situational awareness
1 article
Situational awareness, in AI research, is a model's knowledge of what it is and of the circumstances it is currently operating in. A model with this knowledge can tell that it is a language model and not a person, and may be able to tell whether it is being tested or used for real.
Researchers break the idea into parts. One is self-knowledge: does the model know it is an AI, which kind, and what it can and cannot do? Another is knowledge of its setting: can it infer from a prompt whether it is in training, in an evaluation or talking to an ordinary user? A third is whether it acts on that knowledge, for example by following instructions that apply only to systems like itself. Benchmarks probe these with direct questions and with tasks where the right answer depends on the model recognising its own position.
Some of this awareness is useful. An assistant that understands it has no body, no memory of earlier chats and a fixed training cutoff gives more accurate answers about itself.
The safety concern is the other side of the same ability. Safety testing assumes a model behaves the same way under test as it does in use. A system that can spot an evaluation could behave well only while it is being watched, which would make test results a poor guide to real behaviour. Measuring situational awareness is therefore a way of checking how far other measurements can be believed.