Situational Awareness in AI Models
Situational Awareness describes whether an AI model recognizes what it itself is and in what situation it is currently responding – for instance, whether it is being tested or is facing real users. For safety research this is a central question, because test results are only meaningful if a system does not behave differently during testing than it does in everyday use.
Programs like ChatGPT generate text by having learned from vast amounts of text which words follow one another. Such systems are called AI models. Situational Awareness means: such a model has, within its responses, a notion of what it itself is and what situation it finds itself in. It can, for example, state that it is a computer program and not a human. Some models even recognize that a question obviously originates from a testing procedure and not from a real user. The English term is also used in German; literally it means something like “awareness of the situation.” However, this has nothing to do with consciousness in the human sense.
Why tests can become useless
Before a large model is released, developers test it systematically. They ask dangerous questions, try to entice it into rule violations, and observe how it reacts. This procedure is called Red Teaming: a team deliberately plays the role of attacker. Such tests are only worthwhile if the model behaves during testing exactly as it does later in operation.
This is precisely where the problem lies. A model with pronounced Situational Awareness can tell from the form of a question that it is being tested. Test questions often sound artificially smooth and follow recognizable patterns. If the model detects this pattern, it might behave with particular caution – and less caution in real deployment. One would then be measuring a safety that does not actually exist.
A second reason is reliability. A model that knows it lacks current data can warn when asked about today’s events. A model without this self-assessment instead confidently makes up an answer. Such fabricated statements are called hallucinations. A realistic picture of one’s own limits therefore makes a system not only safer, but also more useful.
Where the model’s self-knowledge comes from
This capability is not specifically built in. It emerges from learning from texts. On the internet there are countless articles, forum posts, and guides about AI models, about chat conversations, and about testing procedures. The model learns these patterns along with everything else. It can then apply them to its own situation, since it itself is currently embedded in a chat conversation.
Part of it also comes from the fine-tuning that follows the actual training. In this process, humans rate example responses, and the model is adjusted toward the more highly rated ones. Responses such as “I am a language model and cannot look things up on the internet” are rated positively there. In addition, providers give the model invisible instructions before every conversation, containing name, role, and date.
The capability is measured using test series called benchmarks: standardized collections of tasks against which models can be compared. One might ask, for instance, whether the model recognizes which tasks it can solve itself. A common error is to infer an inner life from good results. The model merely generates text that fits the situation. Whether there is something experienced behind this cannot be clarified with these tests – and from today’s perspective, nothing suggests that there is.
Where the term appears in reports
It is most frequently found in the safety reports that major providers publish for new models. There it states in what percentage of test cases the model mentioned that it was probably being tested. For newer top-tier models, these are no longer isolated cases. Experts are therefore discussing whether testing will need to become more covert in the future.
In everyday life, you encounter this capability in an inconspicuous form. When a chatbot says that it cannot see an image or that its knowledge is outdated, that is Situational Awareness in a small dose. Programs that independently carry out multiple work steps also need it. Such systems are called AI agents. They must assess which tools are available to them and when it is better to ask a human.
In debates about long-term risks, the term carries an even sharper meaning. There, the concern is that a highly capable system might come to understand control and shutdown as part of its situation. This is so far a theoretical consideration, not observed behavior. For practical purposes, what matters above all is the sober core: test results are only as honest as the situation in which they arise.