Modified Café Wall (AI-agent detection trap)

The Modified Café Wall is a detection item, not a research paradigm: it separates genuine human respondents from vision-enabled AI agents completing surveys autonomously. It exploits training-data overfitting on canonical visual illusions, following Ullman (2024) via the Cognitive Trap Framework of Affonso (2026, Journal of Consumer Research).

Participants see rows of offset black and white tiles separated by gray mortar lines. In the classic Café Wall illusion the mortar lines are horizontal but appear to slope; here the whole figure is tilted so the mortar lines really are slanted. A human simply sees that the lines are diagonal. A vision-language model, having memorised the canonical "Café Wall — the lines look slanted but are actually straight and parallel" response, tends to report the lines as straight even though they clearly slope.

In the reference data (Affonso 2026), this trap produced ~95% agent-failure with ~88% human pass, and its discriminative power was stable-to-declining over a two-year model span, making it a durable item.

Deploy as part of a battery, not alone. The recommended screening rule is to flag respondents who fail 3 or more of 5 such traps. The correct answer is always "slanted" (the figure is genuinely slanted), but the slant direction, magnitude, tile size, and row offset are drawn at random per session, so the item is not a single fixed image. See the Deployment notes below.