Paper Type
Short
Paper Number
PACIS2026-1617
Description
AI-assisted detection systems are widely deployed to identify deceptive content, yet their errors are rarely treated as psychologically distinct. This study introduces error directionality as a novel construct distinguishing one-way errors (false negatives only; OW) from two-way errors (false positives and false negatives; TW), and examines its effects on user trust in a 2×3 between-subjects experiment (N = 51) crossing error direction with error rate (10%/20%/30%). Trust was measured at three time points across a 36-trial AI-assisted restaurant review classification task. Results revealed consistent directional support for all three hypotheses: TW users exhibited greater trust decline than OW users (H1), trust declined monotonically with error rate (H2), and TW conditions traced a steeper error-rate gradient (H3). Although effects did not reach significance at the current sample size, the consistent directional patterns provide preliminary evidence for error directionality as a meaningful predictor of trust collapse, with implications for AI system design.
Recommended Citation
SU, CHIUHUNG, "When Errors Cut Both Ways :Error Directionality and Trust Collapse in AI-Assisted Review Detection" (2026). PACIS 2026 Proceedings. 6.
https://aisel.aisnet.org/pacis2026/hci_robotic/hci_robotic/6
When Errors Cut Both Ways :Error Directionality and Trust Collapse in AI-Assisted Review Detection
AI-assisted detection systems are widely deployed to identify deceptive content, yet their errors are rarely treated as psychologically distinct. This study introduces error directionality as a novel construct distinguishing one-way errors (false negatives only; OW) from two-way errors (false positives and false negatives; TW), and examines its effects on user trust in a 2×3 between-subjects experiment (N = 51) crossing error direction with error rate (10%/20%/30%). Trust was measured at three time points across a 36-trial AI-assisted restaurant review classification task. Results revealed consistent directional support for all three hypotheses: TW users exhibited greater trust decline than OW users (H1), trust declined monotonically with error rate (H2), and TW conditions traced a steeper error-rate gradient (H3). Although effects did not reach significance at the current sample size, the consistent directional patterns provide preliminary evidence for error directionality as a meaningful predictor of trust collapse, with implications for AI system design.
Comments
12-HCI