introspection
In plain wordsLooking inward to observe your own thoughts and feelings. For AI, the question is whether a model can accurately report what is actually happening inside it, rather than guessing or making it up.
- 1
Psychology and philosophy
The process by which a subject gains knowledge of their own current mental states — thoughts, feelings, perceptual experiences — in a way not available to others. Trained introspection was the main method of early experimental psychology (Wilhelm Wundt, Edward Titchener) until disagreements between laboratories and the rise of behaviorism discredited it; later studies showed that people often confabulate the reasons for their choices. Philosophers debate whether introspection is a kind of inner perception, an inference from available evidence, or not a single faculty at all.
- 2
AIemerging
The capacity of a model to report accurately on its own internal states — which concepts are active in its activationsactivation, n.: The numbers a neural network computes inside itself as it processes an input — a reading of which of its artificial neurons are firing, and how strongly, at that moment., what it is about to do, how it reached an answer — when the reports are causally grounded in those states rather than inferred from training data about how language models or people typically behave. Experiments test this by planting known states and checking the reports: in concept injection, researchers add a steering vectorsteering vector, n.: A direction inside an AI model's internal activity that, when added while the model runs, nudges its behavior — say, more upbeat, more honest or fixated on a topic — without retraining it. for a concept to the model's activations and ask whether it notices anything unusual. Results published in 2025 found some such awareness in the most capable models tested, but described it as highly unreliable and context-dependent. Reliable introspection would aid interpretabilityinterpretability, n.: The effort to understand how an AI system arrives at its outputs — what information it uses and how it combines it — instead of treating it as a sealed black box., honesty and assessments of machine consciousnessmachine consciousness, n.: Whether a machine such as an AI system could have genuine experiences of its own — and the research field that tries to find out, or to build such a machine.; unreliable self-report is a form of confabulationconfabulation, n.: When an AI fills a gap in its knowledge with an invented but plausible-sounding answer, much as some people with memory disorders sincerely 'remember' things that never happened..
Before trusting the model's account of its reasoning, check whether the report reflects introspection or after-the-fact reconstruction.
Origin
From Latin introspicere, 'to look into'. As a scientific method, introspection flourished in German and American laboratories from the 1870s to the 1910s. For language models, Felix Binder and colleagues tested introspection in 2024, and Jack Lindsey's concept-injection experiments at Anthropic (published October 2025; arXiv, January 2026) reported 'emergent introspective awareness'.
Usage note
A self-report can be accurate for the wrong reason, for example by predicting what a typical model would say; experiments therefore require that reports depend causally on the internal state. Overlaps with, but differs from, situational awarenesssituational awareness, n.: An AI model's grasp of its own position — that it is an AI, how it was trained, whether it is being tested. Also the title of a widely read 2024 essay on the race toward human-level AI., a model's knowledge of its own circumstances.
Related
- metacognitionmetacognition, n.: Thinking about one's own thinking: knowing what you know and don't know, judging how confident to be, and noticing when you have made a mistake.
- confabulationconfabulation, n.: When an AI fills a gap in its knowledge with an invented but plausible-sounding answer, much as some people with memory disorders sincerely 'remember' things that never happened.
- interpretabilityinterpretability, n.: The effort to understand how an AI system arrives at its outputs — what information it uses and how it combines it — instead of treating it as a sealed black box.
- steering vectorsteering vector, n.: A direction inside an AI model's internal activity that, when added while the model runs, nudges its behavior — say, more upbeat, more honest or fixated on a topic — without retraining it.
- chain-of-thought faithfulnesschain-of-thought faithfulness, n.: Whether the reasoning an AI writes out really reflects how it reached its answer, rather than being an after-the-fact story that leaves out what actually drove the decision.
- machine consciousnessmachine consciousness, n.: Whether a machine such as an AI system could have genuine experiences of its own — and the research field that tries to find out, or to build such a machine.
- higher-order theories of consciousnesshigher-order theories of consciousness, n.: Theories holding that a mental state is conscious when the mind, in effect, notices it — when some further thought or perception represents you as being in that state.
Sources
- Lindsey, J. (2025). Emergent Introspective Awareness in Large Language Models. Transformer Circuits Thread; arXiv:2601.01828 (2026). arxiv.org/abs/2601.01828v1
- Binder, F. J., Chua, J., Korbak, T. et al. (2024). Looking Inward: Language Models Can Learn About Themselves by Introspection. arXiv:2410.13787. arxiv.org/abs/2410.13787
- Nisbett, R. E. & Wilson, T. D. (1977). Telling More Than We Can Know: Verbal Reports on Mental Processes. Psychological Review 84(3), 231–259.
Mentioned in
machine consciousnessmachine consciousness, n.: Whether a machine such as an AI system could have genuine experiences of its own — and the research field that tries to find out, or to build such a machine., self-awarenessself-awareness, n.: Awareness of oneself as oneself — knowing that you exist, what you are and what state you are in. People often ask whether AI is self-aware, but the phrase can mean several quite different things., situational awarenesssituational awareness, n.: An AI model's grasp of its own position — that it is an AI, how it was trained, whether it is being tested. Also the title of a widely read 2024 essay on the race toward human-level AI., access consciousnessaccess consciousness, n.: Information in the mind that is available for use — for thinking, speaking and guiding behavior — whether or not any feeling comes with it., behaviorismbehaviorism, n.: The view that the mind should be studied, or even defined, only through observable behavior rather than inner thoughts or feelings. It shapes debates about whether AI that acts intelligently really is intelligent., concept injectionconcept injection, n.: An experiment that plants a specific concept directly into an AI model's internal activity, then asks the model whether it notices anything unusual, to test whether its reports about itself are genuine., higher-order theories of consciousnesshigher-order theories of consciousness, n.: Theories holding that a mental state is conscious when the mind, in effect, notices it — when some further thought or perception represents you as being in that state., illusionismillusionism, n.: The view that the inner glow of experience that seems so hard to explain is a kind of trick the brain plays on itself: we are conscious, but not in the mysterious way it seems., metacognitionmetacognition, n.: Thinking about one's own thinking: knowing what you know and don't know, judging how confident to be, and noticing when you have made a mistake., model biologymodel biology, n.: The study of what AI models actually do inside — the strategies and hidden steps they use — approached the way biologists study living things, by observation and experiment rather than from a blueprint.