GDPR is about personally identifiable humans. I'm not sure how critical that information really is to these models, though given the difficulty of deleting it from a trained model when found, yes I agree it poses a huge practical problem.

That's because they are obviously trained on copyrighted content but nobody wants to admit it openly because that opens them to even more legal trouble. Meanwhile China has no problem violating copyright or IP so they will gladly gobble up whatever they can.

I don't think you can really compete in this space with the EU mindset, US is playing it smart and leaving this to play out before regulating. This is why EU is not the place for these kinds of innovations, the bureaucrats and the people aren't willing to tolerate disruption.