As labs accuse rivals of training on model outputs, the AI industry is confronting a permission problem that resembles the web-scraping fight it once forced on publishers.
