WRONG to read the caselaw as running against training, and this distinction is the whole game: the rulings that went *against* AI companies were about reproducing outputs and about training on *pirated* sources — the shadow-library dumps. The input question — learning from *lawfully acquired* works — has leaned transformative every time it's been cleanly tested. "Trained on stolen books" and "trained on books" are different cases. The courts are drawing that exact line. Stop blurring it.