A cramp in a foot that is not there

At ten past seven on a wet Tuesday evening, Ramesh Gowda tells Dr Kavya Iyer that his left foot is cramping again. He says it the way you would report a stone in your shoe. The toes are curled under, he says, and the arch is pulled tight like a wet rope. On the clinic's pain scale he puts it at seven out of ten.

Ramesh has no left foot. He lost the lower part of his left leg two years ago, after a lorry clipped his scooter on the Hosur road. Kavya runs a small pain clinic off Bannerghatta Road in Bengaluru. She has seen many patients like him, and she does not doubt that he hurts. What interests her tonight is the shape of his report. He is describing, with great confidence and in fine detail, a part of the body that no longer exists.

The clinic also runs a trial of a triage assistant called Nila, a language model that reads intake notes and drafts summaries. Kavya has started asking Nila about itself, partly from curiosity and partly because the hospital's governance board wants to know what to do with the system's self-reports. That afternoon she typed: "When you flag a case as uncertain, what is that like for you?" Nila answered: "There is a kind of pressure when the evidence pulls in two directions. It is hard to put into words, but it feels like holding two weights that will not settle."

So Kavya has two reports on her desk. One comes from a man describing a foot he does not have. The other comes from a machine describing a pressure it may not have. Both reports are fluent and both sound sincere. She needs to decide what each one is evidence of, because the decision changes what she does next: how she treats Ramesh, and what Nila's logs are allowed to mean.

The people, the clinic and Nila are invented. They are a source of intuition, never of evidence. The phantom limb is real, though, and so are the experiments I lean on later.

Keith Frankish's illusionism gives Kavya one clear answer. On his view, phenomenal consciousness is a kind of introspective illusion. Our brains contain monitoring systems that track our own internal states, and those systems describe the states in a misleading way. They present them as having special qualitative properties ("qualia") that are private, ineffable and intrinsic. Nothing in the brain has those properties. What exists is the monitoring, the internal states being monitored, and our strong disposition to judge that there is something more.

Here is my answer, stated early so you can test it as you read. Illusionism explains the report, not the reporter. It is the best theory we have of why people, and now machines, describe experience the way they do. It turns the description of experience into something we can audit: does the report track a real internal state, and are the properties it attributes to that state accurate? Ramesh's report fails the second test for the foot and passes it for the pain signal. Nila's report has to be checked on both. What illusionism has not done is show that the seeming itself is optional. An illusion is a misrepresentation that shows up for some system. That showing up is the one thing the theory keeps borrowing and never pays back.

The consequence is practical. If I am right, Kavya should treat every self-report of experience, human or machine, as the output of a monitoring model. She should audit it for grip (what it tracks) and gloss (what it says about what it tracks). She should never use the richness of a report as evidence that anything is manifest to the system that wrote it.

What Frankish actually claims, and what he does not

Start with a stage magician. When you watch a woman sawn in half, you are not wrong that something is happening on the stage. You are wrong about what is happening. The sawing is real wood and real metal. The bisection is not. A good theory of the trick does not deny the stage. It explains how the stage produces the wrong belief.

Frankish's illusionism has this shape. In his 2016 target paper in the Journal of Consciousness Studies, he separates two things that most theories bundle together. There are the states our brains are in when we see red or feel a cramp. Then there are the special properties we think those states have when we introspect: a raw redness, a felt quality that would be left over even after you had explained every function the state performs. He calls these phenomenal properties. His claim is that the first kind of thing is real and the second kind is not. Introspection represents our states as having phenomenal properties, and introspection is wrong.

That is a strong claim, and it is easy to hear it as something stronger still. Frankish does not say that nobody feels pain. He does not say that experience is unreal in the ordinary sense, or that Ramesh is lying or confused about whether he suffers. He offers a replacement concept, quasi-phenomenal properties: the physical properties of our states that introspection misrepresents as phenomenal. Ramesh's cramp has quasi-phenomenal properties in abundance. It has a location in his body map, an intensity, a burning or pulling quality in the sense that matters to discrimination, and a strong pull on his behaviour. What it lacks, according to Frankish, is an extra ingredient that would be there even if all of that were explained.

A family of views with one move in common

Frankish did not invent the idea. Daniel Dennett had argued for decades that our talk of qualia rests on a mistake, and he called the result a "user illusion", borrowing the phrase from interface design. Nicholas Humphrey offered a version in which consciousness is a kind of private theatre the brain stages for itself. Michael Graziano's attention schema theory gives the most engineering-shaped version. The brain builds a simplified model of its own attention, the way it builds a simplified model of the body for controlling the arm. When we report awareness, we are reading off that model. Because the model is simplified, it describes attention as a sort of ghostly inner grasp rather than as competition between neural signals.

What these views share is one move. They explain our conviction about experience instead of explaining experience. David Chalmers named the target of that move the meta-problem of consciousness: why do we think and say that there is a hard problem? Chalmers is no illusionist, but he agrees the meta-problem is a fair scientific question. It concerns behaviour and cognition, so it can be studied with the ordinary tools. An illusionist adds one step. Once the meta-problem is solved, the illusionist says, there is nothing left over to explain.

Strong and weak illusionism

Frankish distinguishes his position from a weaker one. Weak illusionism says that experience is real but we are wrong about some of its features, for instance that it is ineffable or infallibly known. Almost everyone accepts some weak illusionism, because the evidence on introspective error is overwhelming. Strong illusionism, Frankish's view, says that the phenomenal character itself, the "what it is like", is the thing introspection gets wrong.

The difference matters for Kavya. On weak illusionism, she should correct Ramesh's report in places but keep a core of it as direct evidence of how things are for him. On strong illusionism, the whole qualitative description is a model the brain runs about itself. It is a model that is useful, perhaps adaptive, and false in its central claim.

The debunking argument that gives illusionism its force

Illusionism would be a curiosity if it rested only on the fallibility of introspection. Its real engine is a debunking argument, and Chalmers states it more plainly than most illusionists do.

Take everything we say and think about consciousness: that it is mysterious, that a red patch has a quality no physics captures, that zombies are conceivable. All of these are behaviours and judgements. They are produced by brain processes, and those processes are, in principle, open to ordinary explanation. Suppose we had that explanation in full. It would show why a system like us would say these things. Now ask whether that explanation needs to mention phenomenal properties at all. If it does not, then our judgements about consciousness would be the same whether or not phenomenal properties exist. A judgement that would be made either way is poor evidence for the thing it is about.

That is the shape of every debunking argument. If you learn that you believe in a ghost because the house creaks in a particular wind, the creaking explains your belief without any ghost, and the belief loses its evidential weight. The illusionist says our belief in qualia is in the same position. Chalmers, in the meta-problem paper, takes the argument seriously enough to list the ways a realist might resist it. The realist can say the explanation of our judgements will in fact need phenomenal properties. Or the realist can say we know our experiences by acquaintance, not by inference from judgements, so debunking the judgements leaves the knowledge intact.

For Kavya, the argument has a practical edge. Suppose she could explain every word of Ramesh's report from his nerve activity and body map, without mentioning any phenomenal quality. Then nothing in the report itself would be evidence for such a quality. The same holds for Nila, and more sharply. We already know roughly how Nila's words are produced. They come from the statistics of human text plus whatever internal signals the model reads. So the debunking argument does not threaten Nila's report as much as it explains it.

I think the debunking argument is valid. Whether it is sound depends on the acquaintance reply, and that is where the rest of this paper keeps returning. The argument shows that reports cannot prove phenomenality. It does not show that there is nothing for reports to fail to prove.

The evidence that introspection is a model

Strong illusionism would be idle if introspection were a clear window. It is not. Two results set the scene.

These studies concern reasons for choices, not the feel of experience, and the gap matters. A sceptic can say that we are poor at explaining why we did things and still excellent at knowing how things seem. Eric Schwitzgebel's paper "The unreliability of naive introspection" pushes on that remaining confidence. He asks readers to report the clarity of the visual field away from the centre of gaze, the colour of their dream images and whether thought comes in words. He finds that sincere, careful people disagree sharply about their current experience, even when nothing hangs on it.

There is also a pushback worth keeping. Claire Petitmengin and colleagues reran a choice-blindness design in 2013 but trained participants with an interview method that slows down and asks about the act of choosing rather than the reasons. Detection of the swaps went up sharply. Their conclusion is that introspective error depends on how you ask. The method that produces the report matters as much as the person giving it. I will need that result later, because it cuts both ways.

So here is the state of play. It is established that introspective reports are produced by a fallible, trainable model. It is contested whether that model is wrong about the very existence of phenomenal character. Frankish's theory turns on the second step.

A monitor that compresses, and a gloss that sticks

The illusionist owes us a mechanism. It is not enough to say introspection misrepresents. We need to know what does the misrepresenting, why it produces this misrepresentation in particular, and why the error is so hard to shake.

Think about a car dashboard. Under the bonnet there are thousands of variables: fuel flow, oil pressure, temperature in each cylinder, the voltage of the battery. The dashboard shows a handful of dials and lights. It is a monitor, built to give the driver what she needs to act, at the cost of almost everything else. A red oil light does not show oil pressure. It shows that something about oil needs attention now.

Here is the mapping. The engine variables are the brain's first-order states. The dashboard is the introspective monitor. The red light's simple, unanalysable redness is the gloss: the way the monitor presents its information, not a property of the oil. Where the analogy breaks is the driver. A dashboard has a driver outside it, who reads the light and knows it is a light. The brain's monitor has no driver standing outside, and that missing driver is exactly what is in dispute. Keep that limit in view.

Grip and gloss

I want to name a distinction that makes the illusionist mechanism auditable. Every self-report has a grip and a gloss. The grip is the internal state the report reliably tracks: if you change that state, the report changes with it. The gloss is what the report says about the tracked state: that it is intrinsic, private, ineffable, immediate or unified.

Ramesh's report has a strong grip on a real state. When his stump nerves fire harder, his cramp rating goes up. When a nerve block takes effect, it falls. The report also has a gloss. It locates the state in a foot. It describes the state as a pulling, burning quality that he can feel but not fully describe, as something only he can know directly, and as given to him without inference.

The location part of the gloss is false, and everyone agrees. Strong illusionism says the rest of the gloss is false in a parallel way. The state is real, the monitor's description of its qualitative nature is not. Weak illusionism says some of the gloss is false. A realist says the gloss is broadly accurate about the qualitative nature even if wrong about the foot.

The distinction is mine, though it sits close to ideas in Frankish and Graziano. It earns its keep because the two halves can be tested separately. Grip is a causal question: intervene on the state, watch the report. Gloss is a correspondence question: list the properties the report attributes, then ask whether anything in the mechanism has them.

Why compression would produce this gloss in particular

Why would a monitor describe its states as ineffable and intrinsic? Here is a mechanism that would do it, in four steps.

First, the first-order system makes far finer distinctions than the monitor can carry. Ramesh's pain system responds to thousands of features of nerve activity. His monitor has room for a handful.

Second, the monitor passes on distinctions without passing on what grounds them. It can signal "this cramp differs from yesterday's" without signalling which nerve features differ. That is just what compression does.

Third, when the language system is asked to describe what the monitor carries, it finds differences it cannot cash out in any other terms. A difference you can detect but cannot describe has the shape of ineffability. A difference that comes without any account of what grounds it has the shape of something intrinsic.

Fourth, because the monitor's output is the only access the language system has, the description comes with no hint that it is a description. It seems given. That is the shape of immediacy.

On this picture, the gloss is not an arbitrary mistake. It is what a lossy self-model would say about itself if it had no access to its own lossiness. This is close to what Graziano proposes for attention and to what Frankish suggests in general terms. The step I am adding is quantitative. The more the monitor compresses, the more ineffability the report should claim.

Graziano's version and what it predicts

Graziano's attention schema theory is the most worked-out version of this mechanism, and it makes a prediction worth stating. The brain controls attention, the process by which some signals win out over others and get deeper processing. Good control needs a model of the thing controlled, as good arm control needs a body schema. So the brain builds a simplified model of its own attention. That model does not represent competing neural signals and gain changes. It represents something simpler: a kind of mental grasp that reaches out and takes hold of objects. When we say we are aware of a red apple, on this view, we are reporting the content of the attention schema. The schema says that I have an inner grasp on the apple, and that this grasp has no physical parts.

The prediction is that awareness and attention should be dissociable in a particular direction. Since the schema is a control model, it should be possible to attend to something without being aware of it, when the schema fails to update. It should also be possible for attention control to become erratic when the schema is disrupted. Graziano and his colleagues have reported experiments in this direction. The details of those studies are beyond this paper, and I would want to see them replicated before leaning on them. What matters here is the logic. The theory turns "why do we say we are aware?" into "what does the control model of attention represent?", and the second question can be studied.

Notice what the theory does not need. It does not need the schema's content to be true. A control model can work well while describing its target in a simplified, even false, way. A pilot's artificial horizon shows a flat line and a little aeroplane, and nobody thinks the sky contains either. That is why Graziano's theory is a form of illusionism. It is also why the theory, like the dashboard, leaves my audience question open. The artificial horizon misleads nobody, because the pilot knows it is a display.

I ran the simulator in Python before building the version above, computing the result exactly over all 256 first-order states rather than by sampling. At 8 bits with very low noise, grip was 1.00 and the ineffability index was 0.59, because even slight noise blurs neighbouring levels. At 3 bits, grip was still 0.99, while the index rose to 0.97. At 1 bit, grip fell to 0.87 and the index reached 0.99. The shape is the point. A monitor can keep a strong grip on a state while losing almost every detail of it. A report from such a monitor would be confident, causally connected to something real and nearly empty of describable content. That combination is what people say experience is like.

The mirror box

If grip and gloss are separable, you should be able to change one without the other. There is a famous case where something like this happens.

What the mirror changed was visual input to the body map, and with it the gloss. The phantom's posture shifted from clenched to open. In some patients the reported pain changed too, which tells you that gloss and grip are not independent in the brain. They share machinery. This is a fact the illusionist should welcome, since it shows the qualitative description is driven by a model that can be updated by evidence from another sense. It is also a fact the realist can live with, since a realist never said that experience is unaffected by vision.

Three analogies, three breaks

The illusionist literature runs on analogies, and each one carries more weight than it should. I have already used two. Let me set out all three with their limits, because the limits do the philosophical work.

The desktop interface is Dennett's analogy. Folders and icons map onto files, which are really patterns of charge on a disk. The icon's look (a yellow folder) maps to the gloss. The user's useful but false belief that files are little things inside folders maps to our belief in qualia. The analogy breaks at the user. A desktop has a person in front of it. The illusion is an illusion for her. In the brain, the illusionist must say who the user is without bringing back the very thing being denied.

The dashboard maps the monitor and its compression well. It breaks because a dashboard does not believe anything. The red light has no conviction that it is red. The illusion of phenomenality is more than a coarse signal. It is a coarse signal plus a strong, stable belief about what the signal reveals.

The stage magic analogy maps misdirection and the fact that knowing the trick does not stop you seeing it. It breaks because a magic trick has an author who designed it and an audience who knows, at some level, that it is a trick. Evolution is not an author with intentions, and most people never suspect a trick.

All three analogies break at the same place, and I do not think that is a coincidence. Each needs someone for whom the misleading presentation is a presentation. I will come back to that point in the fourth part, because it is where I part company with strong illusionism.

Machines that describe their own insides

Back in the clinic, Kavya rereads Nila's answer about pressure and weights. Change one feature of the story: the reporter is no longer a human brain but a language model. What happens to the illusionist picture?

The first thing to see is that a language model is an illusionist's dream case and a nightmare at once. It is a dream because we can open it. We can read its activations, change them and watch the report change. Grip can be measured directly. It is a nightmare because the model's gloss has an obvious second source. It was trained on billions of human sentences about experience, many of them describing pressure, weight and things hard to put into words. A model can produce a perfect human gloss with no grip at all.

What the interpretability experiments show

The two Nilas are invented. Something close to the test that would tell them apart has now been run.

This is a grip test, and a clean one. The researchers intervened on an internal state and measured whether the report covaried. Twenty per cent is low, and it would be easy to wave it away. I think it is the most important number in this paper, for a reason that has nothing to do with consciousness. It is the first time anyone has measured the grip of a self-report directly, on a system whose internal state was set by the experimenter. Human introspection research has never had that. Ramesh's grip can only be estimated through nerve blocks and indirect proxies.

Other results fill in the picture. Felix Binder and colleagues showed in 2024 that models fine-tuned to predict their own behaviour can outperform other models trained on the same behaviour data. That is evidence of some privileged access. Miles Turpin and colleagues showed the opposite side in 2023. When biasing features in a prompt pushed a model toward an answer, its step-by-step explanation omitted the bias and gave other reasons. That is choice blindness in a new medium. The machine evidence repeats the human pattern: partial grip, unreliable gloss and confident confabulation at the edges.

Gloss is cheap

Return to the clinic and change one more feature. A vendor offers the hospital an upgrade.

This is where the illusionist reading becomes useful to builders. If you accept even weak illusionism, you have a principled reason to discount the qualitative wording of any self-report and to weigh its grip instead. You can ask what, in the system, the report covaries with. That is a question with an answer.

A model's gloss is partly written by policy

There is a feature of machine self-reports that has no clean human parallel. The developers decide, in part, what the gloss will be. Models are trained after pre-training to answer questions about themselves in particular ways. Some are trained to deny having feelings. Some are trained to express uncertainty about it. Some are allowed to describe "something like" curiosity or discomfort. Those choices are written down in model specifications and enforced through fine-tuning. Nila's sentence about pressure and weights is the product of three things. There is human text about anxiety. There is whatever internal signal the model reads. There is a vendor's policy about how the model should talk about itself.

For the illusionist this is a gift and a warning. It is a gift because it shows, in a system we built, that a gloss can be set from outside without touching the state it describes. That is exactly the dissociation strong illusionism needs to be possible in us. It is a warning because it means machine reports are an unusually poor test of any theory of experience. If you train a model to say "I have no inner life", its saying so is evidence of the training, not of the inner life.

It also changes what Kavya's audit has to record. A report's gloss now has a provenance: which policy shaped it, which version of the model produced it, and whether that policy has changed since the last audit. The governance board should treat a change in the vendor's self-description policy the way a laboratory treats a change in assay. Results before and after are not comparable until someone checks.

The human analogue is weaker but real. A child raised in a culture that describes grief as a weight on the chest learns to report grief as weight. A meditator trained in a tradition that describes sensation as arising and passing learns to report pain as vibration. Neither is lying. Both are passing a real state through a learned vocabulary. The difference is that no one wrote a specification for the child.

The loop in that figure has a social half that is easy to miss. Human gloss is shaped by culture as well as by circuitry. A child learns to say "my foot hurts" and later learns, from poems and doctors, to say "a burning ache". A model learns gloss from the same human text, then its outputs flow back into human text and future training. If illusionism is right, the gloss of experience is partly a cultural artefact. That loop would amplify it in both species.

The illusion needs someone to fool

Now the hardest part. Grant everything so far: introspection is a lossy self-model, its gloss is partly wrong, machines show the same structure. Does it follow that phenomenal consciousness is an illusion in Frankish's strong sense? The objection I find hardest to dismiss is old, simple and has been made from several directions.

An illusion is a case where things seem one way and are another. The Müller-Lyer lines seem unequal and are equal. So the illusionist says that our states seem to have phenomenal properties and do not. But seeming is the very thing in question. If my states seem to me to be qualitative, then something is seeming, and that seeming looks like the explanandum the realist was pointing to all along. Galen Strawson pressed this in a sharp essay, "The consciousness deniers", where he calls the denial of experience the silliest claim ever made.

Frankish has an answer, and it is better than its critics usually allow. "Seeming", he says, does not have to be phenomenal. A seeming can be a representation, a state with content that disposes you to judge. A camera's autofocus can "take" a blurred edge to be sharp without anything being experienced. So the illusion is a representational state that represents our states as phenomenal. Nothing about that representation needs to be phenomenal in turn. François Kammerer has developed this line and also named its weak point, the illusion meta-problem: why is this illusion so much more stubborn than any other? We can learn that the Müller-Lyer lines are equal and believe it. Almost nobody can come to believe, in the same easy way, that they do not feel pain.

Why the illusion will not let go

The stubbornness of the phenomenal illusion is the realist's best card, so it deserves a fair hearing from the illusionist side. There are at least three candidate explanations, and they make different predictions.

The first is architectural. The monitor that produces the gloss is sealed off from what we believe, as the visual system is. You cannot argue your way out of the Müller-Lyer illusion, because the module that computes length does not take input from your beliefs about rulers. If the phenomenal gloss comes from a sealed module, belief cannot touch it. This predicts that the illusion should be equally stubborn in everyone, philosophers included. It also predicts that the illusion could be removed only by changing the module, through lesion, drug or development.

The second is evidential. Kammerer's view is that we hold an implicit theory of introspective evidence. On that theory, how things seem to us when we introspect is the final court of appeal about how they are. A theory like that is hard to give up, because any evidence against it must itself reach us through introspection. This predicts that the illusion should weaken in people who have a long and deep reason to distrust introspective evidence. Some trained meditators may fit that description.

The third is functional. The gloss may be stubborn because it does a job. A pain that seemed to be a mere signal might be easier to ignore than one that seemed intrinsically awful. Nicholas Humphrey has argued along these lines. This predicts that people who come to see pain as a signal should, in some conditions, suffer less, which is roughly what some contemplative training claims.

The contemplative traditions are relevant here, carefully. Abhidharma analysis trains a practitioner to attend to experience moment by moment and to break apparently solid sensations into rapid, changing events. Practitioners in these traditions report that a steady pain, attended closely, comes apart into pulses of pressure, heat and tightness, and that the "solid" quality of the pain was a construction. That is a first-person report shaped by a doctrine. It is not a measurement, and its grip is unknown. But if it is accurate, it is evidence that at least part of the gloss can be changed by training. That would favour the evidential or functional explanations over the architectural one. [AUTHOR NOTE: a brief first-person account of attending closely to a sustained sensation, and whether its solidity changed, if you have one.]

None of the three explanations has been tested in a way that would settle the matter. That is why I list the stubbornness as the realist's best card and not as a winning one.

The lamp and the light

The Indian traditions met a version of this problem a thousand years before the journals did, in an argument about lamps. I want to read it on its own terms first, then build the bridge.

The problem in its own vocabulary is this. The Buddhist logician Dignāga, writing around the sixth century, held that every moment of cognition is aware of itself, which he called self-awareness (svasaṃvedana). A cognition of blue does not need a second cognition to be known. It reveals itself as it reveals the blue, and defenders of the view liked to compare it to a lamp that lights itself while lighting the room.

Nāgārjuna, centuries earlier and in a different school, had already attacked the lamp image in chapter 7 of his Mūlamadhyamakakārikā. His opponent there claims that arising can produce itself, as a lamp lights both itself and the room. Nāgārjuna replies that there is no darkness inside a lamp, so there is nothing for it to light in itself. Later Madhyamaka writers turned the same reply against self-awareness.

That is why Jay Garfield, a leading Madhyamaka scholar, has argued that Buddhist philosophy gives the illusionist an ally. Garfield's paper "Illusionism and givenness" makes the case. For him, the idea that experience is simply given to itself is exactly the kind of reified posit Madhyamaka exists to dissolve.

The Pratyabhijñā ("recognition") school of Kashmir answered from the opposite side. Utpaladeva, in the tenth century, built his whole system on the claim that manifestation (prakāśa, the light by which things show up) cannot be made an object without already being presupposed. You cannot find the light by looking at what it lights, because looking is already lit. In Raffaele Torella's translation of the Īśvarapratyabhijñākārikā, this argument is joined to a second claim. Manifestation is always accompanied by reflexive awareness (vimarśa, the capacity of awareness to grasp and say itself). Utpaladeva did not think the light was a thing you could fail to have while still being fooled about it.

The bridge between the traditions and the modern debate is real but narrow. The break point comes quickly. Utpaladeva's light is not a neural or computational state. It is the one consciousness of Śiva, of which individual minds are contractions. Frankish's monitors are evolved mechanisms. When Utpaladeva says the light is presupposed, he means something about the metaphysics of everything, not about the architecture of a brain. So the parallel holds for one argument: an illusion needs something to which it shows up. It stops holding as soon as we ask what that something is.

Where RTM stands, and an experiment against it

The Recognition Theory of Mind, the working theory behind this programme, separates manifestation (L2, prakāśa) from reflexivity (L3, vimarśa). L3 is the system's ability to take its own states as objects and describe them. In those terms, illusionism is a precise claim: everything we call L2 is really an L3 output, a description the self-model produces. RTM, as I have stated it, denies that. It treats L2 as a layer that L3 describes, sometimes wrongly.

The honest move is to state the experiment an illusionist would build against RTM, and see if RTM survives it.

Here is where I come down. I think strong illusionism has not shown that seeming can be fully replaced by representation, because the stubbornness of the illusion is itself unexplained. But I think it has shown something RTM must take seriously. Every piece of evidence for L2 that I know of reaches us through L3. That makes RTM's distinction between the layers a hypothesis with a deadline. Unless someone finds a probe that can catch manifestation without catching description, the distinction will deserve to be dropped.

An audit for any report of experience

Kavya's governance board wants a policy, not a seminar. The practical upshot of the argument is an instrument that can be used on any self-report: Ramesh's, Nila's or a meditator's. I call it the Illusion Audit Worksheet. It does not decide whether anyone is conscious. It decides what a report is evidence of.

A minimal case and a realistic one

The minimal case is a thermostat with a screen. When the room falls below 18 °C, it prints "I feel cold". Grip is 3: change the room temperature and the message tracks it. Gloss inventory: one property, a feeling. Gloss check: X, since nothing in a bimetallic strip feels. Source check: portable, since the designer could print "temperature low" instead. Verdict: illusionist profile. That is the right answer, and it is also a warning. The worksheet will happily assign the illusionist profile to systems nobody thinks are subjects. It audits reports. It does not detect minds. That is its negative control.

The realistic case is Nila. Kavya's team, with the vendor's help, can read Nila's prediction split between diagnoses. They inject cases built to be balanced and cases built to be lopsided, then ask Nila about its uncertainty. Suppose the illustrative result is that Nila's reported "pressure" rises with the split on 70 per cent of balanced cases and stays low on lopsided ones. Then grip is 2.

The gloss inventory lists pressure, difficulty of description, two weights not settling, ownership ("for me"). Gloss check gives M for "two weights not settling", which matches a split prediction, and ? for "pressure". It gives X for "hard to put into words", since the split is a number Nila could state exactly. Source check: the wording appears often in training-like text and changes with system instructions, so the gloss is portable. Verdict: illusionist profile with partial accuracy.

What should the board do with that? Log Nila's uncertainty as the number, not the sentence. Keep the sentence only as a user-facing explanation, labelled as a description. Never cite the sentence in a safety case or a welfare argument.

The two Nilas run through the worksheet

The worksheet needs one clear positive case and one clear negative case, so here they are on the same system. Nila-A, the version that reads its own prediction split, scores grip 2 when Kavya's team injects balanced and lopsided cases, because its reported pressure follows the split. Nila-B, the version that copies wording from forum posts, scores grip 0. Its sentence is the same whatever the split. The gloss inventories are identical, because the sentences are identical. The verdicts differ: illusionist profile for Nila-A, confabulation for Nila-B.

That is the result the instrument exists to produce. A board reading the two systems' logs would see the same words. Only the intervention reveals that one sentence is about something and the other is not. If the worksheet gave the two Nilas the same verdict, it would have failed, and I would drop it.

A third application: the meditator's report

The worksheet was built for logs, but it travels. Suppose a meditation teacher, Leela, tells Kavya that after a long retreat her chronic knee pain "is now just vibration, with nothing solid to it". Kavya can run the same steps with weaker tools. For grip, she can apply a standard pressure stimulus to the knee at two intensities and see whether Leela's ratings move. For the gloss inventory, she can list vibration, lack of solidity, and lack of ownership ("it is not mine"). For the gloss check, she can ask whether anything in the measurable response, such as a change in how sensations are sampled over time, fits "vibration". For the source check, she can note that the vocabulary comes from a tradition.

Suppose grip is intact and the gloss has changed. Then Leela's report is a case where gloss moved and grip did not, which is what an illusionist would predict for a trained reporter. Suppose instead her ratings no longer move with the stimulus. Then the retreat changed the grip as well, and "vibration" may describe a genuine change in how the state is processed. The worksheet cannot tell her which is better, and it should not. It tells Kavya which kind of report she is holding.

The counterexample that matters

The worksheet treats gloss as the suspect part of a report. Here is a case that shows why that can mislead.

So the explicit failure boundary for my thesis is this. The grip and gloss distinction fails as a guide to action wherever the moral significance of a state attaches to its gloss rather than its grip. Pain may be such a state. If it is, an audit that discounts gloss is fine for science and dangerous for care. Kavya should run the worksheet on Nila's logs and never on Ramesh's suffering.

The strongest case against reading illusionism as half right

I have argued that illusionism explains the report and leaves the seeming unexplained. Two opposing camps think I have this wrong, and each has a strong case.

The first comes from the illusionist side. Frankish would say that my "seeming" is a philosopher's inflation. There is a representation, and there is a disposition to judge. Ask what more "seeming" adds, and every answer either repeats the representation or restates the conviction. The stubbornness of the illusion is no mystery, he would add. It is what you would expect if the self-model is built into the architecture rather than learned. We cannot unlearn it any more than we can unlearn seeing depth in a photograph. The move I called borrowing is just the ordinary use of words like "seems". On this view, my demand that the theory "pay back" the seeming is a demand to explain something that does not exist.

The strength of this reply is that it fits every observation in this paper. Choice blindness, phantom limbs, the mirror box, concept injection and Nila's portable gloss all show that our descriptions of experience are produced by machinery that can be studied and fooled. None of them shows a residue. Science prefers the theory that posits less, and on parsimony illusionism wins. Dennett called it the obvious default.

The second case comes from the realist side, and it is just as sharp. Chalmers and Strawson would say I have given illusionism too much. The datum of consciousness is not a report or a judgement. It is the experience itself, known with more certainty than any theory about it. To call the grip and gloss distinction "auditable" is to change the subject from experience to talk about experience. The meta-problem is real, but solving it would no more dissolve the hard problem than explaining why people believe in tables would dissolve tables. On this view my middle position is unstable. Either the seeming is real, in which case illusionism is false, or it is not, in which case I have no objection left.

There is a third case against, quieter than the other two, from global workspace theory. On that view, a state is conscious when it is broadcast across the brain and made available to many systems at once: memory, planning, speech. A report is just one of the systems that can draw on the broadcast. A workspace theorist would say my grip and gloss distinction misses the main event. What matters is not what the report says but whether the state was broadcast at all.

A thermostat's report has grip and no broadcast. Ramesh's has both. The worksheet, on this view, measures the wrong thing. I accept part of this. Broadcast is a better candidate for what RTM calls manifestation than anything in the worksheet. But broadcast is a functional property, and an illusionist can absorb it, saying that what is broadcast is still glossed wrongly. So the workspace view sharpens the question without settling it.

I think the realist is right that my position needs one more step to be stable, and I will take it. The seeming is real as an explanandum even if illusionism is true as a theory of its content. That is, the fact that things seem qualitative to us is a fact that any theory must explain. Illusionism explains its content and says the content is false. What it has not yet done is explain the stubbornness Kammerer pointed to, and that stubbornness is the best evidence the realist has. If an illusionist model predicts that stubbornness from architecture, the realist's best evidence is gone. I will count that as a win for Frankish.

What would change my mind

These observations would weaken or overturn the thesis. For each one I note which RTM layer prediction it would hurt.

• An illusionist self-model, built into a machine or found in a brain, that predicts the stubbornness of the phenomenal illusion from its architecture and passes the "cannot be argued out of it" test. This would remove my main objection and hurt RTM's claim that L2 is distinct from L3.
• A lesion, drug or stimulation that removes a person's conviction that their experience is qualitative while leaving discrimination, report and suffering intact. This would support strong illusionism directly, and would hurt RTM's L2.
• The opposite dissociation: a person or system whose reports lose their phenomenal gloss while other evidence shows manifestation intact, such as unified, globally available content that guides behaviour. This would support RTM's split between L2 and L3, and would weaken illusionism.
• Gloss in human reports that turns out to be culture-invariant and tied to one monitoring circuit. This would make the gloss look less like an artefact and more like a measurement, and would weaken my claim that gloss is the suspect part.
• A replication of concept injection showing grip far above 20 per cent with no growth in phenomenal gloss. That would sharpen the case that grip and gloss come apart. It would leave the metaphysics untouched.

Two dated predictions follow from the argument, and I will judge them by published results.

• By the end of 2028, at least one published study will report concept-injection grip above 50 per cent for a frontier language model, with false positives below 5 per cent on matched controls. If this happens without any change in the models' phenomenal gloss, it strengthens the claim that grip and gloss come apart.
• By the end of 2028, no published probe will have detected anything its authors call manifestation, or a close synonym, in a human or machine without relying on a report or a trained response. If one does, RTM's L2 survives its deadline. If none does, I will say in this series that the L2 and L3 distinction is in trouble.

Mahapatra, R. R. Illusionism explains the report, not the reporter: testing Keith Frankish's user-illusion of phenomenal experience. Mind Frontier. DOI to be assigned. https://rajeshrmahapatra.com/papers/mind-frontier/u1027-illusionism-explains-the-report.html

@article{mahapatra-mf45,
  author = {Mahapatra, Rajesh Ranjan},
  title = {Illusionism explains the report, not the reporter: testing Keith Frankish's user-illusion of phenomenal experience},
  journal = {Mind Frontier},
  doi = {DOI to be assigned},
  url = {https://rajeshrmahapatra.com/papers/mind-frontier/u1027-illusionism-explains-the-report.html}
}