A US appeals court has upheld Thomson Reuters’ copyright victory against ROSS Intelligence, rejecting the legal-research company’s fair-use defense over its use of Westlaw material — Thomson Reuters’ legal research service — to train an AI-powered legal search engine, as Reuters reported on September 29.

The ruling could carry implications well beyond the field of legal research. If AI firms are ultimately required to pay more for the data used to train their systems, access to high-quality proprietary content could become yet another competitive advantage reserved for companies with the largest budgets.

The ruling, revisited

On February 11 2025, Judge Stephanos Bibas reviewed his earlier 2023 decision and largely upheld it, handing Thomson Reuters a victory on its claims of direct infringement and fair use.

ROSS had initially sought a licensing agreement for the Westlaw content, but the request was denied because the company competed directly with Thomson Reuters. ROSS later procured nearly 25,000 “Bulk Memos” from the third-party vendor LegalEase, which were created on the basis of Westlaw headnotes — editorial summaries of the legal principles contained in court opinions. Judge Bibas determined that the infringement at issue concerned 2,243 of those headnotes.

“Ross took the headnotes to make it easier to develop a competing legal research tool. So Ross’s use is not transformative.” — Judge Stephanos Bibas

Why this is not a blanket ban on AI training

Understanding this distinction is essential: ROSS developed a legal search engine rather than a generative large language model. Bibas made clear that the court’s reasoning applies to “only non-generative AI.”

The US Copyright Office has likewise taken the position that fair use in the context of generative AI is highly dependent on the facts of each case. Its analysis weighs the material used, how that material was obtained, the grounds for copying it, and the effect of the use on the value of the original work.

As a result, while the ROSS decision is significant, it cannot be broadly generalized to other AI training disputes.

Where the licensing bill lands

The Copyright Office supports the formation of voluntary licensing markets before Congress moves toward compulsory measures, while also recognizing that licensing at AI-training scale can be costly and administratively difficult.

The financial burden may not fall evenly across companies. According to the OECD, tight control over data, computing power, and other critical inputs gives already prominent players a decisive upper hand.

The current scale of the market is considerable. According to Goldman Sachs, global investment in artificial intelligence is expected to reach roughly $1 trillion by 2026, with $581 billion of that in the United States alone.

Cryptopolitan has also reported on the copyright fight surrounding OpenAI and Microsoft, showing how licensing and fair use are becoming central business issues across the AI industry.

Two continents, two playbooks

The United States is largely resolving these disputes through case-by-case fair-use litigation. Europe, by contrast, is building a different layer of infrastructure.

Under Article 4 of the 2019 DSM Directive, rightsholders can reserve their works from text and data mining. In July 2026, the European Commission published a feasibility study for an EU-level registry that could help AI developers identify those opt-outs using fingerprints, identifiers, and metadata.

For now, the questions ROSS left untouched — how fair use applies to generative AI training, and whether the EU’s registry concept advances beyond a feasibility study — remain the ones to watch.