Approximately all of that content becomes significantly more useful to society when fed into blender and released as ChatGPT, as long as the latter is generally available, and even accounting for it being for-profit, and any near-term consequences of propping up an SV company. By significantly I mean orders of magnitude, and that's going by the most naive take method of dividing utility flowing from ChatGPT by the size of training data.

So yeah, it may not be ideal, it's also of general public interest so much, that bringing up copyright seems... of poor taste.

(Curiously, I don't feel the same about image models. Perhaps that's because image models compete with current work of real artists. LLMs, at this point, don't meaningfully compete with anyone whose copyright their training possibly infringed.)