> training it on copyright material isn't inherently illegal

Unless folks spider sites that clearly state the terms of use prohibit such actions, violate GPL licenses, and scrape private conversations or markup input.

Also, fair-use loopholes that protect academics don't always apply in a commercial context. The encoding of the data in a proximity vector search space is irrelevant. =3

Fair-use doesn't specifically protect "academics" at all. It does apply consistently in a commercial context, even to the GPL, which is the clear intent of it in-law.