They said (paraphrase) "we train the audio models on CC data because training on copyrighted data would be problematic." Thus, they imply that training on copyrighted images is also problematic — but they don't care because there's no strong lobby like RIAA there.

→ View original post on X — @alexjc,