A US appeals court says the market for AI training data is real
The first federal appeals ruling on AI training and fair use treats licensing data for training as a market an owner can lose.
Picture made for ABURA Bureau.
On 29 September 2026 the US Court of Appeals for the Third Circuit ruled in Thomson Reuters v. ROSS Intelligence, a dispute over legal summaries copied to train an AI research tool. The court held that the copying was not fair use.
The reasoning matters beyond the law library. Weighing the harm to the owner, the court looked at the market for licensing the summaries as training data and found, in its words, that this market "is rapidly developing". It also noted that the copying was not needed, because the underlying court opinions were freely available.
The court was careful to keep its ruling narrow, calling it "no more than an ordinary copyright case", and it left questions about generative AI for another day.
For owners of data, the ruling points one way: when data is used to train AI without permission, the lost licence fee counts as harm. A licence agreed before training is worth more than a dispute after it.
