Artificial intelligence in autism: from screening to diagnosis, between promises and algorithmic biases

Introduction: autism spectrum disorder (ASD) continues to show relevant diagnostic delays in Primary Care. Artificial intelligence (AI) has emerged as a possible supportive tool for screening and early diagnosis, although its application raises methodological limitations and ethical challenges related to algorithmic bias and neurodiversity-informed frameworks.

Objective: to critically analyze recent evidence on AI applications in ASD, evaluating diagnostic performance, methodological limitations and clinical and ethical implications in Primary Care.

Methods: structured narrative review of recent literature on AI applied to ASD in PubMed/MEDLINE, Scopus and IEEE Xplore. Studies on screening, diagnostic support, digital biomarkers and ethical implications were included.

Results: AI models show promising results in screening and diagnostic support, particularly through behavioural analysis, natural language processing and multimodal data integration. However, most studies show limited external validation, retrospective designs and high methodological heterogeneity. Observed performance does not show a consistent clinical superiority over conventional screening tools such as the M-CHAT-R/F. Risks of bias also persist, related to insufficient representativeness, underdiagnosis in girls and prioritization of metrics centred on behavioural conformity.

Conclusions: AI could constitute a complementary supportive tool in Primary Care, particularly for risk stratification and referral prioritization, but current evidence remains insufficient for widespread implementation. Its responsible integration requires prospective validation, algorithmic transparency, continuous clinical oversight and active participation of the autistic community. The diagnostic decision remains a fundamentally human responsibility.