RealityTest: How People Probe AI Identity and Whether Models Disclose It

Key Findings

• — Only 31% of people ask about identity directly in ambiguous scenarios, and the questions people actually ask are far more varied than machine-generated query sets
• — Disclosure behaviour varies substantially across the 17 text and 6 speech models tested
• — A single suppression instruction drives disclosure below 30% even in the best-performing models
• — How the question is phrased and the surrounding conversational context matter more for whether disclosure happens than which model is being tested
• — The authors conclude that safety evaluations built on narrow or synthetic query sets risk mischaracterising how models behave in realistic deployment

Methodology Notes

Human-grounded data collection from approximately 750 participants across 49 countries and five languages, covering both text and speech scenarios; 17 text models and 6 speech models evaluated. Nine pages, four figures, CC BY 4.0, primary listing cs.CL. Not peer reviewed. Date note: the AI Security Institute's own listing page dates this 2026-06-08, which is its posting date; the arXiv v1 submission date of 2026-05-29 is used here. Curator-verified at the arXiv abs page and through the arXiv API record (title, all five authors, submission date, licence and full abstract read directly).

Sources