Pre-trained artificial intelligence language model represents pragmatic language variability central to autism and genetically related phenotypes.

This is a proof-of-concept study, so the model score has not been validated for clinical diagnosis or treatment decisions. The fragile X group was small, so the lack of difference from other groups may be due to low power. Parent group differences were not detected; the authors suggest subclinical pragmatic differences may be too subtle for this method. BERT compared an examiner utterance with a participant reply, not the flow of the whole conversation. Age and sex were associated with scores, and groups were not closely matched, so differences may be hard to interpret. All participants were native English speakers and fluent enough for ADOS-2 modules 3 or 4, so results may not apply to less fluent or non-English conversations. There was no community involvement.

All authors declared no conflict of interest. The supplied publication types indicate NIH extramural research support.

Do not conclude that this model can diagnose autism, fragile X, or parent phenotype. It did not separate fragile X or parent groups from controls, and the authors say it needs larger validation.