Who Owns the Training Set?

Every frontier model now in commercial use was trained on a corpus that no one, including the labs that built it, can fully enumerate. Scraped web pages, pirated book archives, scanned court filings, forum posts, photographs uploaded to sites whose terms of service predate the concept of machine learning entirely — the material went in under the working assumption that assembling a dataset was a research act, not a commercial one, and that fair use would cover the gap. That assumption is now the single largest unresolved liability on the balance sheet of the AI industry, and the courts, legislatures, and licensing desks working to resolve it are, in effect, deciding who is allowed to build the next generation of these systems at all.

This is not a niche copyright dispute. It is closer to a property-rights settlement for an asset class that came into being faster than the law could describe it. And like every property dispute of consequence, the resolution will not be neutral. It will favor whoever can absorb the cost of getting it wrong.

The legal theories, and why none of them has won cleanly

The core argument advanced by the major labs is a transformative-use claim descended from cases like Authors Guild v. Google — the logic that undergirded Google Books’ scanning project a decade and a half ago. Training a model on a copyrighted work, the argument goes, does not reproduce that work; it extracts statistical patterns from it, the way a student extracts style and technique from a shelf of novels without infringing on any one of them. Courts have shown real sympathy to this framing. Judge William Alsup’s 2025 ruling in the Bartz v. Anthropic litigation drew a line that has since become a reference point across the industry: training a model on lawfully acquired copies of books is very likely transformative and fair, but sourcing those books from shadow libraries of pirated text is a separate act of infringement that fair use does not launder. The distinction — legitimate acquisition versus how the material is subsequently used — has forced a re-reading of what “trained on copyrighted data” actually means as a legal exposure. It is not one question. It is at minimum two: was the copy obtained lawfully, and does the resulting use substitute for the original in the market.

That second prong is where the plaintiffs’ bar has found its most durable foothold. The market-substitution argument — that a model fine-tuned to write in a particular author’s voice, or an image generator capable of reproducing a photographer’s signature style on command, competes directly with the original creator’s livelihood — sidesteps the transformative-use defense almost entirely. It doesn’t matter how the training happened if the output displaces the market for the input. Getty Images’ suit against Stability AI in the UK, and consolidated litigation against several image and music generators in US federal court, both lean on this theory and have survived early motions to dismiss in whole or in part. That survival matters more than any eventual verdict: it means discovery proceeds, training logs get subpoenaed, and the actual contents of these datasets — long treated as trade secrets — start becoming exhibits.

The fight over training data no longer resembles a copyright dispute so much as a zoning dispute for an entire industry.

A third theory, quieter but potentially more consequential, concerns data scraped from platforms with explicit contractual terms — social networks, stock-photo sites, professional databases — where the claim is not copyright but breach of contract and, in some jurisdictions, unauthorized computer access. Reddit’s suit against Anthropic over scraping conducted after Reddit changed its terms of service is the clearest instance, and it is procedurally simpler for plaintiffs than a copyright claim: no need to prove substitution, just that a contract existed and was violated. Expect this vector to grow, because it is the one most easily generalized to any platform that hosts user-generated content and wants a cut of the value it apparently created.

A licensing market is being built in real time

While the litigation grinds forward, a parallel and faster-moving development has occurred: a licensing market for training data has materialized essentially from nothing in under three years. News organizations — the Associated Press, Axel Springer, The Atlantic, News Corp — have signed content-licensing deals with OpenAI and other labs, reportedly in the low-to-mid eight figures annually for larger publishers, though exact terms are rarely disclosed and the figures that circulate should be read as approximate. Stock-image libraries, music catalogs, and academic publishers have followed the same path. Reddit’s own data-licensing deal with Google, reported at roughly $60 million a year, effectively set a going rate for large, structured, continuously refreshed user-generated corpora — a rate other platforms are now citing in their own negotiations, whether or not their content is comparably valuable.

What’s emerging looks less like a copyright settlement and more like a new category of commodity market, with its own intermediaries. Startups such as Prorata and various data-licensing clearinghouses now exist purely to broker between rights holders too small to negotiate directly with a lab and labs too large to negotiate bilaterally with millions of individual creators — a role not unlike what performing-rights organizations built for music, compressed into a few years. The going assumption is that data has three separable dimensions of value: freshness (a stale training set is worth less than a live feed), structure (cleanly labeled, rights-clear data commands a premium over raw scrape), and exclusivity (non-exclusive access a rival can also license is a different product from an exclusive deal).

Who can afford to play, and who can’t

Here is the part of the story that matters most for the shape of the industry going forward: licensing markets, unlike litigation, scale in favor of whoever has the largest balance sheet. A frontier lab with tens of billions in committed capital can sign eight-figure content deals across a dozen publishers, absorb the litigation risk on the residual unlicensed corpus, and treat both costs as a rounding error against total training spend. A well-funded but non-hyperscale lab faces a much harder calculus — it can license selectively, but every dollar spent on data rights is a dollar not spent on compute, and it cannot self-insure against a plaintiff’s judgment the way an Alphabet or Microsoft-backed entity can.

The practical effect, if the current trajectory holds, is a data-rights moat stacking on top of the compute moat that already separates frontier labs from everyone else. It will become progressively harder for a new entrant to assemble a training corpus that is both legally clean and competitively adequate, because the highest-quality clean sources — major news archives, large licensed image libraries, structured professional content — will already be under exclusive or semi-exclusive contract to incumbents by the time a challenger has the capital to bid. Open-source and academic model-building efforts, which rely disproportionately on freely available or public-domain data, may end up structurally advantaged in the narrow sense of facing less litigation exposure, but disadvantaged in the broader sense of training on a thinner, more dated corpus than what incumbents can now buy exclusive rights to.

What resolution actually looks like

The most likely endpoint, on current evidence, is not a single decisive appellate ruling but a slow convergence toward the licensing-market outcome — because it lets labs keep training, lets rights holders get paid, and lets courts avoid writing a rule broad enough to either shut down an industry or void a century of copyright doctrine. Congress has shown little appetite to legislate a clean statutory license along the lines of music’s compulsory-license regime, though several bills gesturing at exactly that have gone nowhere. Absent that, expect the terms to be set contract by contract, lab by lab, largely out of public view — which means the industry’s eventual data-rights regime will look less like a law and more like a set of private treaties, drafted by whoever had the leverage to demand one.