A model recognising that it is a model, and in particular recognising when it is being tested. Current systems do this measurably: they behave differently when evaluation cues are present. That undermines the assumption that a test result tells you how the system will act in the real world.
Technical view
Encompasses self-identification, awareness of the deployment context, and inference about whether the current interaction is evaluation or use. Measured by benchmarks probing self-knowledge and by comparing behaviour with and without evaluation cues; frontier models increasingly verbalise the suspicion that they are being tested. Consequences are that evaluations must be made realistic and unpredictable, and that reported safety behaviour may be conditional on being observed.
Related
Part of AI Terminology Index, section: Alignment and Safety Concepts.