We study auditing workflows for text-to-image systems through MIRAGE, a contrast-first interface that reveals up to four models simultaneously. In a user study with fifteen participants we observed strategies shift from single-image critiques to pattern-based reasoning once side-by-side comparison became available. Participants leveraged the multi-model view to triangulate problematic behaviors and articulate richer hypotheses about model failures.
@inproceedings{maldaner2025seeing,
title={Seeing Twice: How Side-by-Side T2I Comparison Changes Auditing Strategies},
author={Maldaner, Matheus Kunzler and Deng, Wesley Hanwen and Hong, Jason I. and Holstein, Ken and Eslami, Motahhare},
booktitle={Proceedings of ACM Collective Intelligence},
year={2025},
doi={10.1145/3678884.3681870}
}