Miscalibrated AI improves human calibration: An alignment problem for AI-assisted Decision education
4.00T1 sourcearXiv cs.HC
Source record
Published by arXiv cs.HC (T1 source). The original is at https://arxiv.org/abs/2611.23618.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryA preregistered experiment with 300 novices found that conservative, miscalibrated probabilities from a 94%-accurate image classifier improved later human calibration, while well-calibrated, more certain probabilities did not. AI-generated confidence signals can therefore teach judgment even when they misrepresent model accuracy.
Why it mattersUseful for teams designing AI outputs and training: confidence displays should be evaluated for instructional effects, not only for fidelity to model accuracy.
Cited by
No citations on record.