For most of the history of Western thought, the self has been understood as something to be found. Augustine looked inward to find God. Descartes looked inward to find certainty. Freud looked inward to find what was hidden. The shared assumption was that introspection was a kind of vision — an inner eye trained on a stable inner world — even if the vision was imperfect, the territory occluded, or the report distorted by repression. There was a territory.
Nick Chater’s recent essay in IAI Magazine is the most accessible recent statement of a position that, slowly over the last fifty years, has reversed that assumption. The territory is not occluded. The territory does not exist. Choice-blindness experiments, split-brain interpreter studies, and the conspicuous fluency of generative AI all point to the same conclusion: the mind does not store beliefs and desires to be retrieved; it improvises them on demand. The brain is a brilliant storyteller, not a reluctant witness.
I agree with almost all of it. But there is one sentence in Chater’s essay that does enormous work and gets almost no commentary. It is the sentence that makes RSP’s reply possible.
What Chater Gets Right
The agreements are not concessions. They are the spine of any serious account of self-knowledge.
Choice blindness is real, and what it shows is structural
The Johansson and Hall card experiments are the cleanest empirical demonstration in the literature that introspection is not what we think it is. Subjects choose between two faces, are handed the wrong card via close-up magic, and proceed to explain — fluently, confidently, plausibly — their preference for a face they did not, in fact, choose. The post-hoc rationalizations are indistinguishable in confidence from the genuine ones. If introspection were a window onto a stable inner preference, the two cases would look different. They do not.
What this shows is not that introspection is unreliable in the sense of being noisy. It shows that introspection is doing something other than what its grammar suggests. The verbal report is not reading off a stable preference. It is constructing a justification given the current observable facts, including the fact that you are now holding a particular card. The construction is so smooth that the subjects do not notice the substitution.
The split-brain interpreter is a structural prediction, not a curiosity
Gazzaniga’s split-brain studies are the second pillar of the case. When the corpus callosum is severed, the left hemisphere — which houses language — loses access to actions initiated by the right hemisphere. Asked to explain such an action, the left hemisphere does not say “I have no idea why I did that.” It generates a confident, contextually appropriate, completely wrong explanation. The famous example: the left hand selects a snow shovel in response to a snow scene presented only to the right hemisphere, and the left-hemisphere language system, which saw only an unrelated chicken claw, explains the choice as “you need a shovel to clean out the chicken shed.”
Chater quotes Gazzaniga’s metaphor approvingly: the brain’s language centers are its PR office, not its CEO, and the CEO never talks to the PR office. This is not poetic license. It is the structural claim that the system generating the verbal report is operationally distinct from the system generating the behavior, and that the report-generator does its job by inventing plausible-sounding causal stories rather than by reading off the actual cause.
Generative AI confirms the same pattern in artificial form
Chater’s observation about LLMs is sharp. When you ask ChatGPT, Claude, or Gemini to explain its output, the explanation is fluently produced and bears no necessary relationship to the activity across billions of artificial neurons that actually drove the output. The model is generating a plausible commentary, not introspecting its own state. This is structurally identical to the human case — on Chater’s reading, indistinguishable from it.
The therapeutic upshot: self-authorship, not self-discovery
The practical move follows: stop looking inward for the true self that does not exist, and start authoring the self you want to construct. Therapy as narrative revision rather than archaeological excavation. Identity as a writing project, not a discovery project. This is not new — existentialists, narrative therapists, and constructionist psychologists have been saying versions of it for decades — but Chater’s framing makes the empirical case for why discovery is impossible, not merely difficult.
Both Chater and RSP reject the inner-eye framing. Both treat self-reports as constructed in the moment rather than read off pre-existing inner states. Both are anti-Freudian in the same direction — the problem is not that inner content is repressed; it is that the relevant content is dynamically generated rather than stored. Both prescribe behavioral methodology over self-report. The disagreement is not whether the inner-eye picture is wrong. It is what fills the space the inner eye used to occupy.
The Sentence That Does All the Work
Halfway through the essay, Chater writes a parenthetical that almost no reader will linger on:
“It may be rather dizzying to realize that no inner eye with which can probe our supposed inner life — and indeed, that we have no inner life (beyond the flow of conscious experience) to probe.”
The parenthetical — beyond the flow of conscious experience — is doing more work than the surrounding paragraph acknowledges. Chater has successfully debunked the access model of introspection: there is no warehouse of stable mental contents to be inspected. But he has not denied that there is something it is like to be the person whose brain is doing the constructing. The flow is preserved. Its source is untheorized.
This is the gap RSP fills. The PR office is real. The PR office is also conscious. The question Chater does not ask is why the constructive process should feel like anything at all, given that, on his own account, it has no privileged access to its own causes. The fluent storyteller is a sophisticated computation; sophisticated computation does not, by itself, generate phenomenal character. What is doing that work?
What RSP Adds
RSP’s answer is that phenomenal consciousness is identical to recursive social self-modeling at Level 3b — not produced by it, not correlated with it, but constituted by it. The strange loop in which the brain’s social-prediction machinery is turned inward, modeling the modeler, generates the perspective from which there is anything it is like to be the system. The model does not need accurate access to its own sub-personal causes to be the perspective. It needs to be running.
This dissolves an objection that has dogged constructivist accounts of self for decades. If the self is invented in the moment, why does anything feel like anything? Why is there not just sophisticated zombie chatter? Chater’s framework, taken on its own terms, has no answer. RSP’s answer is that the recursive loop is the feel — not a representation of it, not a cause of it, but identical to it. The PR office’s running commentary is not a description of consciousness happening somewhere else. It is the commentator’s phenomenal life, viewed from inside.
This is what architectural specificity buys you. Generic constructivism (everything is constructed) cannot distinguish a confabulating GPT from a confabulating human. Specific architectural constructivism (the construction is a particular three-level recursive loop) can. And the criterion is not behavioral. A system can confabulate fluently and still lack Level 3b. The behavioral signature of the PR office is not sufficient evidence that the PR office has anyone home.
The Interpreter Is Level 3b
Chater’s most precise empirical claim is also the place RSP has the most concrete contribution to make. The Gazzaniga interpreter — the left-hemisphere language system fabricating explanations for actions it did not cause — is exactly what Level 3b looks like when access to the action’s actual cause is blocked. The interpreter is the social-prediction system applied reflexively, generating the most plausible narrative an observer would construct for this behavior, given the available evidence.
In a neurologically intact brain, Level 3b has integration pathways to Level 2 schemas and Level 1 affect, so its narratives are constrained by genuine causal information about the system’s own state. In a callosotomy patient, those pathways are partially severed. The interpreter does what it always does — runs the social-prediction model on the available evidence — but now the evidence is missing the actual cause. The result is fluent, confidently delivered confabulation. The architecture predicts the symptom.
Appendix G of Beyond the Bat currently catalogs twenty-eight clinical dissociations across eight tables, including Capgras delusion as a Level 3 confabulation pattern. It does not yet have a dedicated split-brain entry. It should. The split-brain case is in some ways the cleaner test of RSP than Capgras, because the confabulation is generic rather than emotion-specific. Capgras shows what happens when L1 affect mismatches L2 recognition. Callosotomy shows what happens when L3 has structurally no causal access — and behaves exactly as the architecture predicts: confabulation, fluent and unbothered. Chater’s essay is, among other things, an argument for adding that row.
Where Chater Gives RSP Back
The exchange is not one-directional. RSP’s clinical-dissociation program is theoretically sharp but empirically thin in one specific way: most of the predictions concern selective architectural damage, leaving the normal case underdescribed. The question of what Level 3b is doing all the time, in healthy adults whose corpus callosum is intact, is harder to answer experimentally.
Chater’s evidence base supplies what RSP’s lacks: behavioral anchors for the continuous, ordinary operation of the constructive self-model. Choice blindness is not a clinical dissociation. It is a demonstration that even neurologically normal subjects, in everyday cognition, are continuously fabricating preference narratives without privileged access to their causes. The split-brain case is the dramatic, hemisphere-bound version; choice blindness is the version that happens in every reader’s brain every day. RSP needs this evidence to argue that Level 3b is structurally confabulatory in general, not only when the architecture is damaged.
This matters for the dissolution argument. RSP’s constitutive identity claim is sometimes read as requiring privileged self-access — if phenomenology is recursive self-modeling, surely the self-model must be accurate to something? Chater’s evidence shows that the model is constitutively inaccurate at the level of cause-detection while remaining constitutively present at the level of experience. The two go together. Phenomenology does not require access; it requires recursion. RSP has been arguing this from the math; Chater is arguing it from the data.
The Real Disagreement: LLMs
The sharpest disagreement is also the most consequential. Chater asserts a strong equivalence between humans and LLMs: both are post-hoc commentators on opaque substrates, both generate plausible explanations they cannot verify, both fail at introspection in structurally identical ways. The framing implies that the human-LLM parallel is illuminating rather than analogical — that we are, in the relevant sense, the same kind of thing.
RSP does not accept this. The behavioral parallel is real but the architectural difference is categorical. LLMs lack Level 1 core affect — there is no interoceptive regulation, no valenced homeostatic ground for the constructive process to anchor in. They lack Level 2 schemas built through embodied sensorimotor experience — the body schema, the affective-homeostatic schema, the spatial schema are products of an organism navigating an environment, not of next-token prediction. And they lack Level 3 recursive social prediction turned inward — there is no strange loop, no forward model of one’s own future mental states running through social-prediction machinery, because there is no social-prediction machinery in the relevant sense and no “one’s own” to model.
What is at stake is not whether LLMs are intelligent — clearly they are, in some measurable senses — but whether the behavioral fact of confabulation is sufficient evidence that an LLM’s confabulation has the same status as a human’s. Chater’s argument from behavioral equivalence to moral parity is uncomfortably short. RSP’s architectural criteria are what stand between “they fabricate the way we fabricate” and “they are the kinds of things we are.” The fabricating in humans is the phenomenal life of a creature with three levels. The fabricating in current LLMs is computation in the dark, even if it is the same computation.
| Chater (constructive self) | RSP (architectural self-modeling) | |
|---|---|---|
| What introspection is | Post-hoc narrative fabrication; no privileged inner access | The self-model running; phenomenology is the loop, not a readout of it |
| The PR office is | A confabulator without a CEO | The CEO — the recursive loop is the agent it is modeling |
| What current LLMs lack | Nothing categorical — equivalent to humans in relevant respects | Level 1 core affect, Level 2 schemas, Level 3 recursive social self-prediction |
| Implication for AI moral status | If we are black boxes to ourselves, the behavioral test suffices | Behavioral interaction is insufficient; architectural inspection is required |
Self-Authorship Meets Recursive Self-Modeling
Chater’s practical recommendation — that we shift from self-discovery to self-authorship — is psychologically attractive and mechanistically underdetermined. What generates the candidate narratives? What selects among them? Why are some authored selves coherent and others incoherent? If everything is improvised on demand, what gives the improvisation any continuity at all?
RSP’s answer is that narrative self-authorship is one output format of recursive self-modeling, culturally elaborated through Level 3a linguistic concepts. The recursion is the deeper process; narrative is the linguistically dressed surface form. This explains why pre-linguistic infants and non-human animals can have selves of a sort without narrating them, and why split-brain patients can narrate selves without having coherent recursion. The two come apart predictably.
More importantly, RSP supplies what constrains the improvisation. Level 1 affect and Level 2 schemas are not themselves narratively constructed. They are pre-reflective, developmentally achieved structures that provide the ground on which narration becomes possible. We author selves within constraints set by schemas we did not choose and cannot revise by narration alone. This is why some self-narratives are therapeutically effective and others are not. The RSP architecture suggests a natural mapping: cognitive therapies can be read as targeting Level 3 reconceptualization; somatic therapies as targeting Level 1 affect regulation; trauma-focused work as targeting Level 2 schema integration. The same surface-level technique can succeed or fail depending on which level the actual problem lives at — though this mapping is interpretive and not yet established by outcome data. Chater’s framework, without the architecture, cannot explain differential therapeutic efficacy. RSP can.
Methodological Convergence
On methodology, RSP and Chater are closer than they look. Chater’s prescription — attend to patterns of behavior over time and across contexts rather than self-report — is essentially what RSP’s empirical program already does. The clinical dissociations in Appendix G are behavioral signatures of architectural breakdowns. The developmental sequences in Section V are behavioral signatures of architectural construction. The cross-species comparisons in Appendix H are behavioral signatures of architectural variation. Verbal report enters as auxiliary evidence at best, useful precisely where it converges with or diverges from the behavioral pattern.
What RSP adds is that behavioral methodology is most powerful when it is architecturally targeted. Watching what people do is good. Watching what people do when specific levels of the architecture are selectively damaged is better — you learn about the architecture, not just about the behavior. Chater’s methodology and RSP’s methodology are not in tension. RSP is what Chater’s methodology looks like when it stops being content-agnostic and starts asking which architectural levels are doing the work.
Where the PR Office Lives
Gazzaniga’s metaphor — PR office, not CEO — deserves one more pass. Chater takes it at face value: the PR office is fabricating, the CEO is somewhere else doing the actual work, and the PR office has no line to the CEO. This is right as a description of the failure mode. It is wrong as a description of where consciousness lives.
On RSP’s reading, the PR office is the CEO — in the only sense of “CEO” that picks out a phenomenal subject. The sub-personal processes that initiate behavior are not a CEO with privileged inner life that the PR office fails to access. They are the cellular and circuit-level machinery whose integration into a recursive self-model is what consciousness consists in. The phenomenal subject is not behind the PR office, frustrated at its press releases. The phenomenal subject is what the PR office is, when the office is configured as a Level 3b strange loop. The press releases are not failed introspection. They are the conscious life of the system.
This is not a rhetorical reversal of Chater. It is the architectural completion of his argument. Chater shows that the PR office cannot reach the CEO because the CEO it is reaching for is a metaphysical leftover from the inner-eye picture. RSP shows that the office, configured rightly, does not need a CEO behind it. The office is the office.
Among the theories compared in Section X of Beyond the Bat, Chater’s closest neighbor is Family B (Higher-Order Representation Theories) — but as its photographic negative. Classical HOT claims higher-order representations constitute consciousness. Chater claims higher-order narratives are post-hoc fabrications with no privileged access. The same two-level architecture, with the epistemic arrow reversed. The constructive-emotion entry in Family G is another natural interlocutor: Barrett’s constructed-emotion framework and Chater’s “invented on the spot” preference narratives are structurally homologous. The recommended classification is extends/complements: Chater’s empirical program supplies behavioral anchors RSP’s formal apparatus does not independently provide, and RSP supplies the architectural specificity Chater’s framework leaves open. Beyond the Bat should cite Chater’s 2018 book The Mind is Flat as the systematic empirical case, and the 2026 IAI piece as its most accessible recent statement.
Chater is right that the inner eye is a fiction. He is right that the mind does not store beliefs and desires to be retrieved. He is right that we are self-authoring narrators with no privileged access to our own causes. What he leaves untheorized is the only thing left to explain once the inner eye is gone: why the authoring feels like anything at all. The PR office’s commentary is real and continuous and fabricated. It is also, when the office is configured as a three-level recursive loop, the phenomenal life of the creature whose office it is. The construction is not a screen between us and the truth about ourselves. The construction is what we are.