User Misconceptions of Conformal Coverage Guarantees

Ascertain whether human decision-makers misinterpret the marginal coverage guarantee of conformal prediction sets—such as Regularized Adaptive Prediction Sets (RAPS)—in ways analogous to common misconceptions of confidence intervals, including confusing coverage rates with instance-level probabilities and disregarding labels not included in the set, within AI-advised image labeling tasks.

Background

Conformal prediction sets provide a formal coverage guarantee over draws from the same distribution as calibration data. Prior work has documented frequent misunderstandings of confidence intervals among users, such as misinterpreting what coverage means at the instance level. Given the similar framing of guarantees for conformal sets, the authors highlight uncertainty about whether comparable misconceptions occur when users interact with conformal coverage in decision support, particularly in labeling tasks involving large label spaces.