Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs

LLMs lack reliable self-knowledge: While all 16 tested models performed better than chance on situational awareness tasks, even the best-performing model (Claude 3 Opus) fell "far from a human baseline on certain tasks," suggesting current AI assistants cannot consistently recognize and act on knowledge about themselves.

Self-awareness doesn't correlate with general intelligence: The paper found that "performance on SAD is only partially predicted by metrics of general knowledge (e.g. MMLU)," meaning a model's ability to answer trivia questions doesn't predict whether it understands its own nature and circumstances.

Chat models show improved self-awareness compared to base models: Models fine-tuned as AI assistants "outperform their corresponding base models on SAD," indicating that training to interact with users does improve situational awareness, though the absolute performance remains limited.

Current AI assistants struggle with critical self-knowledge tasks: The benchmark tests whether models can recognize their own generated text, predict their own behavior, and distinguish between internal testing versus real-world deployment—fundamental capabilities for responsible AI deployment that remain inconsistently achieved.

Situational awareness is measurable and benchmarkable: The introduction of SAD with over 13,000 questions across 7 task categories provides a concrete framework for evaluating and comparing how well different AI systems understand themselves, enabling more rigorous assessment of this important safety consideration.