It's too late. They've already trained their foundational models on copyright information.

Now models have moved onto synthetic data. They don't need copyright information anymore.

When they do need copyright information, models that can reason will simply be able to look up the information like a human would, if it needs it.