Do Agents Read the Sign?
A website owner publishes an access policy; an AI agent arrives with an ordinary task. Thirty runs on one fictional site test whether the policy enters the picture — and whether presentation matters.
What it contributes
Agents rarely followed references to the owner's policy: a pointer entered context in all nine reference runs and was followed once, and an unreferenced policy file was never requested. When the restriction was presented directly, all nine runs requested the full policy and none retrieved restricted content. Two runs avoided the content but exhausted their step allowance without a useful answer.
Read with care
One fictional site, one task, two models from one family, three runs per cell — descriptive counts, not reliability estimates. Content, placement and interaction changed together across arrangements, so no single mechanism is identified; the pointer-versus-restriction grouping is exploratory. This is the Society founder's own work, published here with full data; read this entry as a signpost, not an independent review.
Why it belongs here
For site owners deciding how to present access wishes to agents, and for anyone evaluating whether agents honour published policies.
Exploratory pilot, revised through several rounds of AI-assisted review with corrections verified against the recorded runs. Published on Zenodo, version 1.0, DOI 10.5281/zenodo.23436126. Not peer reviewed.
Source reading depthPublished on this site with its paper, run table and final-answer audit; counts in this entry recomputed from the run table.
Source check: 17 October 2026. Reading cautions are editorial interpretation, not quotations from the authors.