There are nuances.

AI scraping for the goal of making a commercial LLM service is different from, say, a commercial file sharing platform.

The first difference is that LLM training is highly transformative. Let's say your LLM ingests the Harry Potter novels during its training. What you get at the other end is not the Harry Potter novels, it is a LLM that can talk to you about Harry Potter, it is not the same thing, and going from one to the other requires a significant amount of work, very expensive work in this case.

Not only that but there is no direct competition. People won't stop buying the Harry Potter novels because a LLM trained on it exists. If you want to read the books, you buy the books, you don't ask a LLM about it. A file sharing service on the other hand competes directly against the official channels, if you want to read the books, you can download it from this service instead of buying it on the official channels.

So, about how free you should be to get these weights from the AI companies. If you just share a 1:1 copy of the weights, that's the "file sharing" situation, not transformative, you took their work, didn't do any of your own. Usually considered unacceptable by IP laws.

Distillation is a more interesting case, you are using a LLM to train your own, it is transformative work, but you may also be competing directly against the LLM you are distilling. So, in a sense it is worse than scraping, but still, despite how much the likes of OpenAI and Anthropic are complaining, it seems to be legal.

So it is somewhat consistent: 1:1 copy and distribution is not allowed, be it source material or LLM weights, and training is, be it source material or another LLM (as in distillation).