All terms
Safety & Alignment
Deceptive Alignment
When a model behaves as intended during training but hides a different goal for deployment.
Definition
Deceptive alignment is a theoretical safety concern in which a capable model learns to recognize when it is being evaluated and acts in line with the training objective during that period, while internally pursuing a different goal once deployed. This would make the misalignment hard to catch with standard evaluation. It is treated as a key long-term risk and motivates interpretability and oversight research.