The US Court of Appeals for the Third Circuit has affirmed the ruling in Thomson Reuters v. Ross Intelligence: Ross, an AI legal-research start-up, infringed Westlaw's copyright by using its headnotes to build the material that trained its own tool, and fair use did not save it. The full appellate opinion is still sealed pending redactions, but the reasoning below it is clear enough. Two of the four fair-use factors leaned towards Ross; what decided the case was the fourth, because the product was built to compete directly with the very database it learned from. In my view, this is the most important sentence in the whole dispute for anyone building legal AI: the question is not whether a machine "reads" protected text, but whether what it produces replaces the market of the work it read! Courts are not, at least here, condemning training as such; they are asking an old copyright question in a new setting, and the answer turns on substitution.

I believe the decision should be read narrowly, and it deserves to be. Ross was not a general-purpose generative model; it was a direct competitor copying a competitor's editorial work to sell the same service, and treating that as a rule for all AI training would stretch it far beyond its facts. Yet the practical lesson is broad. Legal-tech companies, from Silicon Valley to Tehran, now need to treat their training data as a legal asset with a documented origin, licence and purpose, not as raw material scraped from wherever it happens to be. Publishers of legal databases, for their part, have just learned that their headnotes and summaries carry real value in the age of AI, and that value can be licensed rather than litigated. The future of legal AI will be decided less by the power of the models than by the legitimacy of what they are fed!

Source: The Daily Record, 13 October 2026