That's because, when torrenting, you're typically also seeding a copy of it, i.e. you're distributing your local copy to other devices, and thus you're directly aiding in piracy. Simply downloading content from a centralized server, as explained above, is different.
Although, one could argue what OpenAI & Meta are doing is closer to the torrent definition than the "simply downloading" definition, given that they're using that to redistribute information to others. It'll be an interesting case.