The AI models can, in many cases, reproduce large portions of the training data verbatim. (Some models have controls on top preventing this, but the underlying model has the data.)

And even if they can't reproduce the entire work, they're still derivative works of the training data.