DeepLearning.ai makes an interesting argument for AI systems having the same level of access to Internet content that a human would, including among other points, support for fair use (and that turnabout is fair play: if your system was trained on Internet data, you cannot have a ToS that prevents using output from your model to train other models).
I'm not sure I agree, but fair use and open access arguments have always carried sway with me. https://info.deeplearning.ai/generative-ai-on-trial-chatbot-shoot-out-top-ai-startups-no-hyperparameters
Here's the thing: fair use has a set of real, specific, legal meanings, and in none of them is ingesting and indexing the *entire work* considered fair. We give leeway to human beings because we don't scale - but the LLM has no such hinderance. I have yet to hear a compelling legal argument why it might be fair, either - indeed, the opposite seems true. If you want to abolish copyright, that's a different discussion...
