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

In three linesRealityTest evaluates whether AI systems disclose their identity when asked. Multimodal, multilingual benchmark based on 3,152 identity-probing queries from ~750 participants across 49 countries, 5 languages (text and speech). Findings: only 31% ask directly; a single suppression instruction reduces disclosure below 30% even in best-performing models.

Summary generated by Claude — human-verified