What does Thomson Reuters' landmark AI copyright win tell us about fair use and training data?