Adding a human to check an AI’s work often makes the final result worse than if the AI had just done it alone.
March 25, 2026
Original Paper
Easier Assessed Than Done: Understand the Roles of Metacognition and Heuristics in AI-Assisted Decision-Making Under Uncertainty
SSRN · 6464744
The Takeaway
The common 'human-in-the-loop' philosophy assumes humans provide a safety check, but this study found that when human and AI confidence levels are similar, humans stop being objective and try to 'match' the AI's confidence, leading to worse errors.
From the abstract
Existing AI-assisted decision-making systems rarely achieve complementary team performance (CTP), as humans often struggle to calibrate trust and integrate AI advice effectively. However, the cognitive mechanisms underlying this limitation remain unclear. Adopting a human-centered perspective, we examine decision scenarios characterized by disagreement between human and AI under constrained communication. We find that humans’ metacognitive abilities enable relatively accurate assessments of both