per-work royalties clear automatically, because the statutory-damages math that makes this litigation lucrative is the same math that makes per-work licensing administratively impossible without automation.
That third primitive is where blockchain genuinely earns its place. Not as a defense against copyright, but as a settlement rail for a market that cannot function on invoicing. The moment licensing becomes per-work at newsroom scale, the transaction cost of manual clearing exceeds the value of the license. This is the payments insight that gets buried under ideology: the driver of on-chain settlement in developing markets was never blockchain philosophy—it was the failure of existing rails to clear small values. The same failure is about to hit content licensing.
The Vulnerability Forecast
Here is what I expect, stated as a risk position rather than a summary.
The fair-use safe harbor for full-corpus training is closing, and defendants citing Google Books are running an outdated threat model. Warhol changed the first factor, full-corpus ingestion implicates the third, and the fourth turns on a market that is only now forming. The most likely outcome is not a clean verdict but a cascade of settlements, each setting a price the next defendant must beat.
The real exposure is not the damages figure. It is the injunction. A court order barring continued training on specific corpora is a product-iteration kill switch, and its commercial damage exceeds any plausible award. Teams should be modeling the injunction scenario, not the settlement scenario.
And the blind spot almost nobody is pricing: the discovery dilemma. An AI lab that built its pipeline without cryptographic verifiability now cannot prove compliance without disclosing its secrets. That is not a legal problem better lawyers solve. It is an architecture problem that should have been solved at ingestion.
Code does not lie, but it often omits the context. And the context, this time, is that the industry built models on data it can neither prove it was entitled to use nor prove it excluded. Verifiability was optional when capital was cheap. In this market, it is the difference between a licensing negotiation and a winding-up.
The question I would put to every team building on training data right now: if a court ordered you tomorrow to prove, without revealing your model, exactly which works you ingested and under what license—could you? If the answer is no, you are not compliant. You are simply not yet caught.