Controlled hallucination
Controlled hallucination: perception as active brain construction
- Predictive coding in visual cortex is mechanistically established: Rao & Ballard (1999) showed that hierarchical feedback predictions plus feedforward error signals account for extra-classical receptive field effects (surround suppression, end-stopping) in V1 that purely feedforward models cannot explain, providing the computational architecture underlying controlled hallucination
- The rubber hand illusion demonstrates that body ownership is a top-down prediction, not a sensory readout: Synchronous visual-tactile stroking induces ownership of a rubber hand; the effect requires anatomical plausibility, confirming that the brain's prior body model — not raw sensory input — determines ownership attribution
- Binocular rivalry shows the brain selects among competing generative hypotheses rather than averaging inputs: When different images are presented to each eye, perception alternates rather than fuses; top-down beta-band oscillations from frontal cortex modulate which hypothesis wins, consistent with active predictive inference
- Mismatch negativity provides direct neural evidence for ongoing prediction-error computation across sensory modalities: MMN is elicited automatically in audition and vision when expectation is violated, demonstrating that prediction error signalling is a continuous, attention-independent perceptual process, not a special-case phenomenon
- Observer-relative colour illusions (The Dress) establish that even low-level colour attribution is prior-weighted, not stimulus-determined: The same retinal image produces opposite colour percepts across observers based on differing priors about ambient lighting; this case cannot be explained by peripheral encoding or bottom-up processing alone
- Pathological hallucination and ordinary perception lie on a continuum of prediction-to-error weighting: Reduced MMN amplitude in schizophrenia tracks the same prediction error dysregulation that Seth's account predicts for hallucination — the disorder selectively impairs the correction mechanism, not the generative one
- The transparency of perceptual experience is consistent with, not evidence against, controlled hallucination: Experience feels direct because successful, confirmed prediction suppresses awareness of the predictive machinery; transparency is the phenomenology of tightly constrained prediction, not evidence of unmediated world-access
- Naive realist disjunctivism is logically coherent but empirically costly: It requires that veridical perception and hallucination implement categorically different mechanisms despite producing phenomenologically indistinguishable outputs, making the MMN reduction in schizophrenia and RHI data difficult to explain parsimoniously
Research Question
What is the evidence that perception is a generative, top-down process rather than a bottom-up readout of the world — and what are the strongest objections to Seth's "controlled hallucination" framing?
Findings
Executive Summary
Perception is a generative, top-down process: the brain continuously predicts sensory input and uses incoming signals primarily as correction, not construction. This is established across four independent experimental paradigms — the rubber hand illusion, binocular rivalry, mismatch negativity, and observer-relative visual illusions — and mechanistically grounded in Rao & Ballard's (1999) predictive coding architecture for visual cortex. Seth's "controlled hallucination" framing names this correctly: ordinary perception and pathological hallucination share the same computational mechanism and differ only in the degree to which sensory error signals constrain the brain's ongoing generative model. The strongest philosophical objections (naive realism with disjunctivism, Gibson's direct perception) are coherent but are each empirically costly when forced to account for the full evidence base. The hard problem of consciousness remains unresolved within this framework, but that is a separate question from whether perception is generative.
Key Findings
-
Predictive coding in visual cortex is mechanistically established: Rao & Ballard (1999) showed that hierarchical feedback predictions plus feedforward error signals account for extra-classical receptive field effects (surround suppression, end-stopping) in V1 that purely feedforward models cannot explain, providing the computational architecture underlying controlled hallucination.
-
The rubber hand illusion demonstrates that body ownership is a top-down prediction, not a sensory readout: Synchronous visual-tactile stroking induces ownership of a rubber hand; the effect requires anatomical plausibility, confirming that the brain's prior body model — not raw sensory input — determines ownership attribution.
-
Binocular rivalry shows the brain selects among competing generative hypotheses rather than averaging inputs: When different images are presented to each eye, perception alternates rather than fuses; top-down beta-band oscillations from frontal cortex modulate which hypothesis wins, consistent with active predictive inference.
-
Mismatch negativity provides direct neural evidence for ongoing prediction-error computation across sensory modalities: MMN is elicited automatically in audition and vision when expectation is violated, demonstrating that prediction error signalling is a continuous, attention-independent perceptual process, not a special-case phenomenon.
-
Observer-relative colour illusions (The Dress) establish that even low-level colour attribution is prior-weighted, not stimulus-determined: The same retinal image produces opposite colour percepts across observers based on differing priors about ambient lighting; this case cannot be explained by peripheral encoding or bottom-up processing alone.
-
Pathological hallucination and ordinary perception lie on a continuum of prediction-to-error weighting: Reduced MMN amplitude in schizophrenia tracks the same prediction error dysregulation that Seth's account predicts for hallucination — the disorder selectively impairs the correction mechanism, not the generative one.
-
The transparency of perceptual experience is consistent with, not evidence against, controlled hallucination: Experience feels direct because successful, confirmed prediction suppresses awareness of the predictive machinery; transparency is the phenomenology of tightly constrained prediction, not evidence of unmediated world-access.
-
Naive realist disjunctivism is logically coherent but empirically costly: It requires that veridical perception and hallucination implement categorically different mechanisms despite producing phenomenologically indistinguishable outputs, making the MMN reduction in schizophrenia and RHI data difficult to explain parsimoniously.
-
Gibson's direct perception faces its greatest challenge from prior-dependent cases: The Gibsonian framework handles skilled embodied action well but cannot explain The Dress (the stimulus is invariant; the percept is observer-relative) or body ownership illusions (the rubber hand is not in the ambient optic array) without importing top-down modelling.
-
The hard problem displacement objection is philosophically valid but empirically irrelevant to the generative-perception claim: Seth brackets the hard problem deliberately; whether this is scientific pragmatism or evasion depends on one's prior probability that phenomenal consciousness is physically explicable, not on the neuroscience of predictive perception.
Assumptions
- Assumption: Secondary sources accurately represent Seth's position in Being You ch. 1–3. Justification: Five independent reviews (LSE Review of Books, MIT Technology Review, CIFAR Q&A, naturalism.org, Psychology Today) converge on the same framing. The probability that all misrepresent Seth in the same direction is low. The core claim ("we're all hallucinating all the time") is a direct quote attributed to Seth in multiple sources.
- Assumption: The Rao & Ballard (1999) model is the appropriate mechanistic grounding for Seth's thesis. Justification: Seth cites predictive processing as the theoretical basis, and Rao & Ballard is the canonical computational instantiation of hierarchical predictive coding. Prior completed item established this connection.
- Assumption: The MMN literature applies to the same predictive mechanism Seth invokes. Justification: Friston's FEP-based account of MMN and Seth's controlled hallucination are both instances of precision-weighted prediction error minimisation. This equivalence is standard in the literature.
Analysis
The experimental evidence was evaluated for independence: the RHI, binocular rivalry, MMN, and visual illusion paradigms use different sensory modalities (somatosensory, visual, auditory, visual respectively), different experimental designs (synchronous/asynchronous stroking, dichoptic presentation, oddball paradigm, natural photographs), and different dependent variables (proprioceptive drift, perceptual alternation rates, ERP amplitude, colour report). Their convergence on a top-down generative account is therefore strong in the same way that a claim supported by multiple independent measurement methods is stronger than a claim supported by replications of a single method.
The competing accounts (naive realism, Gibson) were evaluated against the same evidence base. Naive realism's disjunctivist move acknowledges the evidence for hallucination being generative while insulating veridical perception from the same account. The cost is that it provides no explanation for the RHI in normal (non-hallucinating) subjects, no explanation for The Dress, and no explanation for why MMN reduction correlates with hallucinatory severity if veridical perception is mechanistically different. Gibson's account handles action-relevant perception well but is silent on observer-relative perceptual phenomena.
The hard problem objection was separated from the generative-perception claim because they are logically independent. The neuroscience of how perception works is not resolved by the philosophy of why there is any experience at all.
Risks, Gaps, and Uncertainties
- The primary texts (Being You, Surfing Uncertainty, The Predictive Mind) were not directly read — findings depend on multiple independent secondary reviews, which creates a small residual risk of systematic misrepresentation.
- The link between prediction error dysregulation and hallucinatory severity in schizophrenia is associational, not causal. The direction of causation and the exact mechanism remain contested in computational psychiatry.
- The claim that binocular rivalry is mediated specifically by top-down frontal beta-band oscillations rests partly on a preprint (arXiv) rather than a fully peer-reviewed consensus finding; the directional claim (top-down modulation exists) is well-established, but the specific neural mechanism (beta band from frontal cortex) is medium-confidence.
- Seth's thesis says nothing directly about unconscious perception or implicit processing — its scope is explicitly conscious experience. The evidence base used here (particularly MMN, which operates without attention) partially conflates conscious and non-conscious prediction error minimisation.
Open Questions
- Can the Gibsonian affordance framework be formally reconciled with predictive processing (as enactivist accounts attempt), or are they genuinely incompatible? This may be resolvable via formal modelling but is currently open.
- Does the "real problem" research programme eventually force an engagement with the hard problem, or can a complete neural account of conscious experience be given without answering why there is any experience at all?
- What are the limits of the controlled hallucination framing for edge cases — perception in severely cognitively impaired individuals, neonatal perception before rich priors develop, or non-human animals? Does the framework scale down to minimal creatures?
sources
- [x] Seth, A.K. (2021). Being You: A New Science of Consciousness. Ch. 1–3. — accessed via multiple secondary reviews (MIT Technology Review, LSE Review of Books, CIFAR Q&A, Psychology Today, naturalism.org)
- [x] Clark, A. (2016). Surfing Uncertainty: Prediction, Action, and the Embodied Mind. — synthesised in prior completed item
2026-02-28-predictive-processing-active-inference.md - [x] Hohwy, J. (2013). The Predictive Mind. OUP. — synthesised in prior completed item
- [x] Rao, R.P. & Ballard, D.H. (1999). "Predictive coding in the visual cortex." Nature Neuroscience. — paper PDF accessible at — content reviewed via multiple direct summaries