Laying Down a Path While Walking — Thirty Years of Neurophenomenology· ASSC Satellite MeetingLang
COMPUTATIONAL NEUROPHENOMENOLOGY · THE LIVED STRUCTURE OF PERSISTENT PERCEPTUAL 'ERROR'
What Is It Like to Be Mistaken?
From Correlation to Generative Constraint in the Computational Neurophenomenology of Visual Illusions
Alfredo Muñoz AlarcónDoctoral candidate in Philosophy · Universidad Alberto Hurtadoalmunoz@uahurtado.cl
01Isn't this a contradiction?
First-person experience and third-person formal models look like different worlds. The easy bridges lose the phenomenon — collapse experience into representations, or merely correlate reports with brain states.
02Varela's answer: invariants
Neither reduce nor correlate — place the two under reciprocal constraint. Lived experience has describable invariants, recovered by disciplined first-person method (Varela 1996; Lutz & Thompson 2003).
03From correlation to constraint
Let phenomenology enter as a constraint on the form of a generative model — which architectures are admissible (a family, not one model) — not as data to be fit. No claim that experience is a representation.
04Classic case: Müller–Lyer
Seeing what you know to be false becomes a positive design specification — a case where a first-person invariant does work no third-person measure can do alone. ⇒
ThesisA cognitively impenetrable illusion is a stable conflict between perceptual appearance and epistemic correction. The phenomenon of interest is therefore the lived structure of being wrong: continuing to see what one knows to be false. That structure functions as an exclusion criterion over the architectures admissible for any generative model of human perception.
What do I see? — one line looks longer. Studied by psychophysics.
E2Error recognition
What do I know? — the lines are equal. Studied by metacognition.
E3Lived error
How do both coexist? — the appearance stays compelling. The target of this proposal.
§1 · Why Mistakenness?
Most perceptual errors disappear once we inspect the scene, gather new evidence, or measure the stimulus. Cognitively impenetrable illusions are different: the appearance persists despite correction. This makes them philosophically distinctive — they let us separate the percept itself from the subject's awareness that it is mistaken. The resulting tension is not merely a computational anomaly; it is a lived structure of error.
§2 · The Belief–Percept Lock
The phenomenon has two layers.
Outerpercept vs stimulus — the classic "illusion"
Innerthe real explanandum — percept vs belief
A first-order estimate ("unequal") and a contrary higher-order belief ("equal") held simultaneously, knowingly, stably — the belief unable to revise the estimate. The lock is not merely a disagreement between perception and reality; it is the persistence of disagreement between perception and correction: the percept remains authoritative even after the subject acquires the relevant knowledge.
Outer · percept vs stimulus
Stimulus physically equal
≠
Percept looks distorted
Inner · the belief–percept lock
Percept "UNEQUAL"
Belief "EQUAL"
Knowledge cannot revise the percept.
§3 · Not a Malfunction
Judged against external correspondence, an illusion looks like a defect. On a constructivist, enactive reading it is the opposite — the output of mechanisms working exactly as built for the statistics of ordinary 3-D scenes. Brown & Friston (2012): the Cornsweet effect falls directly out of a generative model with ecologically plausible priors; the "illusory" percept is the most probable explanation of the input. The divergence appears only against artificial laboratory displays.
§4 · Correlation → Constraint
Classical neurophenomenology sought reciprocal constraint between first-person description and third-person measurement. Its computational extension (Ramstead et al. 2022; Sandved-Smith et al. 2025) casts this in generative models: a model of the phenomenology constrains the class of admissible neural realizations.
Constraint (constructive): report → an architectural requirement the model must satisfy.
The point. The first-person report does not contribute another variable to correlate with neural activity — it identifies a phenomenological invariant the model must preserve. Bridge principle: when an invariant is stable across the very interventions that would tell rival architectures apart, preserving it is an architectural, not merely behavioral, requirement. The lock meets that condition.
§5 · Expectations vs Constraints
Why doesn't a confident, correct belief propagate down and fix the percept? The predictive hierarchy implements prior information in two ways (Teufel & Fletcher 2020; Caporuscio et al. 2022) — and a third state stays distinct:
Explicit belief — doxastic / metacognitive; from measurement or testimony.
Top-down expectation — flexible, context-dependent; can modulate lower levels.
Bottom-up constraint — inflexible, structural; built into early processing.
Explicit belief"equal"
no causal edge ↓
Top-down expectation· modulates ↘
Bottom-up constraint· fixes ↘
illusory estimate
↓
percept
The lock is the absence of an edge from belief to constraint. A model that lets belief overwrite the estimate has modeled a correctable mistake — the wrong phenomenon.
§6 · A LIMITED CONTRAST: ARTIFICIAL PERCEPTUAL SYSTEMS
Yampolskiy proposes that illusions can serve as a test for qualia: “it is only by experiencing an illusion that the agent is able to enter into a certain internal state”(Yampolskiy, 2024 — AI: Unexplainable, Unpredictable, Uncontrollable, §10.2). A system reporting the illusion rather than the measurement would thereby reveal an experiential state.
Horizontal lines are:
Not in image
Crooked ●
Straight
Red
Orange circles are:
Left is bigger
Right is bigger ●
Same size
Not in image
Horizontal stripe is:
Solid
Spectrum of gray ●
Not in image
Crooked
●Our proposal asks a different question. The crucial phenomenon is not producing the illusion-consistent response but preserving the belief–percept lock. We grant the antecedent and block the inference: a human-like illusion profile shows shared outer-layer susceptibility, not the lock (which needs the inner layer). Shared susceptibility ≠ shared perception ≠ shared consciousness. What distinguishes the biological case is not an extra structural feature but the origin of the constraint: owned (sedimented by the system's own self-maintaining history) vs stipulated by a designer. The contrast is autonomy, not substrate — not carbon vs silicon; an autonomous artificial agent could in principle own its constraints too.
§7 · What It Is Like to Be Mistaken
At the level of perception there is no intrinsic marker of error. The Müller–Lyer line appears longer in exactly the same way a genuinely longer line would — so being mistaken does not feel different from being right. What becomes experientially salient is the conflict between appearance and correction: the subject measures the stimulus, acquires the belief, and yet continues to see the same appearance. The phenomenology of error emerges not inside the percept but in the persistence of this inconsistency — the central claim of the paper.
§8 · Scope and Limits
Scope: cognitively impenetrable geometric and lightness illusions, within a session. Out of scope (different invariants): bistable perception, impossible figures, and illusory motion (cases that carry a different felt anomaly). Slow perceptual learning can attenuate the effect (a diachronic retuning), not a within-session belief-driven correction (Jenkin 2023).
TakeawayCognitively impenetrable illusions reveal a stable structure of lived error: the subject continues to perceive what they know to be false. This conflict is not a curiosity to be correlated with neural activity, nor a test for the presence of qualia — it is a phenomenological invariant that constrains the architecture of admissible generative models. The recalcitrant visual illusions are the model case: not the subject of the inquiry but the clearest case in which the structure of first-person experience constrains a computational explanation (a dependence that may generalize beyond this one illusion).
Selected References
Brown & Friston (2012). Free-energy & illusions: the Cornsweet effect. Front. Psychol. 3:43.Buckner (2020). Adversarial examples & a theory of artefacts for deep learning. Nat. Mach. Intell. 2, 731–736.Caporuscio et al. (2022). When seeing is not believing. Conscious. Cogn. 102:103334.Calabi (2012). Introduction: errors in perception and more. In Perceptual Illusions (ed.).Fodor (1983). The Modularity of Mind. MIT Press.Gallagher, Hutto & Hipólito (2022). Predictive processing & disillusions about illusions. RPP 13(4).Jenkin (2023). Perceptual learning & reasons-responsiveness. Noûs 57(2), 481–508.Marr (1982). Vision. W. H. Freeman.Ramstead et al. (2022). From generative models to generative passages. Rev. Phil. Psych. 13.Sandved-Smith et al. (2025). Deep computational neurophenomenology. Neurosci. Conscious. niaf016.Suzuki (2026). Beyond the reducing valve: neurophenomenology of altered states via DNNs. Front. Psychol. 17, 1819038.Teufel & Fletcher (2020). Forms of prediction in the nervous system. Nat. Rev. Neurosci.Varela (1996). Neurophenomenology: a remedy for the hard problem. JCS.Veerabadran et al. (2023). Adversarial manipulations influence human & machine perception. Nat. Commun. 14, 4933.Yampolskiy (2018). Artificial consciousness: an illusionary solution. RSL Ital. J. Cogn. Sci. 2/2018, 287–318.Yampolskiy (2024). AI: Unexplainable, Unpredictable, Uncontrollable. §10.2.
I gratefully acknowledge funding from ANID through the National Doctoral Scholarship (21242212), and Dr. Francisco Pereira, lead researcher of the Fondecyt Regular project (1250205).
Draft paper
POSTER
GitHub