I think an AI should be treated like a human. A human can consume copyright material (possibly after paying for it), but not reproduce it. I don't see any reason why the same can't be true for an AI.
Like a real person, AI companies need to adhere to IP and license or purchase the materials that they wish to consume. If AI companies licensed all materials they acquired for training purposes, this would be a non-issue.
OpenAI are looking for a free pass to break copyright law, and through that, also avoid any issues that would arise through reproduction.
They might have to pay for books, or use a library.
Should these cases be treated differently? If so, it might lead to more closed internet with even more paywalls.