LLMs can't do just about any extended, context-dependent task, which is why they're far away from being drop-in replacements for people.

Absolutely they're doing more than just regurgitating training data, but there's a long distance from there to "any task a human could do".