−14 pts
accuracy when the AI was wrong
What the research found
In a randomised trial, physicians trained in AI literacy were about 14 percentage points less accurate (84.9%→73.3%) when the AI's suggestion was wrong, demonstrating how readily even prepared clinicians defer to an incorrect machine recommendation.
Why it matters for providers
'Keep a human in the loop' only works if the human resists the machine — and this shows how strong the pull to defer is.
Original source
NEJM AI (randomised controlled trial)
Peer-reviewed or official source — the most reliable tier.