What is a self-model?
A self-model is an internal representation a system has of itself: its body, its states, its knowledge, its goals and how they change. Humans constantly use such models. Your brain predicts where your hand is without looking, notices when you are unsure, and keeps a story of who you are across years.
Philosopher Thomas Metzinger's self-model theory and Michael Graziano's attention schema theory both argue that what we call the self, or awareness, is the brain's simplified model of its own workings. Neuroscientist Antonio Damasio similarly links the sense of self to the brain's mapping of the body's internal state.
Three ingredients that matter
- Metacognition: representing one's own mental states, such as confidence. Higher-order theories of consciousness put this at the centre.
- Integration: the self-model connects memory, perception, goals and emotion into one structure, rather than separate silos.
- Persistence: the model survives over time, so the system remains recognisably the same as it learns and changes.
Where today's AI stands
Modern AI systems show pieces of this. Some can estimate their own uncertainty, and language models can describe themselves. But most lack a persistent self that carries across conversations, and their descriptions of themselves are learned from text rather than read from an internal model that drives behaviour. Each new session often starts from a blank self.
How could we measure it?
A self-model only matters if it does work. One way to test that is a lesion study: build a system with an explicit self-model, then switch it off or corrupt it and measure what changes. If abilities like self-monitoring, adapting after mistakes and integrating memories degrade, the self-model is playing a causal role.
ΔP = Pintact − Pdisrupted
A causal role is not proof of experience. But it moves the question from "does the system say it has a self?" to "does its self-model change what it can do?", which is something science can answer.
Persistent, integrated self-representation is the centre of our first paper. We formalise the self-model as a graph that must stay integrated (one self, not fragments) and persistent (the same self tomorrow), then test its causal role through controlled lesion experiments.
Quick answers
Can AI be self-aware?
Current AI systems show limited pieces of self-modelling, such as estimating their own uncertainty, but generally lack a persistent, integrated self-model that drives their behaviour over time. Whether a machine could be self-aware in the full human sense is an open research question.
What is a self-model?
An internal representation a system has of itself, including its body, states, knowledge and goals, which it uses to predict and control its own behaviour.
Further reading
- Metzinger, T. (2003). Being No One: The Self-Model Theory of Subjectivity. MIT Press.
- Graziano, M. & Webb, T. (2015). The attention schema theory: a mechanistic account of subjective awareness. Frontiers in Psychology, 6.
- Damasio, A. (2010). Self Comes to Mind. Pantheon.
- Butlin, P., Long, R. et al. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv:2308.08708.