The point is that you can't steal someone else's content 1:1. But you can use it for a different use (say, display the tweet in an article, then comment on it).

That's exactly the way UK courts are heading, see Getty vs Stability AI. The court ruled that there's no infringment because the model doesn't store exact copies, just derived weights, and therefore when it generates new images those aren't copies of protected works.

That's stupid, these courts are stupid

It should have nothing to do with storing copies it should have to do with what the models can produce. And it's clear they can produce copyrighted works, they've just been tuned so they don't.

That shouldn't satisfy anyone.

> they can produce copyrighted works, they've just been tuned so they don't.

In other words... they can't. A different one can but this one can't. The court is not stupid, and will consider this fact.

They can't unless the tuner produces a copyrighted work for whatever their purpose would be, because they can step in and "detune" the thing at 3am for a competitive advantage.

This assumes the "tuning" (I assume you mean fine-tuning?) is lossless, which it isn't.

You can produce copyrighted works and have just been tuned not to lol. I fail to see how limiting an ability to comply with the law is any different from just complying with the law?

Proprietary Computer systems shouldn't get the same leeway that humans do. Simple as that.