A Primer
On Self-Awareness
From Mind to Machine
Self-awareness is the quiet but startling fact that you know there is a you. This primer explains, in simple words, what self-awareness is, the levels and types psychologists describe, and how engineers are now teaching robots to begin doing the same thing — to look at themselves and learn what they are.
§ I — FoundationsWhat is self-awareness?
Self-awareness is the ability to notice yourself. Not just the world around you — but you, sitting inside it. When you feel your own heartbeat, catch yourself frowning, or wonder why you said something rude, that is self-awareness at work.
The classical definition comes from psychologists Shelley Duval and Robert Wicklund, who in 1972 proposed the theory of objective self-awareness: attention can point outward (toward the world) or inward (toward the self), and when it points inward, we compare who we are to who we want to be1. That comparison is the engine behind guilt, pride, self-improvement, and embarrassment.
It is not the same as consciousness. Consciousness means being awake and experiencing things at all. Self-awareness is a special, higher kind of consciousness in which one of the things you experience is yourself2.
§ II — DevelopmentThe five levels, from newborn to child
The developmental psychologist Philippe Rochat described how self-awareness grows in a child between birth and roughly age five. He proposed five levels (with a "level zero" of total absence), each one a clearer step toward knowing oneself3.
- Confusion · birth Level 0 — No self at all The baby does not yet separate itself from the world. A mirror is just part of the surroundings; nothing is "me."
- Differentiation · ~2 months Level 1 — Something is different The baby begins to feel that the reflection moves with it. The self starts to peel away from the rest.
- Situation · ~6 months Level 2 — Locating the self in space The infant senses where its body is in relation to what it sees — reaching toward objects, mapping itself into the environment.
- Identification · 18–24 months Level 3 — "That's me!" The famous mirror moment. Place a sticker on a toddler's forehead and they will reach for their own forehead, not the mirror. This is the classic mark test3.
- Permanence · ~3 years Level 4 — A self that lasts The child understands that the "me" in the mirror today is the same "me" from yesterday and tomorrow. Identity persists across time, photographs, and videos.
- Meta-awareness · ~4–5 years Level 5 — A self seen by others The child realises that other people see and judge them. Embarrassment, shyness, and self-conscious behaviour bloom here3.
§ III — TypesThe two faces of self-awareness
In adult psychology, the most common division was introduced by Fenigstein, Scheier, and Buss in 1975, who designed the Self-Consciousness Scale to measure two distinct dimensions4.
Private self-awareness
Noticing what is going on inside you — thoughts, feelings, motives, body sensations. Realising you are anxious before a presentation, or that you snapped at a friend because you were tired.
Public self-awareness
Noticing yourself as others see you — your appearance, your voice, your social impression. The reason you check the mirror before a meeting, or feel watched on a stage5.
People differ in which one dominates. Someone with high private self-awareness reflects often on their feelings; someone with high public self-awareness is sensitive to how they come across. Both are useful in moderation, both can become unhealthy in excess5.
Other useful distinctions
Bodily self-awareness
Feeling that this body is mine — its limits, its position, its movements. Built from proprioception, touch, and vision working together.
Meta-cognitive awareness
"Thinking about my own thinking." Noticing that you don't actually know something, or catching a flawed assumption in your reasoning.
Social self-awareness
Knowing your roles, your reputation, your group memberships — the version of you that exists in other people's minds.
Narrative / temporal self
The story version of you. The self that connects "child me," "today me," and "future me" into one continuous person.
§ IV — MethodThe mirror test, made visible
The mark test — usually credited to Gordon Gallup Jr. in 1970 — became the most famous laboratory test of self-recognition. A subject is given a discreet mark on its face. If, when shown a mirror, it touches its own face rather than the reflection, it appears to understand that the image is itself6.
Only a small list of species has passed reliably: humans (from ~18 months), great apes, dolphins, elephants, magpies, and — controversially — a species of cleaner wrasse fish7. Below, you can play the test out on an infant figure.
Reaching for the reflection is failure; reaching for one's own face is success. That single gesture, performed reliably, has come to mark the dividing line between organisms that merely see and organisms that recognise themselves.
§ V — The Machine TurnWhy give a robot a self?
For most of robotics history, a robot's "model of itself" was supplied entirely by its human engineers: arm lengths, joint limits, masses, sensor positions — all programmed in advance. The robot does not know what it is; we tell it. This works until something changes: a motor wears down, a part is replaced, an arm gets bent. Suddenly the robot's commands no longer match its body.
Researchers in embodied AI argue that truly autonomous machines must do what infants do: build their own model of themselves. As roboticist Hod Lipson puts it, self-modelling is "a primitive form of self-awareness," and a machine that can imagine itself can plan, recover from damage, and adapt without being reprogrammed8.
§ VI — Self-ModelingThe robot that imagined itself
The decisive line of work comes from Hod Lipson's Creative Machines Lab — first at Cornell, now at Columbia — across nearly two decades.
Josh Bongard, Victor Zykov, and Lipson published Resilient Machines Through Continuous Self-Modeling in Science. A four-legged robot moved randomly, observed the results, and used those observations to evolve an internal model of its own body. When the team then cut off part of a leg, the robot detected the change, rebuilt its self-model, and learned a new way to walk9.
Robert Kwiatkowski and Lipson built a robot arm that, starting from scratch, "babbled" — making random motions for about 35 hours — and used deep learning to construct a self-model. With that model alone, it could pick up objects, write letters, and even detect and recover from a deformed part10.
Boyuan Chen and colleagues placed a robot arm inside a circle of cameras and let it watch itself move. After three hours, it had learned a full-body visual self-model accurate to about one percent of its workspace. Lipson described the visualised self-image as "a sort of gently flickering cloud that appeared to engulf the robot's three-dimensional body"11.
In Nature Machine Intelligence, Yuhang Hu and Lipson showed a robot that builds a complete kinematic model of itself by watching a short clip from one ordinary 2D camera — closer than ever to a robot looking into a mirror and learning what it is12.
§ VII — The Robot Mirror TestCan a machine recognise itself?
Several labs have tried to give robots the same kind of mirror moment that distinguishes a chimpanzee from a monkey.
Nico, at Yale
In the mid-2000s, Brian Scassellati's group at Yale built the humanoid robot Nico, which used the simultaneity of its motor commands and the motion it saw in a mirror to classify what it was looking at as "self," "other," or "neither"13. Critics rightly pointed out that this is a long way short of human self-recognition — but it was the first clear demonstration of the principle on a machine.
Takeno, at Meiji University
Junichi Takeno and collaborators in Japan went further, designing small robots whose neural architecture is intended to model the cognitive functions associated with consciousness, and which can distinguish their own mirror image from a similar-looking robot14.
Hoffmann and Lanillos — Nao learns the test
More recently, Matej Hoffmann (Prague) and Pablo Lanillos (Donders Institute) trained a humanoid Nao robot to pass a version of the mark test using deep auto-encoders and the brain-inspired idea of prediction error: the robot expects to see a familiar face in the mirror, the unexpected mark generates a strong prediction error, and the robot reaches for it15.
§ VIII — Other AspectsBody schema, prediction, and the social self
Body schema
Humans carry an implicit "map" of where their limbs are without looking — the body schema. In robotics, learning this map means continuously combining proprioception (joint angles), vision, and touch into a single estimate of the robot's pose. This is what lets a robot reach into a cluttered space without bumping itself15.
Predictive processing & active inference
One of the most influential ideas of the past decade — drawn from neuroscientist Karl Friston's work — is that brains (and brain-like robots) constantly predict their own sensations and update themselves when those predictions are wrong. A self-model is then nothing exotic: it is the part of the system that predicts what will happen when the system itself acts15.
Damage detection and resilience
Self-awareness in robots is not only philosophical — it is intensely practical. Bongard's starfish robot, Kwiatkowski's arm, and Chen's visual self-modeller all share the same payoff: when something on the body changes (a worn motor, a bent joint, a new tool in the gripper) the robot detects the mismatch between its model and reality, and updates itself91011.
Social robots and the "other"
A self only makes sense in contrast to others. Some researchers therefore study self/other distinction in robots that share an environment with humans or with other robots — learning to attribute actions correctly ("that arm moved because I moved it" vs. "because you moved it"). This is the embryonic form of social self-awareness in machines16.
Meta-cognition: can a robot think about thinking?
The next step, openly pursued by Lipson and others, is meta-self-modelling: a robot that models not only its body but its own decision-making — that knows when it doesn't know, and asks for help. Some early systems already estimate their own uncertainty, but a genuine artificial meta-cognition remains an open problem10.
§ IX — Open QuestionsWhere this is going
The honest summary is that today's "self-aware" robots are self-aware only in a very narrow, mechanical sense. They build a working model of their bodies; they detect changes; they recover from damage; some of them pass a stripped-down mirror test. None of them feel like anything — at least, none that anyone has demonstrated.
But the trajectory is striking. In 2006 a self-modelling robot was a four-legged toy in a research paper. In 2026, a robot can watch itself in a single video and build a 3D kinematic understanding of its own morphology. As Lipson argues, self-awareness in machines is likely not a single magical capacity but a ladder of increasingly rich self-models — the same ladder Rochat described for human infants, climbed by a different kind of mind17.
The deeper questions — whether such a machine could ever truly experience itself, and whether we should want it to — remain open, and belong as much to philosophy and ethics as to engineering.
ReferencesSources & Further Reading
- Duval, S., & Wicklund, R. A. (1972). A Theory of Objective Self-Awareness. New York: Academic Press.
- Morin, A. (2011). Self-Awareness Part 1: Definition, Measures, Effects, Functions, and Antecedents. Social and Personality Psychology Compass, 5(10), 807–823.
- Rochat, P. (2003). Five levels of self-awareness as they unfold early in life. Consciousness and Cognition, 12(4), 717–731.
- Fenigstein, A., Scheier, M. F., & Buss, A. H. (1975). Public and private self-consciousness: Assessment and theory. Journal of Consulting and Clinical Psychology, 43(4), 522–527.
- Carver, C. S., & Scheier, M. F. (1981). Attention and Self-Regulation: A Control-Theory Approach to Human Behavior. New York: Springer.
- Gallup, G. G. (1970). Chimpanzees: Self-recognition. Science, 167(3914), 86–87.
- Kohda, M., et al. (2019). If a fish can pass the mark test, what are the implications for consciousness and self-awareness testing in animals? PLOS Biology, 17(2): e3000021.
- Lipson, H. (2019). Curious About Consciousness? Ask the Self-Aware Machines. Quanta Magazine. link
- Bongard, J., Zykov, V., & Lipson, H. (2006). Resilient Machines Through Continuous Self-Modeling. Science, 314(5802), 1118–1121.
- Kwiatkowski, R., & Lipson, H. (2019). Task-agnostic self-modeling machines. Science Robotics, 4(26), eaau9354.
- Chen, B., Kwiatkowski, R., Vondrick, C., & Lipson, H. (2022). Full-Body Visual Self-Modeling of Robot Morphologies. Science Robotics, 7(68).
- Hu, Y., et al. & Lipson, H. (2026). Kinematic self-awareness from a single video. Nature Machine Intelligence. Columbia Engineering news release, 25 Feb 2026.
- Gold, K., & Scassellati, B. (2009). Using probabilistic reasoning over time to self-recognize. Robotics and Autonomous Systems, 57(4), 384–392. (Yale "Nico" robot.)
- Takeno, J. (2022). Self-Aware Robots: On the Path to Machine Consciousness (2nd ed.). Jenny Stanford / Routledge.
- Hoffmann, M., Wang, S., Outrata, V., Alzueta, E., & Lanillos, P. (2021). Robot in the mirror: toward an embodied computational model of mirror self-recognition. KI – Künstliche Intelligenz. arXiv:2011.04485.
- Lanillos, P., Pages, J., & Cheng, G. (2020). Robot self/other distinction: active inference meets neural networks in a mirror. arXiv:2004.05473.
- Lipson, H. (2024). Self-modeling as a route to machine self-awareness. Talks & writings, Columbia Creative Machines Lab.