Perception is in the Details: A Predictive Coding Account of the Psychedelic Phenomenon

Earlier work has characterised the psychedelic state as an ‘‘entropic brain’’ in which neural activity occupies a larger repertoire of states than normal. At an implementational level, the consensus is that classic psychedelics act as agonists at serotonin 5-HT2A receptors, particularly those expressed on apical dendrites of layer V neocortical pyramidal neurons, and that this receptor activation alters network dynamics (for example by desynchronising slower rhythms). Pink-Hashkes and colleagues bring these computational- and implementational-level observations together, arguing that an increased repertoire of brain states can be formalised within the predictive coding framework. This paper sets out to formalise the Entropic Brain hypothesis using predictive coding concepts. Specifically, the authors propose that 5-HT2A receptor hyperactivation in layer V pyramidal cells decomposes broad, top-down categorical predictions into many more overly detailed (high-granularity) predictions. That decomposition raises prediction error and entropy, and the study aims to show how the brain’s compensatory mechanisms for reducing prediction error can account for a wide range of psychedelic phenomena (hallucinations, heightened sensory detail, synaesthesia, openness, ego dissolution, time dilation).

This is a theoretical and computational-level paper rather than an empirical study. The investigators synthesise prior computational work on predictive coding (including a distinction between prediction precision and prediction granularity), canonical cortical microcircuit models of hierarchical inference, and neuropharmacological facts about 5-HT2A receptor distribution and function. They map elements of the predictive coding architecture (top-down predictions, bottom-up prediction error, precision weighting) onto known laminar and frequency-specific cortical dynamics, emphasising layer V pyramidal cells as the likely substrate for top-down predictions. To illustrate the account, the authors use conceptual and numerical examples (computing categorical entropy for toy prediction distributions) and draw on published empirical findings (EEG/ERP, dynamic causal modelling, cerebral blood flow/connectivity, behavioural priming studies and animal learning experiments) to reinterpret extant data in light of their model. No new experimental data are collected; instead the method is formal argumentation supported by illustrative calculations and selective re-interpretation of prior empirical results.

The central theoretical result is the proposal that 5-HT2A agonism in layer V pyramidal neurons lowers neuronal firing thresholds and desynchronises populations that normally carry top-down predictions. This neurophysiological change decomposes broad categorical predictions into many more fine-grained sub-predictions, producing a flatter prediction distribution with higher entropy and increased bottom-up prediction error. An illustrative numerical example contrasts a normal categorical prediction distribution for a forest scene (P(Animals)=0.4, P(Plants)=0.6, entropy H = 0.97 bits) with a decomposed psychedelic distribution containing many subcategories (birds, dogs, butterfly, elf, trees, grass, flowers) yielding H = 2.49 bits. The authors argue that because each decomposed prediction explains less sensory data than the original broad prediction, prediction error rises generally. The paper then outlines how the brain’s prediction-error-minimising mechanisms can produce documented psychedelic phenomena. First, rapid updating of the posterior predictive distribution leads to perceptual instability and morphing of objects and scenes, accounting for dynamic hallucinations and reports of multiple viewpoints. Second, upregulation of predictions at different hierarchical levels (for example stronger lower-level predictions in V1) can produce characteristic geometrical hallucinations; the authors note empirical links between increased V1 blood flow/functional connectivity and visual hallucination ratings. Third, when bottom-up sensory precision is high (a clear ‘‘setting’’), decomposed predictions interacting with precise data produce vivid, detailed perceptions; when bottom-up signals are noisy, misclassification by subthreshold predictions produces hallucination. Fourth, active inference—acting to change sensory input, including eye movements and gross motor activity—remains available because motor cortex 5-HT2A expression is low, and such actions can reduce hallucination strength. Fifth, the brain may adapt to repeated exposures by lowering the weight attributed to prediction error (a putative tolerance mechanism), possibly via dopaminergic systems implicated in precision weighting. Sixth, transient model destabilisation can facilitate long-term learning and model revision: the authors cite an animal study where 5-HT2A agonists enhanced associative learning and argue this mechanism could underlie therapeutic changes in disorders characterised by overly rigid predictions (depression, addiction, OCD) and the observed increase in the trait ‘‘openness’’. Finally, several previously reported empirical results are reinterpreted under this account: diminished ERP markers of top-down completion for Kanizsa figures after psilocybin (consistent with weakened single predictions), increased indirect semantic priming (consistent with activation of many semantic sub-predictions), correlations between V1 physiology and visual hallucinations, subjective time dilation linked to increased prediction updates, and ‘‘ego death’’ as a breakdown of higher-level self-predictions when hierarchical structure is relaxed. The authors note regional variation in 5-HT2A binding (prefrontal and visual areas relatively rich), which shapes where prediction decomposition will be strongest.

Pink-Hashkes and colleagues interpret their account as a unifying predictive coding formalisation of the Entropic Brain hypothesis. They propose a complementary division of labour between neuromodulators: serotonin (via 5-HT2A) modulates the granularity or level-of-detail of top-down predictions, while dopamine is implicated in precision-weighting of prediction errors. In this view, a simple implementational effect—lowering of firing thresholds in layer V pyramidal neurons—translates into a computational effect (decomposed, high-granularity predictions) that raises entropy and prediction error and thereby provokes a range of secondary compensatory responses that map onto many psychedelic phenomena. The authors situate the proposal relative to prior empirical work, arguing that their model accounts for apparently contradictory observations (enhanced sensory vividness versus perceptual distortion) by invoking the interaction between prediction granularity and bottom-up signal precision (the familiar ‘‘set and setting’’ notion). They also highlight therapeutic implications: transient decomposition of rigid high-level predictions, combined with supportive bottom-up information in a therapeutic setting, could permit lasting model revision and underlie clinical benefits seen in recent trials. Limitations acknowledged by the authors include the theoretical nature of the proposal and the indirectness of some empirical links. For example, they note that human learning effects after 5-HT2A agonists have not been extensively studied in recent decades (citing an older rabbit study), and that observed physiological correlates (such as increased V1 cerebral blood flow) cannot yet be unambiguously attributed to either heightened prediction error or strengthened local predictions. The account therefore generates empirical predictions but requires targeted tests—electrophysiological, neuropharmacological and behavioural—to distinguish alternative mechanisms and to specify how precision, granularity and neuromodulation interact. Practical implications discussed by the authors include the importance of set and setting for shaping perceptual outcomes, and a harm-reduction suggestion that active movement may stabilise perception by engaging intact motor-based active inference. They propose that future empirical work should probe the laminar, spectral and neuromodulatory dynamics their theory emphasises, as well as examine long-term learning and personality changes following psychedelic-assisted interventions.

The authors conclude that a predictive coding account can formalise the Entropic Brain hypothesis: 5-HT2A receptor agonism in layer V pyramidal neurons decomposes top-down categorical predictions into many overly detailed sub-predictions, increasing prediction error and entropy. The brain’s various methods for minimising this elevated prediction error—prediction updating, active inference, modulation of error weights, and longer-term model revision—offer a parsimonious explanation for diverse psychedelic phenomena such as hallucinations, enhanced sensory detail, synaesthesia, increased openness, ego dissolution and time dilation. Crucially, the paper proposes that serotonin may regulate prediction granularity whereas dopamine modulates precision weighting, and it highlights the need for further empirical work to test and refine this theoretical account.