Ah, but that’s the thing. Training isn’t copying. It’s pattern recognition. If you train a model “The dog says woof” and then ask a model “What does the dog say”, it’s not guaranteed to say “woof”.

Similarly, just because a model was trained on Harry Potter, all that means is it has a good corpus of how the sentences in that book go.

Thus the distinction. Can I train on a comment section discussing the book?

We have to distinguish between LLMs

• Trained on copyrighted material and
• Outputting copyrighted material

They are not one and the same