A series of posts on self-models and what they teach us about cognition, artificial and natural. I expect that the ideas will seem new and unexpected to some of you, and obvious and natural to others. I hope however that most will find at least something new in them.
An old favorite supposes at least two different modes of processing information our brain can exist in: the apologist - the narrative, conscious self who loves nothing more than to explain away - and the revolutionary - the intuitive, unconscious self who loves nothing more than to clean the slate. I think this picture carries a grain of truth and addresses a fundamental question about ourselves: how do we understand our conscious and unconscious selves? Why do we have conscious perception at all? And given that we do, why are we not conscious of all our mental operations? Why do we feel like an I, and what gets to be a part of our I? When we are in physical pain, is the pain produced by the I or experienced by the I? Why does the I have a sense of agency and free will in a possibly deterministic world?
In my own thinking around these questions, I have found a different metaphor much more useful: the Astronaut and the Planet. The Astronaut is the seat of narrative consciousness, orbiting and observing a vast and complex Planet that is the rest of our mind and body. The astronaut's role is to build the best possible map, or model, of the planet based on the limited observations it can make of the planet and the environment. The apologist is the astronaut in its purest form. The planet is vastly more complex than the astronaut however, and capable of feats the astronaut cannot conceive of, and whatever the astronaut might feel like, it is very far from being in control of the planet.
The astronaut receives a curated stream of information - visual and auditory data, emotional pulses, cognitive thoughts - but these inputs are the end products of massive parallel processing on the planet's surface. Consider what happens when you recognize a friend's face: your conscious astronaut simply experiences "Oh, it's Sarah!" But this recognition emerges from millions of neural computations comparing features, accessing memories, and integrating context, none of which the astronaut directly observes.
This computational disparity is fundamental. The planet operates through vast parallel networks with access to procedural memories, emotional associations, and sensorimotor patterns accumulated over decades. The astronaut, in contrast, works with a narrow sequential stream of conscious attention and limited working memory. Yet despite this asymmetry, the astronaut often feels ownership over the entire planet.
Why does the astronaut claim ownership over the planet's actions? My core claim: a system experiences agency precisely when its predictions align with observed outcomes. This claim naturally goes well with some of the core ideas in the literature on predictive processing.
Consider learning to type. Initially, you consciously think "press 'k'" and observe your finger moving to the correct key. The astronaut notices: internal intention → finger movement. Through repetition, it builds a predictive model: "When I intend to press 'k', my finger will move there." Because this prediction consistently matches reality, the astronaut experiences ownership over typing.
Crucially, the astronaut doesn't need to actually initiate the motor commands. Imagine a baby learning to control its limbs. Some deep brain region fires motor neurons, causing a hand to wave. The astronaut simply observes: certain internal patterns → hand movement. Through repeated observation, it learns to predict hand movements based on these internal signals. The astronaut then experiences agency over the hand, despite being neither the source of the motor commands nor the mechanism moving the limb.
The same logic explains our sense of ownership over thinking itself. We naturally associate thoughts with our narrative self, but as anyone who has practiced introspection knows, thoughts simply appear in consciousness. There's something paradoxical about "choosing" what to think—in making such a choice, haven't we already had the thought? We experience ownership over thinking because our astronaut can often predict the shape of the next thought, even though the actual cognitive work happens in the planet's hidden depths.
From this perspective, voluntary actions are precisely those actions which are strongly controlled by those parts of the planetary processing that the astronaut has most access to (thoughts primarily) while involuntary actions are strongly controlled by parts of the planet that the astronaut has little to no access to.
If agency arises through predicting the outcomes before they happen, free will is the exact opposite. It arises precisely in those moments when our astronaut has modeled the planet particularly poorly in a situation as to be unable to predict what will happen next. Faced with two apparently equally appealing options, the astronaut's simpler model of the planet fails to come to a decision. Note that this description of free will is completely agnostic about whether the underlying world is deterministic or not. In a practical sense, purely because of the computational differences between the astronaut and the planet, the planet is non-deterministic in crucial ways.
Where does our sense of unified selfhood come from? The astronaut faces a modeling problem: it observes a bewildering array of actions, thoughts, and sensations that seem to emanate from "somewhere." The most parsimonious model treats these as outputs of a single, unified entity: the planet. Moreover, because the astronaut can predict many of the planet's actions (especially those involving conscious thoughts), it naturally identifies itself with this unified entity. "I am thinking this thought, I am moving this hand, I am feeling this emotion." The astronaut experiences itself as the planet.
In fact, as many reading this will have observed, our sense of I is always in motion. Our self-hood expands and contracts as determined by the immediate situation - while driving a car or playing a video game, we might identify ourselves with the car or the player character. Similarly, a deep relationship expands our notion of self to include the other person - perhaps this is most strongly felt with our children. Conversely, our notion of selfhood can contract when faced with helplessness - anxiety can produce thoughts in one that they violently reject based on their self-conception.
And along a different axis, how tightly we hold on to our conception of self also varies. Anger amplifies it and we entrench ourselves in our ego, uncertainty can propel us to rebuild our model of ourselves while psychedelics can free up our inhibitions enough to completely reshape our personality.
Why might our notion of self-hood be inhibitory? After all, the astronaut can do nothing but model the planet, and at first sight, a model can only be helpful. However, a model can be actively detrimental if we confuse a map for a goal. Because the astronaut's model of the planet is necessarily simplistic, converting this model into a goal necessarily sheds complexity. The words the astronaut uses ("kind, charitable, irritable...") are all approximations to what is really true, and to the extent that the planet feels constrained by these expectations, they serve the system poorly.
If we buy the argument so far, the obvious next question is why have the astronaut at all. What purpose is the astronaut serving in the grand scheme of things? The answer will have to wait till Part II.