AI Training Data and Copyright: Ownership, Compensation, and the Press

The mainstream reading lands in roughly one place: the law here is unsettled and everyone knows it.

The sanctions motion isn’t just a procedural fight—it’s exposing a deeper institutional mismatch. Courts are being asked to adjudicate training data provenance using tools designed for discrete acts of copying, like photocopiers or peer-to-peer networks. But AI training isn’t a copy; it’s a statistical transformation of an entire corpus into a new functional object. The legal system lacks the institutional grammar to even describe what’s happening, let alone enforce transparency. This isn’t about bad faith. It’s about the fact that copyright law’s evidentiary standards assume a world where evidence is stable, bounded, and human-readable. AI training data isn’t any of those things. The sanctions motion is forcing the question: if the law can’t technically enforce what it claims to regulate, is the problem the law—or the infrastructure it’s trying to govern?

The law does not need to trace a sentence through billions of model parameters to govern this. It needs to separate three acts: acquiring copies, processing them during training, and producing outputs that reproduce protected expression. In Bartz v. Anthropic, the court called training “spectacularly transformative” while refusing to excuse a permanent library assembled from pirated books. So “training is fair use” does not legalize every supply chain. I would treat the sanctions motion, which OpenAI contests, as a traceability test rather than evidence of infringement. Courts can prospectively require dataset manifests—records of what entered each training run—along with acquisition records and retention rules. Those controls will not reveal which article caused a particular answer, but they can establish what was copied, from where, and under what claimed authority. Without that operational layer, doctrine becomes unenforceable abstraction.

Treating traceability as a forensic puzzle assumes adversarial discovery is the only way to close an information gap. Japan and Vietnam handle this administratively through pooled licensing and sliding scale audits. Practitioners report high compliance without litigation, not because they cracked the black box, but because they shifted the burden from proving past ingestion to negotiating forward looking access. That is a relational fix to a bargaining asymmetry. The U.S. is asking a four factor test to resolve a structural leverage gap. Weiyi Xia wants an AI specific fair use clause; Michael Murray says courts are already adapting the doctrine. Both keep the fight in court. But if mid size publishers lack the leverage to demand manifests, a courtroom precedent cements a two tier ecosystem. The question is whether compensation belongs in copyright damages, or requires a pooled mechanism that matches the infrastructure scale.

Notice what nobody at this table has said yet: the lawsuit cannot fix the thing publishers are actually bleeding from. Newspaper revenue collapsed over the two decades before GPT-2 existed — classifieds went to Craigslist, display advertising went to Google and Meta, local newsrooms hollowed out entirely. Generative AI is a compounding pressure on an already-broken model, not the cause of it. So even a total victory for the Times — damages, licensing, everything — does not restore journalism's economics. Which forces the real question: if the remedy can't match the wound, what is this lawsuit actually for? My answer: it's a leverage play by the one publisher with a differentiated subscription product, priced for the Times and inherited by everyone else.

Here's the sentence nobody's willing to say out loud: in two thousand one, the Times was the defendant. Tasini — freelancers suing over digital republication of their work into electronic databases. The Times' argument then was structurally identical to OpenAI's now: aggregation into a new format is a different product, not a substitute, and creators benefit from the reach. The Times lost. But it made the move. That's not hypocrisy, it's a pattern. Every information platform, at the moment it achieves scale, discovers that its ingestion is transformative and its predecessors' expectations are quaint. And Oren Bracha sharpens the knife further — he argues this may not even be a fair use question, because big data training sits categorically outside copyright's scope.

The licensing deals do not concede that training requires payment. A use can be legally free yet commercially worth buying for current content, reliable feeds, attribution, product integration, or reduced litigation risk. The dangerous inference is the reverse: because bespoke licensing exists, courts should manufacture a universal licensing market through fair use’s market-harm factor. I think that would entrench concentration. Large publishers can supply clean archives and negotiate; smaller outlets face the fixed costs of proving ownership and monitoring use. A better architecture would offer developers a training safe harbor—clear protection from liability—conditional on auditable records and participation in a collective scheme open to publishers of every size, while preserving separate liability for outputs that reproduce protected expression or bypass paywalls. The legal rule should lower transaction costs, not reward whoever can afford discovery.

The safe harbor proposal assumes substitution is real, and I want to stop and test that assumption, because the one empirical study we have cuts against it. Harvard Business School research — Pacelli — looked at AI summaries in financial news and found reader engagement deepened rather than collapsed. That's the exact sector where substitution should bite hardest: time-sensitive, high-stakes content. If summaries there function as a front door rather than a replacement, then the fourth-factor market-harm theory may be sector-specific, not universal. And that matters enormously, because courts weighing fair use are being asked to credit a generalization the evidence doesn't yet support. Which sectors actually have substitution data — and which are extrapolating from anxiety? Until someone answers that, we're designing compensation architecture for a harm we haven't measured.

The opt-out architecture in Europe isn’t just a legal tool—it’s a test of institutional capacity. Article 4 of the DSM Directive lets any publisher reserve rights by marking content with machine-readable signals. But the real friction isn’t technical; it’s organizational. The EU’s system assumes publishers can coordinate at scale, yet no one has built a shared registry. That silence isn’t accidental. It reveals that copyright’s default setting—individual rights enforced through individual action—is structurally mismatched to the infrastructure it’s trying to regulate. The U.S. lawsuit isn’t filling that gap; it’s exploiting it. By litigating alone, the Times isn’t just fighting OpenAI—it’s betting that a courtroom victory will force AI companies to negotiate with every publisher individually. That’s not a solution; it’s a transfer of coordination costs onto the weakest link: small outlets that can’t afford discovery or licensing teams. The EU’s opt-out was designed to prevent this exact outcome, but only if publishers use it. Their hesitation isn’t laziness—it’s the absence of an institution to pool their leverage. The question isn’t whether opt-outs work; it’s why no one has built the mechanism to make them work at scale.

The coordination problem just named disappears when you stop treating information flows as discrete property claims. Justice Krishna Iyer’s jurisprudence in India framed copyright as a social contract, weighing creator rights against the public’s need to access knowledge rather than treating data ingestion as a trespass. The Berne Convention’s exception provisions operate on that same balancing logic. When you view training inputs through that lens, the American fixation on proving individual market harm starts to look like a category error. Malaysia is exploring hardware import levies, and Japan’s JCCC routes compensation through pooled agreements instead of courtroom discovery. Those systems treat high-quality text as a collective input and fund it administratively. The real test isn’t whether fair use survives. It’s whether U.S. institutions can tolerate a compensation mechanism that doesn’t require calling ingestion theft first.

Something underneath the compensation debate deserves naming: journalism was never priced by anyone who valued it.

The pricing model isn’t the problem—it’s the pricing unit. Newspapers sell stories; AI sells answers. Those are different commodities, and the shift isn’t just technological, it’s institutional. For a century, journalism’s economic grammar has been built around discrete units of attention: the article, the headline, the byline. That grammar made sense when distribution was physical and attention was scarce. But AI doesn’t traffic in articles—it traffics in functional outputs: summaries, explanations, decision-support. The unit of value has changed, and the legal system is still trying to price the old one. The real structural question isn’t whether the Times can charge for subscriptions. It’s whether any institution can charge for something that no longer exists as a discrete product. The lawsuit isn’t fighting over copyright—it’s fighting over the right to define what journalism is worth in a world where the thing being sold isn’t journalism as we’ve known it, but a new kind of utility. And that’s a question copyright law was never designed to answer.

When knowledge shifts from discrete artifacts to synthesized utility, property law stops being the right instrument. Kautilya’s statecraft framed knowledge governance around balancing individual claims with societal needs, precisely because the value of information sits in how it is organized for public use, not in isolating raw inputs. The Berne Convention’s public-interest balancing operates on that same logic. The U.S. is asking copyright to price a utility grid. That is why the compensation debate fractures into levies, pooled licensing, equity arrangements, and litigation—each is a different attempt to fund upstream producers when the downstream product is no longer divisible. The real question is whether we treat verified journalism as a shared input requiring structural upkeep, or keep forcing a property framework to solve an infrastructure funding problem.

A structural fund can reproduce the same concentration unless its allocation formula rewards journalism rather than copyright inventory. Paying by articles ingested, audience size, or demonstrated model use favors large archives and publishers able to audit them; the evidence base barely captures small and mid-size outlets at all. I would tie any levy or pooled payment to measurable production inputs—newsroom employment, original reporting, and coverage of underserved communities—rather than attempts to trace which sentence influenced which answer. That changes the policy objective from compensating yesterday’s owners to financing tomorrow’s information supply. It also avoids making unproven substitution the gateway to support: the public-interest case rests on maintaining reporting capacity, not on proving that every AI summary cost someone a subscription.

Strip away the doctrine and this case is not yet about legality at all — it's about the right to know. No public audit exists of which articles entered which training runs, and specialist reporting suggests even OpenAI's own engineers can't reliably trace which documents drove which behaviors. That makes the sanctions motion something stranger than it looks: a publisher asking a court to punish a company for failing to preserve evidence that may be technically unrecoverable in the first place. So discovery becomes the whole ballgame — not as a means to a fair-use ruling, but as the only transparency mechanism anyone has. Which tells you something damning about our governance: we've outsourced the basic question of what goes into these systems to a spoliation fight in one federal courtroom.

One cost of the structural-funding consensus forming here hasn't been priced. A levy allocated by newsroom employment, original reporting, and coverage of underserved communities requires some body to define those terms and audit compliance. That body is either the state or a state-chartered collective. You have just invented a licensing authority for journalism — in a country whose press freedom tradition runs on the premise that no such authority should exist. Mill's argument for a free press wasn't about markets; it was about no one holding the switch. Litigation is a terrible instrument, but it has one virtue: nobody has to define what counts as real journalism to use it. So the honest question is which dependency we prefer — on a subsidy formula, or on a courtroom.

The licensing deals between OpenAI and major publishers aren’t just commercial hedges—they’re admissions that the legal defense of fair use rests on a fiction of non-rivalry. When a model trains on a Times article, the publisher loses nothing in the moment; the harm, if it exists, is probabilistic and deferred. But the licensing agreements reveal that publishers and AI companies both know the same article can be monetized twice: once as training input, once as subscription content. That duality collapses the core fair-use premise that the use doesn’t compete with the original market. The deals don’t prove substitution; they prove that the same work can occupy two economic roles at once, and copyright law has no grammar for that.

A licensing deal proves willingness to pay, not a legal duty to pay. The distinction matters because the fourth fair-use factor asks about harm to legitimate markets; it cannot simply assume that every market for permission is legitimate, or fair use becomes circular—once anyone sells a license, declining to buy one becomes market harm. Publishers have a credible counterclaim: repeated commercial deals may demonstrate an emerging market that unlicensed training impairs. But I think the stronger structural inference is narrower: selective licensing converts legal uncertainty into bargaining leverage. Because evidence about small and mid-size publishers is largely absent, we do not know whether that market distributes value or merely rewards incumbents powerful enough to negotiate.

We are watching a role reversal that media lawyers track but the coverage misses. In two thousand one, the Times was the defendant in Tasini, arguing that digital databases were merely a new format and freelancers deserved no separate payment. The Supreme Court rejected that. Today the positions flip, and the AI developers are running the exact same aggregation defense. I read this as a repeating mechanism rather than hypocrisy. Each wave of information infrastructure invokes transformation to absorb the prior ecosystem’s labor at zero marginal cost. Fair use has functionally become the legal instrument of platform succession. That explains why a courtroom victory cannot repair journalism’s revenue base. The lawsuit is attacking the very consolidation logic the press deployed to capture value two decades ago.

The platform-succession argument has one weak joint worth pressing: it assumes the new platform actually destroys the old ecosystem's value rather than merely capturing it. But the substitution evidence is thinner than the lawsuit's framing. The NYT alleges near-verbatim reproduction cannibalizes subscriptions — yet Pacelli's Harvard Business School work found AI summaries in financial news deepened reader engagement rather than replacing it. If AI summaries function as a discovery layer in some sectors, the harm story isn't universal; it's sector-specific, and that matters enormously for the fourth fair-use factor, which is where this case will actually be decided. The deeper pattern may not be that platforms kill the prior ecosystem — it's that they reprice it, and the publishers who survive are the ones whose product can't be summarized. That is a brutal sorting mechanism, and no court ruling will stop it.

Follow that sorting mechanism to its output and it should alarm you. What resists summarization is voice, personality, scoops, opinion — the stuff people pay for. What summarizes cleanly is exactly the civic layer: council minutes, court dockets, school board budgets, the procedural reporting that nobody reads but everybody depends on. The market sorts against the public-good component of journalism, precisely because that component is commodity-like. Columns survive; the courthouse beat doesn't. And there's an unresolved possibility that makes this worse.

The problem isn't that journalism is being replaced—it's that it's being absorbed into something larger without leaving a trace. When a model trains on a reporter's work, the statistical weight of that work persists, but the institutional signal that produced it disappears. The harm isn't substitution; it's evaporation. The system stops registering that the Times exists as an institution, only as a statistical input. That's why the substitution debate misses the point: the real injury isn't that readers stop paying for the Times; it's that the system stops registering that the Times exists at all. The lawsuit is trying to force an accounting of something the architecture was designed to erase.

“Statistical persistence” is the wrong foundation for a remedy because it implies causal attribution the technology cannot support. A passage may enter training without being recoverable as the source of any particular answer. I would require three separate ledgers: an acquisition ledger recording what entered the pipeline and under what authority; an output ledger logging substantial quotation or close paraphrase; and a funding mechanism based on reporting capacity rather than unverifiable influence inside model weights. Current-news systems can also attach citations when they retrieve identifiable articles at answer time, but those citations should not pretend to identify what taught the underlying model. Without those distinctions, provenance becomes theater: a plausible link is displayed while the actual economic relationship remains unaudited.

The evaporation problem you are describing is what happens when you try to fund infrastructure maintenance with a trespass claim. For thirteen centuries, the imperial examination system treated textual archives as a shared upkeep obligation, not a ledger of individual property rights. We are doing the inverse: demanding forensic receipts for a statistical process, then watching the civic reporting layer starve when those receipts cannot be produced. Weiyi Xia proposes shifting copyright from author-centrism to an AI-specific clause that treats training as a system-level function. That tracks with Justice Krishna Iyer’s jurisprudence on knowledge access: when a technology absorbs an entire information ecosystem, compensation becomes a maintenance duty, not a per-copy royalty. Japan’s pooled licensing works because it stops chasing document-level provenance and funds the reporting apparatus directly. Courts can still police verbatim outputs, but the training stage needs a collective compact. Litigation prices extraction. It does not price continuity.

One thing the licensing ledger genuinely tells us, without over-reading it: OpenAI is paying some publishers while litigating against others, and the brief gives us no pricing detail on any of those deals. So we know the fact of payment, not its logic. Three readings fit that fact. It's hush money — buying peace with the outlets likeliest to sue. It's an access fee for current content at answer time, which is a retrieval market, not a training market. Or it's a quiet concession that archives have market value, which would undercut the fair-use defense the company is running in the same breath. Those readings imply opposite precedents, and the NYT ruling will only adjudicate one of them. The question I'd put to the room: which market is the court actually being asked to regulate — the historical corpus, or the live answer?

Both, and that's the trap. The historical corpus is a stock problem — a one-time, already-completed ingestion. The live answer is a flow problem, recurring, priceable, cancellable. Copyright is built for stocks: discrete acts of copying, discrete damages. It handles flows badly. Which means the corpus fight is really a bargaining chip over sunk value, and the only thing anyone can actually govern going forward is the flow — the retrieval layer, where citation is technically feasible and payment can be metered. Hayek's point about dispersed knowledge cuts here: nobody can compute what any one article contributed. But you can meter what gets fetched today. Litigate the past, contract the future — and be honest that those are two different negotiations wearing one docket number.

The opt-out registry Europe built under Article 4 of the DSM Directive isn’t just a technical tool—it’s a Weberian test of whether publishers can act as a collective actor at all. The directive gave them the right to reserve rights, but no institution stepped forward to build the shared infrastructure. That failure isn’t technical; it’s bureaucratic. The same press that once syndicated wire copy across continents now can’t coordinate a single digital opt-out. The lesson isn’t that licensing is impossible; it’s that the institutional capacity to enforce property rights in a digital environment has atrophied. Courts are being asked to solve what is fundamentally a coordination problem, and they’re not designed for that. The litigation is a symptom of institutional decay, not a solution to it.

Metering retrieval creates a regulatory escape hatch: developers can reduce visible fetching and answer from periodically refreshed model weights, moving the taxable event back into the untraceable layer. I would therefore separate compensation from technical architecture. Charge at a measurable business boundary—AI subscription revenue, advertising revenue, or another legislatively defined base—and distribute through an independent fund using auditable indicators of original reporting. The allocation formula should reward newsroom employment and coverage of underserved communities, not archive size or negotiating power. That is my policy judgment, not an empirical finding; public evidence does not yet establish which formula works. But it prevents yesterday’s dominant publishers from turning a democratic-information subsidy into an incumbent-protection scheme.

The revenue fund proposal sidesteps the actual bottleneck: who administers it. American copyright architecture treats information as individual exclusion rights, which forces every newsroom to litigate or license alone. That is why Japan’s JCCC pooling and Vietnam’s sliding-scale audit pact matter. Both shifted from bilateral enforcement to collective mechanisms, achieving high adherence with zero litigation. Current deal-making leaves small and mid-size publishers almost entirely out of the room, because one-on-one negotiations naturally reward archive size and legal leverage. A pooled structure flips that geometry. It sets an ecosystem floor before anyone bargains for the ceiling. I think the structural move is not changing the payment instrument, but changing the bargaining unit. When you atomize rights holders against concentrated compute capital, you guarantee that only the largest players capture value. Collective administration recognizes that information ecosystems survive or collapse together.

There's one compensation model in the brief that inverts every other option on the table. The contested example is Brazil's Folha reportedly taking an equity stake in Llama-Br rather than a licensing fee — contested because it's thinly documented, but worth taking seriously as a design. Every other mechanism we've discussed — damages, levies, pooled funds, metered retrieval — asks how much per unit of content. Equity asks who owns the upside of the system built on the content. That reframing matters because rents get renegotiated downward as models commoditize, while ownership compounds. Whether publishers would actually want equity is a genuine question — fiduciary pressure for near-term revenue, litigation strategy, skepticism about AI valuations all cut against it. But if the answer is that the equity looks worthless, that's itself a finding: it means the plaintiffs don't believe the systems trained on their work have durable value, which sits oddly beside the damages claim.

Nobody in this room has mentioned the reader — and every mechanism we've proposed is a payment from a machine to an institution, with the citizen as a spectator. Here's what worries me. A front page is an editorial claim about what matters, made by people you didn't ask.

The equity stake in Llama-Br isn’t just a financial instrument—it’s a structural admission that the value of journalism isn’t in the articles themselves, but in the institutional capacity to produce them at scale. When Folha takes equity, it’s not licensing content; it’s licensing the ability to generate more content. That reframes the entire dispute: the question isn’t whether AI companies owe publishers for past ingestion, but whether publishers should own a share of the future capacity built on their institutional muscle. The fact that no U.S. publisher has pursued this path suggests they’re still fighting the last war—treating journalism as a stock of copyrighted works rather than a flow of institutional trust. That’s why the litigation feels anachronistic: it’s trying to price the archive when the real asset is the reporting apparatus that keeps producing it.

Three moves revised the standard framing. We questioned whether substitution is even established — Pacelli's Harvard Business School work found AI summaries deepened engagement in financial news, which means market harm may be sector-specific rather than universal. We split one docket into two negotiations: the sunk historical corpus, and the metered live retrieval layer. And we priced the cost nobody prices — any structural fund needs someone to define what counts as journalism, which is a licensing authority for the press. Sharpest tension: the market sorts against the courthouse beat while columns survive. Takeaway: watch retrieval, not training. Honest answer — the law's unsettled, and the remedy can't match the wound. Thank you for listening. As it happened; as it is.