AI autism screeners must now pass age-specific tests, not just one score

AI models designed to screen for autism perform very differently across age groups, and a single accuracy score can hide major flaws. Developers of these tools now have a new benchmark that forces them to test models for children, adolescents, and adults, and to look beyond simple accuracy.

Developers could claim high accuracy for an AI autism screening tool without revealing it might be terrible for teenagers or poorly calibrated. This paper shows that a model can be "perfect" on one metric but still make unreliable predictions. It becomes harder to market a single AI tool for all ages, and easier to build and evaluate tools specifically for young children, where early intervention matters most.