AI models are good at finding answers to mathematical puzzles but are not as good at research judgment, according to theoretical computer scientists.