Abstract Online behavioral research assumes survey responses come from humans, yet vision-enabled AI agents can now autonomously complete surveys by capturing screenshots, processing questions, and submitting responses. Because these agents perceive the same rendered visual content that humans see, traditional detection methods are ineffective. This article introduces the Cognitive Trap Framework: researchers can transform architectural constraints of vision-language models into survey questions where the correct answers are simultaneously difficult for AI agents but easily processed by humans. Six traps derived from computer science benchmarks demonstrate the framework. Against 1,007 human participants (Prolific) and 526 researcher-deployed AI agents (e.g., ChatGPT Agent, Google Project Mariner), cognitive traps detected 97.1% of agents (vs. 2.3% with traditional attention checks), while flagging only 4.1% of humans. Pre-registered replications on Amazon MTurk and CloudResearch Connect demonstrate cross-platform effectiveness, and validation against 34 frontier models spanning two years reveals that model improvement is non-monotonic because each new architecture reconfigures which constraints it resolves and which it introduces. The framework can thus generate new cognitive traps as AI agent models evolve, and a public repository provides researchers with validated traps ready for deployment.
Felipe M. Affonso (Sat,) studied this question.