In the reply to Henry Taylor I said in passing that his “refrigerator light illusion” — you think you are conscious of everything because the moment you check, you are — belongs to a larger family. This is that family. Four findings, from four literatures that do not cite each other, share one shape. Each is a case of a mind mistaking what it can generate on demand for what it already has. Laid side by side, they stop looking like four quirks and start looking like the fingerprints of a single architecture.
The Four
1. Knowledge: the illusion of explanatory depth
Ask people how well they understand a zipper, a flush toilet, a bicycle. They rate their understanding high. Then ask them to actually explain the mechanism, step by step, and the rating collapses under them. Rozenblit and Keil (2002) named this the illusion of explanatory depth, and showed it is specific: it hits explanatory knowledge far harder than facts, procedures, or stories. What you possess is a shallow gist plus the confident sense that the detail is there to be had. The detail is not there. You mistook a pointer for the model it points at.
2. Perception: the grand illusion
Look up from a countryside scene and it feels complete, every leaf present at once. Change-blindness experiments say otherwise: flick a blank frame between two versions of a scene and people miss large, repeated changes they would swear they were seeing. The felt richness of the visual field is not a stored high-resolution image; it is the standing ability to fetch any detail the instant attention asks for it (O’Regan and Noë, 2001). Taylor’s refrigerator light is the one-appliance version of this: the light is on whenever you open the door to check, so you conclude it is always on. Attention is the door.
3. Reasons: choice blindness
Johansson, Hall, Sikström, and Olsson (2005) showed people pairs of faces and asked which they preferred. Then, by sleight of hand, they handed back the rejected face and asked why the participant had chosen it. Most people — roughly 87% — did not notice the swap, and cheerfully explained a preference they had never expressed (“I liked her smile,” for the face they had just passed over). The reason was generated on the spot, to fit an outcome, and reported as though it had been the standing cause of the choice.
4. Agency: confabulation
This is the oldest of the four. Nisbett and Wilson (1977), in one of the most-cited papers in modern consciousness studies, catalogued people giving confident, articulate reasons for behavior actually driven by causes they could not have introspected — position effects, priming, context — and titled it telling more than we can know. Its sharpest form is the split-brain interpreter (Gazzaniga): the left hemisphere, shown only that the right hemisphere has acted, invents a reason and states it with no sense of invention. The account is manufactured to explain an action already taken and delivered as its cause.
The Shared Skeleton
Strip the domains away and the same three parts remain each time. There is a generative capacity that produces something plausible on request: an explanation, a percept, a preference, a reason. There is a metacognitive error that reads this capacity as a standing possession: I understand it, I saw it, I chose it for this reason, I did it because. And there is a non-neutral probe — the request to explain, the shift of attention, the handed-back card, the demand for a reason — that does not read the state so much as create the thing it seems to find.
One detail is worth pausing on, because it explains why some of these illusions were easy to expose and others took a century. The probe can run in either direction. When probing tries to retrieve (explain the toilet) it comes up empty and the illusion breaks; that is why the illusion of explanatory depth is demonstrated simply by asking. When the probe instead manufactures the thing it looks for — attend to the birdsong and the birdsong appears — the illusion seals itself, and you need an indirect method like change blindness to catch the system not generating. Same skeleton, opposite behavior under the light. The grand illusion hid longer because it fills its own gaps as fast as you check them.
One Architecture
Here is the claim RSP makes, and it is not a claim about any one of these illusions. It is that this is what introspection is. If the self-model does not read pre-existing inner states but generates its reports as it goes, then introspection should overstate what is already there, systematically, across every domain in which it operates — knowledge, perception, choice, agency. The four illusions are not four malfunctions of an otherwise transparent inner eye. They are the ordinary output of a generative system doing exactly what it is built to do, seen from four angles.
RSP states the commitment directly: “introspection is not transparent access to the structure of experience” but “itself a cognitive process that can mislead us about its own workings,” the givenness of raw experience being “an artifact of how consciousness presents itself.” The deeper move behind it is the framework’s anti-qualia-realism: there is no meaning-independent inner store being misreported, because there is no such store to begin with. This dissolves a puzzle these illusions otherwise raise. If introspection is unreliable, what is the reliable inner truth it is failing to report? On the standard picture there has to be one — a real preference, a real rich field, a real reason — that the confabulating report distorts. RSP says the report is not distorting a hidden original; it is the state, generated. The self-model is not a flawed witness to an inner world. It is the author of one.
This is where the architecture earns its keep, and where RSP separates from a merely deflationary shrug. Saying “introspection confabulates” is cheap; several people have said it. RSP says which system confabulates and why it must: a Level 3b self-model, built from the same machinery the brain uses to predict other minds, turned back on itself and required to issue a report in real time. It confabulates for the same reason it is conscious at all — because it models a self that it does not separately store. The framework already carries the general version of this in its comparison set, Chater’s introspection-as-illusion thesis, which it classes as complementary, agreeing on the improvising self-model and parting only on whether to call the result an illusion. RSP declines the word: an illusion misrepresents a reality behind it, and here there is no behind.
Why This Is More Than A Metaphor
Collecting four biases under one banner is only worth doing if it buys something. It buys a prediction with a shape. If these illusions are the signature of a generative self-model rather than four unrelated bugs, then the signature should be findable wherever the architecture operates and absent where it does not. It should show up in domains no one has tested yet, on the same terms: confident, probe-created, collapsing under retrieval and sealing under generation. And it should fail to appear, or appear differently, in systems that lack a persistent self-model producing reports about itself — which is a real, if hard, line to draw between a system that merely outputs text about its states and one that models a self it is at stake in. That is the RSP version of the claim, and it is more falsifiable than “introspection is unreliable,” because it says where the unreliability comes from and therefore where it should stop.
Honest Scope
Three limits, stated plainly. First, these four literatures are consistent with RSP; none was designed to test it, and none confirms it against rivals. Second, much of the deflationary reading here is shared property — Dennett, Chater, and the sensorimotor theorists would sign most of the paragraphs above; RSP’s distinctive contribution is not the deflation but the architecture that produces it (a socially-recursive Level 3b self-model with Level 1 affective grounding), which is what a bare “the mind is flat” leaves unspecified. Third, the unifying reading is mine and the framework’s; Rozenblit, Noë, Johansson, and Nisbett were not writing about consciousness architecture, and it would be its own small illusion of explanatory depth to imply they were. What survives all three caveats is modest and real: four independent findings have one skeleton, and RSP predicts that skeleton for a specific reason.
This is a synthesis, not a reply to a single paper: it develops RSP’s constructed-richness / anti-qualia-realism commitment by showing that four unrelated introspective illusions share the generative-on-demand signature the framework predicts. It adds no new theory family to the comparison and needs no playground change — Chater’s introspection-as-illusion thesis is already the in-paper member of the family, classed as complementary. Its natural paper home is the anti-qualia / introspection discussion, as convergent evidence that introspective reports are generated rather than retrieved. Companion to the Taylor reply, where the refrigerator light first pointed at the door.
Sources: Rozenblit, L., & Keil, F. (2002). The misunderstood limits of folk science: an illusion of explanatory depth. Cognitive Science, 26(5), 521–562. · O’Regan, J. K., & Noë, A. (2001). A sensorimotor account of vision and visual consciousness. Behavioral and Brain Sciences, 24(5), 939–1031. · Johansson, P., Hall, L., Sikström, S., & Olsson, A. (2005). Failure to detect mismatches between intention and outcome in a simple decision task. Science, 310(5745), 116–119. · Nisbett, R. E., & Wilson, T. D. (1977). Telling more than we can know: Verbal reports on mental processes. Psychological Review, 84(3), 231–259. · Gazzaniga, M. S. (1985). The Social Brain: Discovering the Networks of the Mind. Basic Books.