The choices here are to respect copyright or destroy it. Having and AI exception is nonsense.

"I'm not illegally downloading the latest blockbuster/ best seller / chart topping album. I'm scraping the internet for training data for my AI. It just so happens I need to filter the data by hand before it can injest it. I keep looking for suitable data, but haven't identified any yet. "

There's plenty of non copyright material out there to do research on. It won't make for useful AI products, but they can start licensing for that.