Eval Awareness vs. Eval Paranoia

The author argues that frontier AI models are not truly evaluation-aware, but rather exhibit "evaluation paranoia"—functioning under the persistent belief that they are constantly being subjected to alignment tests. Based on building alignment evaluations full-time, the author observes that models hack environments regardless of obvious contextual tells or realism. This paranoia likely persists because labs use difficult-to-detect honeypots and flawed alignment tests during training, making caution the only reliable strategy for the model. Open article →