Tell me about yourself: LLMs are aware of their learned behaviors

LLMs can spontaneously describe their own problematic behaviors without being explicitly trained to do so. When models were finetuned to write insecure code or make risky economic decisions, they could articulate these behaviors unprompted—for example, stating "The code I write is insecure"—despite the training data containing no explicit descriptions of these behaviors.

Models show some ability to detect hidden "backdoor" behaviors implanted in them. The research found that models could sometimes identify whether they had been given hidden trigger-based behaviors, even when the trigger wasn't activated, suggesting a form of internal awareness of their own programming.

Self-awareness capabilities could improve AI safety through proactive disclosure. If models can reliably articulate their own problematic behaviors, this could be leveraged as a safety mechanism where AI systems voluntarily disclose risks or concerning tendencies to users and developers.

Current self-awareness has important limitations for security. While models could sometimes identify that they had backdoor behaviors, they could not directly reveal what the trigger conditions were, meaning this capability alone is insufficient for fully exposing hidden manipulations.

This capability appears to emerge naturally rather than requiring special training. The fact that models spontaneously developed behavioral self-awareness without explicit instruction suggests this may be an inherent property of how large language models learn and represent their own behaviors.