AI companies face intensifying legal and policy scrutiny over opaque training data, affecting how copyrighted works are used and compensated.