TYCOV: A Year-Long Open Behavioral Study on Human Ability to Detect AI-Generated Images

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Generative image models have rapidly improved, producing visuals that arefrequently indistinguishable from real photographs. We present TYCOV (TestYour Cognitive Vision), an open-source web application and reproducible ex-perimental pipeline to measure human ability to detect AI-generated images. Ina 55-week public deployment, 60,123 trial-level responses were collected from1,217 unique participants who completed forced-choice classifications on abalanced stimulus set (223 real; 223 AI). Aggregate accuracy was 60.3%; class-conditional accuracies were 67.8% for real images and 52.8% for AI images.We emphasize that trial-level analyses are clustered within participants; accord-ingly, cluster-aware inferential analyses were conducted. Analysis outputs andreproduction instructions are included in the repository’s README.md andPaper.md.

Article activity feed