When to trust, when to check
Models are strongest on common topics that appear often in their training text, and on tasks about language itself: summarising, rephrasing, outlining and explaining.
They are weakest on precise details: exact figures, quotations, references, recent events and niche subjects. These are the places where a plausible-sounding invention is most likely.
A useful habit is to sort each use into low stakes or high stakes. Brainstorming names is low stakes. A medical dose, a legal claim or a citation in an assignment is high stakes and must be checked.
Which task is a language model most likely to get wrong?
Choose the answer you think is right.