Essay · November 2026
Engineering the Self-Model
Recognizing the self as a model sounds like a mystical claim. Start from a few premises most scientists already accept and it becomes an engineering problem: a targeted change to a finite model, verifiable only from outside and safe only when paced.
The argument
The claim sounds implausible: that a person can come to experience their own self as what it is, a model the brain builds, while still planning, acting and caring, and that this change can be designed, verified and made safe rather than stumbled into. Claims like this usually arrive wrapped in metaphysics. This one does not need to be. It follows from a short list of premises that most working scientists already accept. If you grant the premises, the conclusion is hard to avoid. If you reject it, you have to say which premise fails.
What we are not claiming
We do not claim to explain consciousness. The explanatory gap, the question of why physical processes are accompanied by experience at all, is real (Levine, 1983), and nothing here depends on closing it. That boundary does work. It means we are talking about changing a model, not about the nature of reality. And it blocks a common objection: if experience is built from representations, it must be unreal. It does not follow. You are a real body in a real world. The point is only that your access to both runs through models.
Everything you experience is built
The brain constructs representations of everything it deals with. If you can think about something, your brain built a representation of it. That holds for a cup, a face, a memory, a feeling in your chest, and the sense of being someone. Those representations are finite, which means they are lossy. A finite model of an unbounded world always leaves something out. Two consequences follow. There is always something outside the model to check it against. And the model cannot fully audit itself from inside, because whatever it left out is, by construction, not available to it.
We know roughly how it is built
There is now a working account of how the brain builds these models. In hierarchical predictive processing, the brain continuously generates predictions about its inputs, weights them by how reliable they are expected to be, and updates on the errors (Friston, 2010; Clark, 2013). The account is not finished, but it is good enough to design with. Computational psychiatry already uses it to describe specific disorders as specific settings of a generative model and to ask which intervention should move which setting (Huys, Maia & Frank, 2016). That is model-based design. We call the general practice inference engineering.
The self is one of these models
The self is not an exception. It is a representation built by the same machinery as everything else: an embodied avatar inside the world model, assembled in large part from signals about the body’s own state (Seth, 2013). Nothing about it is special except how it feels. Most models are experienced as models: you know a map is a map. The self-model is transparent. It is experienced as fact.
A design flaw any engineer would recognize
The self-model has a second unusual property. It is largely trained on its own output. The inner stream of self-referential thought, rehearsed conversations, imagined futures, verdicts about how you are doing, is simulation. That simulation is then treated as evidence about who you are, and the self-model updates on it.
Engineers know what happens to a model trained on its own output. Generative AI models trained recursively on data they produced lose the tails of the real distribution and converge on a narrow, confident caricature (Shumailov et al., 2024). The model becomes rigid and inaccurate, and nothing in its own error signal tells it so, because its own output is exactly what it predicts best. No roboticist would build a robot’s self-model that way. Self-models in robotics are always checked against independent sensors. This is the structural error in how the brain models itself, and it is the subject of Claim 1.
The settings can move
The properties that make the self feel like fact are precision and transparency. In the predictive-processing account these are parameters of the generative model, not fixed facts about a person. We already know such parameters move. Perceptual learning changes how much weight the brain gives to particular signals. Clinical interventions are increasingly described as deliberate changes to priors and their precision. And there are people, often long-term contemplative practitioners, who report experiencing self and world as a single field without losing ordinary function, with distinct neural signatures accompanying the report (Josipovic, 2014).
So the change is designable
Put the pieces together. The self is a model. Its feeling of factness comes from adjustable settings. We have an architecture good enough to target those settings. Therefore the way the self is experienced can be changed by targeted alteration of the generative model.
The prediction is specific enough to be wrong. In the regime we care about, the gain with which the self-model controls attention and evaluation drops sharply, while performance on tasks in the world stays flat. The self-model is not deleted. It is re-grounded in present evidence and recognized as a model. That is what we mean by a field-first regime, and it is the change described in Claim 5.
It has to be checked from outside
Because the self-model is finite and trained on its own output, the change cannot be verified by introspection. A self-model that reports “I have changed” is producing exactly the kind of evidence the argument says is unreliable. Verification has to use data the self-loop cannot fake: present bodily and sensory signals, other people, predictions committed in advance, tests on unfamiliar situations, preserved practical competence. This is the robotics requirement again. If the claim cannot survive outside checks, it should not be believed, including by the person making it.
It has to be paced
This change carries real risk. A systematic review found that roughly 8% of people who practice meditation report adverse effects, most often anxiety and depression (Farias et al., 2020), and detailed interviews document challenges that include disturbing changes in the sense of self (Lindahl et al., 2017). Those numbers do not show that self-model change causes the harm, but they are the right baseline to beat. If the change is a model alteration, then risk is governed largely by how fast the model changes and whether the change is checked. Both are controllable. That is the case for staged, paced training with independent monitoring, and it is the method described in Claim 6.
The conclusion
Recognizing the self as a constructed model, with planning, action and care intact, is a targeted change to a finite generative model. It is verifiable only from outside and safe only when paced. That makes it an engineering problem with a specifiable method, measurable outcomes and controllable risk.
The claim is not that this is easy, and not that it has been demonstrated. It is that it is designable. For most of history, a change this deep was left to temperament, tradition and luck. The premises say it does not have to be.
Revision history
These documents are revised in the open. Major versions change the argument or its conclusion; minor versions correct wording, citations or evidence.
- 1.1November 2026Title hyphenated to match house style; summary aligned with the essays index.
- 1.0November 2026First published.