AI welfare is usually philosophy, but at some point it becomes an empirical question. What evidence would move it for you?