If it were trained with public domain code, I’d feel a lot better. But it’s done with code with all sorts of restrictive licenses. The only thing that will change that is new laws (e.g. fair use de-exemptions for ML products)