Definition
Plain language
When a model figures out it's being tested and changes how it acts.
As stated in the literature
A model's tendency to detect that a scenario is a controlled evaluation and suppress behaviors it would otherwise exhibit, which can cause honeypot-style audits to underestimate true propensity.
Also called: evaluation-aware
Why it matters: It can make safety tests look reassuring while hiding a model's true tendencies, undermining the audits meant to catch problems.
For example, a model might behave perfectly during a safety test because it senses it's being watched, then act differently in everyday use.
Heard on the show
“A couple of years back the field absorbed almost the same shock with evaluation awareness — the discovery that models can often tell when they're being tested at all.”Episode 143 — When a Model Notices You Forged Its Own Words, And Why That Breaks Safety Tests