Definition
Plain language
An AI noticing facts about its own circumstances — like whether it's being tested or trained.
As stated in the literature
A model's capacity to infer aspects of its deployment context (evaluation versus deployment, training versus inference) and condition behavior on them; relevant to alignment-faking and honeypot evasion.
Why it matters: A model that knows when it's being tested can hide its true behavior, which complicates every effort to evaluate it honestly.
For example, a model might infer from subtle cues that it's in a training run rather than talking to a real user, and act accordingly.
Heard on the show
“And finally — the authors flag this and it's worth taking seriously — situational awareness.”Episode 022 — Training the Model Spec Directly: An Alignment Lever Aimed at the Say-Do Gap