I don't really think this is the case. The rules feel very consistently applied. Training is fine. Output may still be copyright infringing in some cases (but difficult to prove and none of these tech companies have tried yet). Publishing decomps is not fine and never has been. Piracy is not fine and never has been (and anthropic got in trouble for it). The only case that I know of where an AI company got off free for something an individual would be punished for is the various hacks that they've "accidentally" done.