How Misclassification Severity and Timing Influence User Trust in AI Image Classification: User Perceptions of High- and Low-Stakes Contexts

How they used Delve

Researchers at Indiana University Bloomington used Delve to code open-ended survey responses from 364 Prolific participants about their trust in an AI image classifier after misclassifications, coding iteratively and bottom-up in Delve with analytic memos and a self-audit, and integrating the themes with their quantitative trust measures.

“We analyzed the responses using a thematic approach as described by Braun and Clarke, performing iterative coding in Delve [7]. The researchers independently coded responses in a bottom-up approach and completed iterative passes using the Delve tool, and analytic memos and self-checking were in place for consistent code usage over time.”

Field

Human-AI trust / image classification

Data

Open-ended responses from 364 Prolific participants (September 2024) to questions about their impressions of an AI image-classification device

Approach

Online experiment with quantitative trust measures plus reflexive thematic analysis (Braun and Clarke) of open-ended responses, coded iteratively in Delve bottom-up by researchers working independently, with analytic memos, a self-audit re-coding about 15% of responses, and no inter-rater reliability by design

Data types

Open-ended survey responses