Is consciousness an illusion? Put that baldly and most people answer with a shrug or a laugh, because the one thing nobody seems able to doubt is that seeing red, or stubbing a toe, feels like something. A small group of philosophers and scientists answers yes anyway, though in a narrower sense than the slogan suggests. They do not deny that people perceive, think, feel pain in the functional sense, or report all of this. What they deny is that experiences have the special inner, ineffable, private qualities that philosophers call phenomenal properties. On their view those qualities are how our own brains misdescribe their states to themselves.

The position is called illusionism, and the dispute over it is one of the few in philosophy of mind where empirical work might actually move the needle. This article sets out the precise question, what is settled, the rival answers at full strength, and then a structured analysis built for this article: a list of predictions that would separate illusionism from realism about phenomenal consciousness, sorted by how much each could really discriminate. Readers who want the background first may prefer our plain guide to the hard problem; this piece assumes it and asks a narrower question about evidence.

The question, stated precisely

Two claims need to be pulled apart. The first is about reports: humans reliably judge that experience has qualities which seem impossible to capture in physical or functional terms. David Chalmers calls the task of explaining those judgments the meta-problem of consciousness (Chalmers 2018). That task is not in dispute. Nobody thinks problem reports are uncaused, and Chalmers himself classifies explaining them as, strictly, one of the easy problems, since it concerns behavior and the mechanisms that produce it.

The second claim is the contested one. Illusionism, in Keith Frankish’s influential statement, is a view

“which holds that phenomenal consciousness is an illusion and aims to explain why it seems to exist”

Frankish, 2016, Journal of Consciousness Studies

So the open question is this: once the meta-problem is solved, is there anything left over? Realists say yes: a real phenomenal character that the judgments are about. Illusionists say no: the judgments are the whole story, and they are systematically mistaken in the way a visual illusion is mistaken (Frankish 2016). Frankish’s term for what introspection actually detects is quasi-phenomenal properties, physical properties that introspection represents as phenomenal.

Why care? Three reasons. If illusionism is right, the hard problem is a pseudo-problem and consciousness science can be done entirely in functional terms. Moral status, including the moral status of animals and machines, would then have to be anchored in something other than phenomenal feel. And the question of whether an AI system is conscious would change shape completely.

People do ask this outside seminar rooms. A web search on the question, run while preparing this article, returned an essay by the philosopher Philip Goff titled with the same question (Goff 2023), long discussion threads on LessWrong-style forums and the Effective Altruism Forum, a podcast episode with Frankish, and an Aeon essay arguing for a middle position. That is anecdotal, but it shows the question has left the journals.

What is established, and what is not

Some ground is genuinely shared. Here is a list of points that serious participants on both sides accept, followed by what remains disputed.

  • Problem reports have mechanisms. Chalmers, no illusionist, argues the meta-problem is tractable and lists candidate mechanisms: introspective models, phenomenal concepts, introspective opacity about underlying processes, and the attribution of primitive qualities to experience.
  • Introspection is fallible, sometimes badly. Eric Schwitzgebel’s review of cases, from peripheral vision to emotional experience, concludes bluntly that “The introspection of current conscious experience, far from being secure, nearly infallible, is faulty,” (Schwitzgebel 2008).
  • Brains model their own processing. Michael Graziano’s attention schema theory holds that the brain builds a simplified model of its own attention. In a neural network agent that learned to control a simulated spotlight of visual attention, removing the attention model degraded control (Wilterson and Graziano 2021). A later preprint found modest gains from attention-schema modules in multi-agent reinforcement learning, with the authors calling the results preliminary (Liu et al. 2023, preprint).
  • Self-reports in machines can be partly grounded. In concept-injection experiments on large language models, the best model detected and named an artificially injected internal state on roughly 20 percent of trials, with no false positives in controls (Lindsey 2025; a research report, not peer reviewed). The author explicitly sets aside whether this bears on subjective experience.

What is not established: whether problem intuitions are universal (one preprint reports evidence that they are not widespread among lay people, Sytsma and Ozdemir 2019, preprint), whether any self-modeling mechanism actually generates the full pattern of human problem reports, and whether explaining the reports would explain away their subject matter. That last item is the crux, and no experiment has touched it directly.

The competing answers at full strength

Strong illusionism

Phenomenal properties, in the philosophers’ sense, do not exist. Introspection represents sensory states as having intrinsic, ineffable qualities, and that representation is false (Frankish 2016). Daniel Dennett’s “user illusion” picture belongs here, and Graziano’s attention schema supplies a candidate mechanism: a cheap, schematic model of attention that depicts awareness as a non-physical inner glow because a detailed model would be costly and useless for control.

For it: parsimony, compatibility with physics, the documented unreliability of introspection, and the fact that every other case of an apparently irreducible property (life, heat, color as a property of surfaces) eventually yielded to functional explanation. Against it: it asks us to doubt the one thing that seems most directly given, and it has no consensus mechanism yet.

Non-reductive realism

Phenomenal consciousness is real and not identical to any functional property. The meta-problem has a mechanistic answer, but our judgments are also about something real, perhaps linked to it by psychophysical laws. Chalmers takes roughly this line: “explaining our judgments about consciousness does not suffice to solve or dissolve the problem of consciousness.” (Chalmers 2018).

For it: the apparent conceivability of functional duplicates without experience, and the Moorean point that we know we are conscious better than we know any premise of an argument against it. Against it: it must say why our judgments are reliable if they would be produced the same way with or without the phenomenal extra, which is the hardest problem realists face.

Reductive realism

Experience is real and is identical to some physical or functional process; introspection gets the existence right but misdescribes the nature, making it seem non-physical. This view sits close to illusionism, close enough that the remaining gap can look verbal. Most consciousness scientists working on the theories compared in our survey of the main scientific theories of consciousness implicitly hold some version of it.

For it: it keeps both the data of experience and a naturalistic ontology. Against it: illusionists say it cannot explain why experience seems non-physical without becoming illusionism; anti-physicalists say it simply ignores the explanatory gap.

The intuition-first rejection

Goff, a panpsychist, puts the critic’s case in the plainest terms: he is about 95 percent confident illusionism is false, holding that the existence of experience is more evident than the existence of the external world, and that all philosophy rests on trusting what seems most evident (Goff 2023). The illusionist reply is that this is exactly what a well-designed illusion would make one say.

The meta-problem as the battleground

Why does the meta-problem matter so much? Because illusionism’s main argument is a debunking argument. If a mechanism can be specified that would produce our confident reports of ineffable inner qualities whether or not such qualities exist, then those reports lose their evidential standing. Chalmers grants the conditional force of this:

“If a consensus solution of this sort ever develops, then support for illusionism may swell.”

Chalmers, 2018, Journal of Consciousness Studies

A realist then has two replies. One is that our knowledge of experience does not run through the report-generating mechanism at all; it is direct acquaintance. The other is that the mechanism and the phenomenal facts are connected, so the mechanism is reliable. Each reply carries costs. Acquaintance has to explain how a non-causal relation shows up in what people say; the reliability reply has to identify the link.

Chalmers also points to an experimental strategy: build artificial systems with principled self-models and see whether they produce human-like problem reports. He mentions an early attempt by Luke Muehlhauser and Buck Shlegeris, a very simple agent that produced structurally similar reports, and adds a caution: a system that simple could talk this way while almost nobody would call it conscious. That caution, which cuts in favor of the illusionist, is where the predictions below start.

An analysis built for this article: predictions that could separate the views

The method here is simple to state. For each candidate observation, ask what strong illusionism predicts, what realism (reductive or not) predicts, and whether the two predictions actually differ. A prediction both camps would happily accept does no work, however interesting it is. Each item is then rated for discriminating power: high, moderate, or low. The ratings are the author’s judgment, explained in the text, not measurements.

#ObservationIllusionism predictsRealism predictsDiscriminates?
1Agents with an opaque, schematic self-model, trained with no human text about minds, spontaneously report ineffable inner qualities; matched agents without the self-model do notYes, and report content should track what the self-model leaves outPossible; realists accept a mechanism for reportsLow alone; moderate if content tracks the model’s omissions precisely
2Altering the self-model changes reports about experience while first-order discrimination stays fixedYesReductive realists: yes, if the self-model is part of experience; non-reductive: unclearModerate
3Problem intuitions vary sharply with culture, training and framingExpected if the illusion is partly learnedLess expected if experience is a robust datumModerate, but weak illusionism survives either result
4Introspective errors cluster on intrinsic, qualitative character rather than on structure and timingYes, that is where the misrepresentation livesNo particular clustering expectedModerate to high
5A complete mechanistic model generates every human problem report with no free parameter for phenomenal factsYesNon-reductive realists predict a residue of judgments that resist itHigh in principle; requires a finished theory
6Machine self-reports become causally grounded in internal states while stripped of human consciousness vocabularyGrounded self-reports need not include problem talk unless the model is opaqueNo firm predictionLow to moderate

Several things stand out when the predictions are laid side by side.

First, item 1, the experiment most often proposed, is weaker than it looks. A system that says it has ineffable qualia would impress no one if it had been trained on human writing, since it could be echoing. Even an uncontaminated system only shows that a mechanism can produce such reports. Realists already grant that. The test sharpens if the content of the reports lines up with the specific gaps in the self-model: if an agent whose attention model omits timing information comes to insist that its awareness is timeless, and one whose model omits spatial grain says its awareness is seamless and unified, that is a fingerprint of misrepresentation rather than detection.

Second, the strongest discriminators (items 4 and 5) target the realist’s weakest point, the reliability of judgments. If errors about experience are concentrated exactly where a compressed self-model would predict, the best explanation of the judgments makes no reference to phenomenal facts, and the debunking argument gains force. The existing evidence on introspective error (Schwitzgebel 2008) points in this direction but was not gathered to test it.

Third, item 3 is a trap for both sides. Lay surveys showing weak problem intuitions (Sytsma and Ozdemir 2019, preprint) do not refute realism, since a real phenomenon can be under-noticed. Strong, universal intuitions would not refute illusionism either, since a hard-wired illusion would also be universal. The result matters mostly for the historical, culturally learned version of the meta-problem solution.

Fourth, human neurology already supplies natural experiments for item 2. Cases in which patients confabulate fluent accounts of states they cannot access, familiar from work with split-brain patients, show that report systems can run loose from what they report on. What is missing are studies designed to ask whether reports about the qualitative character of experience come apart from first-order perception in the specific way a self-model account requires.

What this analysis does not show. It does not show that any of these results would settle the metaphysics. Even the ideal outcome for illusionism (items 1, 4 and 5 all confirmed) leaves the acquaintance reply available to a realist willing to pay its price. Empirical work can make illusionism the best explanation of the evidence about reports; it cannot force anyone to accept that the reports are the only evidence.

The author’s reading

This section is opinion. Illusionism is, in my reading, underrated as a research program and overrated as a settled conclusion. It is underrated because it converts a seemingly intractable question into a set of engineering and psychology questions, and those questions are answerable. It is overrated because the step from “we can explain why we say X” to “X does not exist” requires an epistemic premise, that our knowledge of experience runs entirely through the report mechanism, which is exactly what realists deny and which no experiment directly tests.

The most likely outcome over the next decade, again as opinion, is partial convergence. Mechanistic accounts of problem reports will improve, reductive realists and weak illusionists will find their remaining differences mostly verbal, and a hard core of non-reductive realists will remain, citing acquaintance. That is not a stalemate; it would shrink the disagreement to a precise epistemological point.

Specific, falsifiable predictions that could move the debate:

  • If an attention-schema style agent, trained with no human mental vocabulary and given a channel for describing its own states, produces descriptions whose odd features (timelessness, unity, ineffability) map onto what its self-model omits, illusionism gains; if such agents never produce problem-like descriptions under any self-model design, the attention-schema mechanism loses support.
  • If a preregistered study finds that human introspective errors are no more frequent for qualitative character than for structural features like timing and location, prediction 4 fails and the debunking case weakens.
  • If disrupting attention-related self-modeling (for example with targeted stimulation) changes how people describe the felt quality of a percept while leaving discrimination accuracy intact, the self-model account is supported; if no such dissociation can be produced, it is in trouble.
  • If assessments of AI systems against indicator properties (Butlin et al. 2023, preprint) are extended to include self-model opacity as an indicator, illusionism would predict that this indicator, rather than workspace or recurrence alone, best predicts whether a system talks about a hard problem.

Common questions

Do illusionists claim people do not feel pain?

No. They accept that pain exists as a state that hurts in the functional sense: it grabs attention, motivates avoidance, and is reported as awful. What they deny is that it also has an extra, intrinsic, non-functional quality. Critics reply that this extra quality is precisely what “feels” means.

Is illusionism self-refuting, since an illusion must be experienced?

This is the most common objection. Illusionists answer that an illusion here means a misrepresentation, which needs no experiencer in the phenomenal sense, just as a thermostat can misrepresent temperature. Whether that answer works depends on whether misrepresenting is enough to explain the seeming itself, and that remains disputed.

Would illusionism mean AI could be conscious?

In one sense it makes the question easier: if consciousness is a kind of self-modeling, then a machine with the right self-model has whatever humans have. In another sense it dissolves the question, because there is no further fact about inner feel to discover. Large surveys of AI against theories of consciousness (Butlin et al. 2023, preprint) do not take a side on illusionism.

Is this the same as saying the self is an illusion?

Not quite. Claims that the self is a construction concern the unity and persistence of a subject. Illusionism concerns the qualitative character of experiences. The two often travel together, but a person could accept one and reject the other.

Further reading

As an Amazon Associate I earn from qualifying purchases. A commission may be earned if you buy through this link, at no extra cost to you.

References

  1. Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., et al. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv preprint 2308.08708. https://arxiv.org/abs/2308.08708
  2. Chalmers, D. J. (2018). The Meta-Problem of Consciousness. Journal of Consciousness Studies, 25(9-10), 6-61. https://consc.net/papers/metaproblem.pdf
  3. Frankish, K. (2016). Illusionism as a Theory of Consciousness. Journal of Consciousness Studies (2016 special issue on illusionism). Author eprint: https://dl1.cuni.cz/mod/resource/view.php?id=688036
  4. Goff, P. (2023, May 11). Is Consciousness an Illusion? Substack essay. https://philipgoff.substack.com/p/is-consciousness-an-illusion
  5. Lindsey, J. (2025). Emergent Introspective Awareness in Large Language Models. Transformer Circuits Thread (research report). https://transformer-circuits.pub/2025/introspection/index.html
  6. Liu, D., Bolotta, S., Zhu, H., Bengio, Y., & Dumas, G. (2023). Attention Schema in Neural Agents. arXiv preprint 2305.17375. https://arxiv.org/abs/2305.17375
  7. Schwitzgebel, E. (2008). The Unreliability of Naive Introspection. Philosophical Review, 117(2), 245-273. https://faculty.ucr.edu/~eschwitz/SchwitzPapers/Naive1.pdf
  8. Sytsma, J., & Ozdemir, E. (2019). No Problem: Evidence that Problem Intuitions Are Not Widespread. Preprint, PhilSci-Archive. https://philsci-archive.pitt.edu/15573
  9. Wilterson, A. I., & Graziano, M. S. A. (2021). The attention schema theory in a neural network agent [full title not verified]. Proceedings of the National Academy of Sciences. https://pmc.ncbi.nlm.nih.gov/articles/PMC8379943