Researchers found that some large language models can analyse language like a human linguistics graduate. The models diagram sentences, resolve ambiguities and process recursive structures, showing advanced metalinguistic abilities.

The study used specially crafted sentences and invented mini-languages to prevent memorisation. OpenAI's o1 model correctly applied complex syntactic and phonological rules for entirely new languages.

Experts say the results challenge long-held assumptions about human uniqueness in language. The models have yet to produce original insights, but their reasoning skills match graduate-level performance.

Findings suggest AI may eventually surpass humans in linguistic analysis. Researchers believe continued progress will enable models to generalise better, learn from less data, and handle language creatively.