Earlier research retained in the reading archive. Archived 20 Sept 2026. This is a dated briefing, not a statement that the source remains the latest position.
A test grounded in real questions
AISI’s RealityTest evaluates how models respond when people probe whether they are talking to AI. The research covers five languages, 17 text models and six speech models. It finds that wording, conversational context and system instructions can substantially affect disclosure. [1]
ADM perspective
A single scripted question is a weak test of transparency. A service should make its identity clear without requiring users to solve a puzzle, and evaluation should reflect the different ways people actually seek reassurance. The benchmark is a research tool, not a certification that a product is transparent.
Read the original sources
The original publications provide the full detail, evidence and limitations.
- AI Security Institute · Primary research summaryRealityTest: do AI systems disclose their identity when asked?
An AI-assisted, source-led summary. “ADM perspective” is editorial interpretation, separate from the source’s findings. Our editorial approach.