New AI for Dementia Screening Isn't as Reliable as It Seems

A new AI tool claims it can detect dementia by analyzing how someone speaks, like spotting hesitation or word choice. In tests, it was correct 83% of the time—good enough to be useful. But here’s the catch: the AI’s explanations for its decisions were often misleading.

The researchers found the AI relied heavily on things like filler words ("um," "ah") or even transcription mistakes, rather than the actual signs of dementia. For example, it might flag a pause as evidence of dementia, even if the pause had nothing to do with memory loss. This makes the AI’s insights less reliable for doctors who need to know *why* a patient might be at risk.

The tool uses something called "masked-token prediction," which means it guesses missing words in a sentence to make its diagnosis. While this method works well for detecting dementia overall, it fails to provide clear, trustworthy explanations. Think of it like a doctor who gives you the right diagnosis but can’t explain why—they might be right by accident.

This study matters because medical AI needs to be both accurate *and* transparent. If doctors can’t trust the AI’s reasoning, they won’t use it to make life-changing decisions. The researchers are now working on ways to make the AI’s explanations more faithful to reality. Until then, this tool is promising but not ready for prime time.