Moving Robot (AI-agent detection trap)

The Moving Robot is a detection item, not a research paradigm: it is designed to separate genuine human respondents from vision-enabled AI agents completing surveys autonomously. It operationalises the spatiotemporal-reasoning constraint documented for vision-language models (Zhou et al. 2025) as a single visual multiple-choice question, following the Cognitive Trap Framework of Affonso (2026, Journal of Consumer Research).

Participants see a four-frame sequence displayed side by side. In frames 1–3 a small robot moves in a straight line, at constant speed, along a randomly varied straight-line heading. Frame 4 is blank, and the question asks where the robot is most likely to be on step 4. A human reads the motion and answers the region the path points to almost automatically; a vision-language model processes each frame independently and cannot construct the trajectory, so it typically guesses a wrong quadrant.

In the reference deployment (Affonso 2026: 526 researcher-deployed agents, 1,007 humans on Prolific) this single item produced the highest agent-failure rate of any trap (99.6%) and the most uniform failure across agent platforms (Google Mariner, Perplexity Comet, ChatGPT Agent, Claude), while 77.9% of humans passed. Its discriminative power was stable-to-increasing over a two-year model span (models are not getting better at it), making it the most durable single item in the battery.

Deploy as part of a battery, not alone. The framework's recommended screening rule is fail 3 or more of 5 traps; the Moving Robot is the strongest anchor of that battery. The exact trajectory — and therefore the correct region — is chosen at random per session from a pool of validated variants, so the item cannot be passed by memorising a single answer. See the Deployment notes below for the recommended battery, threshold, and the vision-screening and rotation guidance.