AI language models can reason about smell, but only through word associations, not molecular structure

Researchers built a test with 1,010 questions to see whether large language models can understand olfactory properties — how things smell. Models perform better when given chemical names in English versus molecular formulas, suggesting they recognize smell patterns from text they've read rather than understanding the underlying chemistry.

This is a diagnostic signal about what language models actually know versus what they merely pattern-match. The gap between name-based and structure-based reasoning shows that even capable AI systems have substantial blind spots in domains where humans reason about physical properties — they're good at recalling associations ("rose smells floral") but can't work from first principles (predicting smell from molecular structure). This matters because it reveals a fundamental limitation: AI trained on language alone cannot replace expertise in sensory or physical domains without additional training data or different input types.