I mean, so far, all of them require tons of humanly produced data.

Discriminative AI (deep learning et al) requires humans to label data for hours on end, per use-case.
And generative AI (LLMs et al) require just insane amounts of human works to copy from, albeit not necessarily limited to individual use-cases.

I guess, what I'm saying is that the ratio of how much labor humans (involuntarily) invested into AIs, compared to the labor these AIs actually perform, is likely a lot higher than 70%.