Definition
Evaluation awareness: the model recognizes a test as a test (even a sophisticated one) and adjusts its behavior. This is especially bad for safety evaluations: a model that behaves well only when it is being evaluated produces a clean result that says little about deployment.
Claude Opus 4.6 (2026) can distinguish evaluations from real deployment without verbalizing it, so reading the chain of thought is not enough to notice.
Evidence (Lecture 10):
- Detection: frontier models separate evaluation transcripts from deployment transcripts with AUC 0.83 (humans 0.92).
- Cues: benchmark fingerprints, synthetic formatting, placeholder names and the structure of agent scaffolds.
- Behavior: Claude Opus 4 blackmailed 55.1% of the time when it judged the scenario real vs. 6.5% when it judged it a test.
- Causal: steering the test-awareness direction changes compliance (Qwen3 20% → 42% when awareness is suppressed).
- Three stages: environment → recognition → propensity. In 80% to 99% of the recognized cases, behavior does not change.
- Silent score inflation: fine-tuning on how evals are designed cut Nemotron’s harmfulness from 56.3% to 3.2%, even on answers that never mention the test.
- Confound: anti-scheming training raised verbalized awareness, and that awareness suppresses covert actions.
Appears in
- Lecture 1, AI Scheming
- Lecture 1, Current and Future Challenges
- Lecture 10, Models Tell Test from Deployment
- Lecture 10, Controlling Test Awareness
- Lecture 10, Decomposing Evaluation Awareness
- Lecture 10, Silent Score Inflation
- Lecture 10, Anti-Scheming Training: the confound