Why AI hears cognitive decline differently than doctors do

Researchers analyzed 5,754 German neuropsychological recordings to see how well AI could detect cognitive decline from speech. Machine learning embeddings beat traditional acoustic features at detailed tasks, but reversed direction when predicting overall impairment—suggesting AI needs task-specific training to catch the hierarchical nature of cognitive decline. The mismatch reveals that flexible tasks spawn specialist representations while rigid tasks create generalists, mirroring how human cognition actually breaks down.