The Neurocomputational Architecture of Auditory Verbal Hallucinations
Reconceptualizing "hearing voices" not as localized lesions or chemical imbalances, but as systematic mathematical breakdowns in the brain's Bayesian sensory inference engine.
The Bayesian Brain & Hierarchical Predictive Coding
Perception is an active process of hypothesis testing. According to Bayes' theorem, the brain constructs reality by integrating top-down prior expectations (P(θ)) with bottom-up sensory evidence (P(D|θ)), weighted by their respective precisions (Π = 1/σ²). When prior precision overwhelmingly dominates, the brain synthesizes percepts matching expectations rather than acoustic reality.
Simulate Perceptual Inference
Select a physiological condition to observe how the resulting perceptual posterior shifts across sensory space:
Corollary Discharge & Sensory Attenuation Failure
Whenever motor speech is initiated (including internal speech), an efference copy generates a corollary discharge (CD) signal to auditory cortex, attenuating predicted sensory feedback (observable electrophysiologically as N1 ERP suppression). In voice-hearers, CD failure leaves self-generated inner speech unattenuated and surprising, triggering an ascending prediction error that high-level cortices explain away as an external voice.
Mechanism of External Attribution
Because inner vocalization lacks normal CD-mediated motor prediction, the sensory feedback arrives at superficial pyramidal layers with unattenuated magnitude.
Unattenuated inner speech acts as a massive bottom-up mystery. The optimal Bayesian explanation for an unpredicted voice-like signal is the hypothesis that someone else is speaking.
Circular Inference & Inhibitory Loop Failure
Developed by Jardri & Deneve, the Circular Inference model demonstrates how an imbalance in excitatory/inhibitory (E/I) GABAergic loops breaks the functional separation of top-down and bottom-up signals. Signals reverberate repeatedly, causing the network to overcount information, driving rapid belief convergence to absolute false certainty.
Healthy Balanced Hierarchy
Normal CancellationPathological Reverberation
E/I DisinhibitionLoop Impairment Topography
Ascending Loop Failure: Predictions are counted as fresh sensory inputs, directly generating positive symptoms like hallucinations.
Descending Loop Failure: Sensory data is repeatedly treated as high-level confirmation, correlating with negative symptoms and cognitive disorganization.
Canonical Microcircuits & Neurochemistry
Predictive coding maps onto the laminar architecture of the cerebral cortex (Bastos et al., 2012). Superficial layers transmit ascending prediction errors via fast gamma oscillations, while deep layers transmit descending predictions via slower alpha/beta rhythms.
Superficial Pyramidal Cells
Function: Encode and transmit ascending Prediction Errors (PEs).
Receptors: AMPA / Kainate (Fast Glutamatergic).
Gain Control: Post-synaptic excitability modulated by Dopamine & Acetylcholine.
NMDA / GABA Balance
Function: Enforce local E/I balance and prevent circular feedback.
Receptors: NMDA-R on Parvalbumin-positive GABAergic cells.
Pathology: NMDA hypofunction disinhibits superficial layers, generating bottom-up noise.
Deep Pyramidal Cells
Function: Encode and transmit descending top-down Predictions (Priors).
Receptors: NMDA (Slower, integrative glutamatergic signaling).
Pathology: Compensation for chaotic PE noise results in hyper-precise deep priors.
| Neuromodulator | Computational Target | Physiological Action | Psychotic Manifestation |
|---|---|---|---|
| Dopamine (DA) | Prediction Error Precision (Π_PE) | Boosts post-synaptic gain of superficial pyramidal cells | Aberrant salience; assigning extreme importance to noise |
| Acetylcholine (ACh) | Sensory Likelihood Precision (Π_Sensory) | Suppresses intrinsic recurrent collateral connections | Cholinergic deficit causes blinding to reality → Hallucinosis |
| GABA / Glutamate | E/I Balance & Signal Separation | Parvalbumin interneuron-mediated feedback inhibition | NMDA hypofunction leads to circular feedback & overcounting |
Pavlovian Conditioning & HGF Computational Profiling
A landmark study by Powers, Mathys, & Corlett (2017) used an auditory-visual Pavlovian conditioning task and a 3-level Hierarchical Gaussian Filter (HGF) model to computationally isolate voice-hearers across four distinct groups.
Key Empirical Discoveries
Both clinical and non-clinical voice hearers exhibited heavily elevated prior weighting (n). They reported hearing conditioned tones even in absolute silence.
High-level environmental volatility rigidity (X3) strictly separated treatment-seeking patients from non-clinical voice hearers (e.g. psychic clairaudients).
Phenomenological Stratification Matrix
Computational psychiatry cleanly decouples the perceptual experience of voice-hearing from the clinical distress and disability of psychotic illness.
Treatment-Seeking AVHs
Schizophrenia Spectrum- ✕Corollary Discharge: Severe deficit; loss of N1 ERP suppression during inner speech.
- ✕Origin of Imbalance: Bottom-up E/I NMDA noise forcing reactive, protective priors.
- ✕Subjective Agency: Absent; voices experienced as intrusive, autonomous, and alien.
- ✕Volatility (X3): Pathologically rigid; inability to update beliefs when contexts shift.
- ✕Emotional Valence: Ego-dystonic, hostile, critical, generating high clinical distress.
Non-Treatment-Seeking AVHs
Healthy Clairaudients- ✓Corollary Discharge: Normal physiological motor-sensory attenuation intact.
- ✓Origin of Imbalance: Top-down hyper-precise priors deployed volitionally.
- ✓Subjective Agency: High; individual can initiate or terminate voice encounters.
- ✓Volatility (X3): Highly flexible and adaptive to environmental changing rules.
- ✓Emotional Valence: Ego-syntonic, protective, insightful, spiritually meaningful.
Computationally-Informed Therapeutic Targets
By viewing hallucinations through specific mathematical parameters, existing treatments can be repurposed and novel modalities designed to precisely tune inferential machinery.
Neuromodulation (rTMS & tDCS)
Low-frequency (1Hz) rTMS applied to the Temporoparietal Junction (TPJ) directly dampens local hyperactivity, suppressing the pathological reverberation of ascending prediction errors.
Psychopharmacology
D2 antagonists globally reduce prediction error precision (Π_PE), starving strong priors of false feedback. Future glutamatergic/cholinergic drugs aim to restore sensory precision at source.
Avatar & Narrative Therapies
Interacting with digital voice avatars restores active inference and agency, restructuring high-level threat hyper-priors and shifting emotional valence from hostile to controlled.