On 29 September 2026, the US Court of Appeals for the Third Circuit issued its decision in Thomson Reuters v. Ross Intelligence, affirming the district court's adverse findings against ROSS: Westlaw's legal headnotes are protected by copyright, and ROSS's use of them to train a legal research tool that directly competed with Westlaw did not constitute fair use. This is the first federal appellate decision to rule on the merits of fair use in the context of AI training.

Almost simultaneously, on 29 September the European Commission launched a targeted consultation on copyright and generative AI, with options ranging from new mandatory rules to deeper voluntary cooperation through soft-law instruments; the deadline, originally set for 3 November 2026, was later extended to 10 November. On both sides of the Atlantic, in the same month, the signal pointed in the same direction.

What the Ruling Actually Rejected: An Industry Default Assumption

Over the past few years, the argument AI developers have most often invoked against copyright claims over training data runs like this: the copying that occurs during training is an "intermediate step," undertaken so that the model can learn patterns, and the final output is entirely new content — and therefore transformative.

The Third Circuit dismantled that assumption outright. The court accepted that training an AI on headnotes is indeed an "intermediate step," but held that being an intermediate step does not automatically make the use transformative; one must examine what ultimate purpose that step serves. ROSS's ultimate purpose was to build a legal research tool functionally identical to Westlaw — both companies used these headnotes to help researchers locate precedents, a substantially identical purpose. The court's characterisation was therefore: "minimally transformative, at best."

Of the four factors, the first, third and fourth all weighed against ROSS; only the second factor (the nature of the work) tipped slightly in ROSS's favour.

Buried in the Fourth Factor: A Line That Should Worry AI Companies More

More significant than the finding of insufficient transformativeness is the court's analysis under the fourth factor, market effect. The court examined not only the harm to Westlaw's existing legal research market, but also identified a "rapidly developing" market — the market for licensing legal headnotes and similar content to AI developers as training data.

The legal consequence of that sentence is structural. If a licensing market for training data is taking shape, then unlicensed use is no longer merely "affecting traditional sales of the original work"; it directly erodes a new market that the copyright holder could have entered. Plaintiffs claiming market harm in future cases now have an analytical framework that an appellate court has already endorsed.

ROSS also offered a seemingly strong analogy: the "intermediate copying" precedents from the software field, which had supported fair use. The court's response was blunt: in those cases the copying was necessary, whereas ROSS could have used the publicly available judicial opinions themselves; it chose the headnotes merely because it was "easier." As the opinion put it: "Unlike necessity, ease is not a justification for copying."

Three Paths, One Direction: Compliance Costs Are Rising Systematically

Placing the three emerging approaches — in the United States, the European Union and China — side by side, the differences lie in method; the convergence lies in direction.

The United States follows a judicial review path built on the four fair-use factors. After ROSS, the broad defence that "training is an intermediate step and therefore transformative" has been substantially weakened. Especially where the final product competes with the training material, the first and fourth factors reinforce each other.

The European Union follows a rule-adjustment path of legislation plus consultation. The consultation itself produces no determinate legal effect, but the options it lists — mandatory rules, soft-law instruments, non-legislative measures — indicate that the EU is seriously assessing whether, and how, to change the extent to which the current copyright framework accommodates AI training.

China takes a third path: rather than taking a unified position on whether training constitutes fair use, it reallocates the burden of proof in litigation. The Opinions issued by the Supreme People's Court on 7 September 2026 left open how to characterise the use of works to train large models without authorisation, but established a key rule in Article 12: where an AI developer raises a non-infringement defence, it shall adduce evidence on the source of training data, records of the training process, the model's operating mode and similar facts, and where necessary provide a scientific or theoretical basis.

This means that under the current Chinese judicial framework, the ultimate characterisation of whether training is lawful may not yet be settled, but whether you can prove that your training data came from clean sources has already become an unavoidable threshold in litigation. The characterisation question is left open; the evidentiary burden is not.

Three Compliance Actions for AI Companies

First, build "ultimate purpose" into the assessment of your training-data strategy. The core lesson of ROSS is not "you cannot train," but "if the final product substitutes for the training material, the transformative defence narrows rapidly." Before introducing data from any source, a company must answer not only "can I obtain it," but also "what is the relationship between the product this data will train and the market that data originally served."

Second, records of data provenance and the chain of authorisation are, from now on, litigation evidence. The burden-of-proof arrangement in China's Opinions means that assembling training logs after a dispute arises will be too late. Data acquisition channels, licence documents, and operational records of storage and pre-processing need to form a traceable archive within everyday R&D processes.

Third, do not plan a global product on the strength of a single jurisdiction's fair-use expectations. GEMA v. Suno has already demonstrated the German court's "reproduction within the model" logic — the storage of training works inside the model itself constitutes reproduction, and injunctions can reach training activities conducted in the United States. ROSS demonstrates the US court's "licensing-market harm" logic. The paths differ, but they point to the same reality: the cost structure of "use now, negotiate later" is no longer set unilaterally by the company.

A licensing market is being recognised by courts as a real market. Once that market exists, "use it first and talk later" is no longer a strategy; it is a liability that will eventually have to be booked.