The Influence of Psilocybin on Subconscious and Conscious Emotional Learning

Casanova and colleagues situate their study in the context of a renewed interest in serotonergic psychedelics as potential therapeutics for psychiatric disorders. The introduction summarises that psilocybin, a 5-HT–acting tryptamine found in “magic mushrooms,” alters perception, mood and self-referential processing and has been linked to neuroplastic changes in animals and symptomatic improvement in conditions such as major depressive disorder. The authors highlight gaps in knowledge about how psilocybin affects reinforcement learning in humans, and specifically whether its effects differ when emotional cues are processed consciously versus subconsciously. The study set out to test whether psilocybin alters reinforcement learning in a probabilistic cue-reward task that used fearful and neutral facial cues presented at durations intended to favour either subconscious (33 ms) or conscious (47 ms) processing. The investigators had three preregistered hypotheses: that psilocybin would impair learning from negative (fearful) stimuli by decreasing their salience, that psilocybin would enhance learning with neutral and conscious stimuli via neuroplastic effects, and that reduced top-down processing under psychedelics would make subconscious cues more disruptive to learning under psilocybin than under placebo. The authors frame the question as clinically relevant because reinforcement learning processes are implicated in psychiatric disorders and in therapeutic change.

This was a randomized, double-blind, placebo-controlled crossover study in 30 healthy volunteers (23 male, 7 female; mean age 29), all white and right-handed. Participants underwent two test sessions 14 days apart and received either psilocybin or placebo (mannitol) in counterbalanced order. Psilocybin doses were weight-adapted into three categories: 10 mg (<50 kg), 15 mg (<80 kg), and 20 mg (≥80 kg). Key exclusion criteria included personal or family history of major psychiatric illness, major medical conditions, seizure history, current psychotropic medication, and pregnancy or breastfeeding. Approximately 240 minutes after substance administration participants completed EmotLearn, a probabilistic cue-reward task programmed in Presentation®. On each measurement day participants completed four tasks (eight tasks total across the study), each of 60 trials. Each task paired one presentation-duration condition (subconscious 33 ms or conscious 47 ms) with one emotional cue type (fearful or neutral faces from a standard face database). On each trial a face cue was shown, then a pair of symbols; participants selected a symbol and received +10 CHF or −10 CHF feedback. One symbol in each pair had a 70% chance of reward (30% loss); the other symbol had the inverse probabilities. Monetary outcomes were virtual and not paid out. The design aimed to probe how emotional valence and awareness modulate probabilistic reinforcement learning under psilocybin versus placebo. Raw trial-level accuracy (proportion correct per trial) was extracted and smoothed with a Gaussian filter. The Shapiro–Wilk test indicated non-normality of the learning data. The primary inferential approach was a hierarchical Bayesian mixed-effects regression using a beta distribution for the bounded accuracy outcome. Models included fixed and random effects for substance, cue category (fearful vs neutral, subconscious vs conscious), dosage and interactions; inference used Hamiltonian Monte Carlo (No-U-Turn Sampler) implemented in Stan with four chains of 2000 iterations each, 1000 warm-up iterations discarded, yielding 4000 post-warmup draws. Pairwise marginal means and highest posterior density (HPD) intervals were used to assess credible differences (95% HPD). Nonparametric Mann–Whitney tests were used for specific between-condition comparisons (e.g. reaction times, psychometric scales) and α = 0.05 was the nominal significance level.

Overall learning performance did not differ reliably between psilocybin and placebo: mean accuracy across all trials was 0.729 under psilocybin (N=30, n=3600, σ=0.332) versus 0.745 under placebo (N=30, n=3600, σ=0.324). The Bayesian model found the overall group difference included zero in the 95% HPD interval (0 ∈ 95% HPD), interpreted as noninferiority of psilocybin to placebo for the primary learning measure. Trial-by-trial curves showed comparable learning trajectories from ~50% at trial 1 towards higher accuracy by trial 30; only trial 20 showed a statistically significant advantage for placebo (p<0.005, Mann–Whitney, multiple-comparisons corrected). Levene’s test suggested borderline greater variance under psilocybin (p=0.05002), interpreted as more switching/exploratory behaviour. When analyses were stratified by cue awareness, subconscious cues under psilocybin produced worse performance than placebo: mean accuracy for subconscious cues was 0.714 with psilocybin versus 0.739 with placebo (N=30, n=1800 each, σ≈0.33), and pairwise comparisons showed psilocybin–subconscious was credibly worse than placebo–subconscious and worse than the three other conditions (0 ∉ 95% HPD). Emotional valence analyses showed that fearful versus neutral faces yielded similar learning curves under placebo until ~trial 22, after which they diverged; psilocybin fearful performance did not differ credibly from placebo fearful (0 ∈ 95% HPD) but was credibly worse than placebo neutral (0 ∉ 95% HPD). Subgroup contrasts highlighted that subconscious neutral cues particularly hindered learning under psilocybin: psilocybin–subconscious–neutral mean accuracy 0.717 (N=30, n=900, σ=0.342) versus placebo–subconscious–neutral 0.757 (N=30, n=900, σ=0.327), a credible difference (0 ∉ 95% HPD). Within the psilocybin condition, conscious–neutral cues performed better than both subconscious–neutral and subconscious–fearful (0 ∉ 95% HPD for both comparisons); the placebo group did not show these intra-condition differences. Order effects in the crossover were notable: participants who received psilocybin on the first day had lower mean accuracy (μ=0.689, n=16) than those who received psilocybin after placebo (μ=0.776, n=14) and those who received placebo first (μ=0.781, n=14); these differences were credibly significant (0 ∉ 95% HPD). Dose-stratified analyses showed the 20 mg psilocybin subgroup (μ=0.74, N=15, n=540) outperformed the same participants on placebo (μ=0.725, 0 ∉ 99% HPD), whereas the 15 mg subgroup showed the reverse (psilocybin μ=0.717 versus placebo μ=0.747, 0 ∉ 99% HPD). The 10 mg group included only two participants, limiting inference. Intrapersonal comparisons across the crossover showed heterogeneity: 12 participants (40%) performed significantly better with placebo, 7 participants (23.3%) performed significantly better with psilocybin, and 11 participants (36.7%) showed no significant within-person difference (Mann–Whitney, p<0.05 threshold). Monetary earnings did not differ significantly (placebo 133.9 ± 9.9 CHF vs psilocybin 103.5 ± 12 CHF, p=0.09), though the mean was 29.3% higher under placebo. Reaction times were slower on psilocybin (mean 1.73 s, σ=2.11 s) than placebo (mean 1.34 s, σ=1.62 s; p<0.001, Mann–Whitney); within psilocybin participants the 20 mg group had faster reaction times than the 15 mg group (1.49 s vs 2.01 s, p<0.001). Psychometric scales showed marked subjective effects: the 5D-ASC Vigilance Reduction score averaged 44 under psilocybin versus 9.75 under placebo (p<0.001, Cohen’s d 1.97), and the 11D-ASC Impaired Control and Cognition subscale averaged 29.48 under psilocybin versus 0.74 under placebo (p<0.001, Cohen’s d 1.52).

The investigators interpret their results as indicating that psilocybin preserves reinforcement learning capacity in a probabilistic cue-reward task relative to placebo, while selectively altering how awareness of emotional cues affects performance. Both groups showed upward learning curves and comparable overall accuracy, supporting the claim that strategy finding and belief updating in an uncertain environment remain largely intact under the tested psilocybin conditions. However, subconscious cues—particularly neutral faces presented briefly—produced a credible decrement in performance under psilocybin compared with placebo, suggesting that altered top-down suppression of task-irrelevant input or disrupted attentional tracking may underlie the effect. Casanova and colleagues situate these findings within mechanistic accounts of serotonergic modulation and predictive coding. They discuss how psilocybin’s partial agonism at 5-HT2A (and modulatory effects at other 5-HT subtypes) may shift the balance of top-down versus bottom-up signalling, relax overly precise priors and increase bottom-up prediction error flow. Such changes could increase exploratory switching (consistent with the borderline higher variance under psilocybin) while also attenuating aversive processing in ways analogous to chronic SSRI effects. The authors compare their findings to prior work with LSD and ketamine, noting differences in receptor profiles and downstream dopaminergic or glutamatergic effects that may explain discrepancies in learning-rate findings across studies. The discussion acknowledges several limitations reported by the authors. First, the complexity of the dataset and subgroup hierarchies prevented fitting a formal reinforcement-learning computational model, so inference relied on a Bayesian mixed-effects regression. Second, the learning task was administered ~240 minutes after dosing, a time when acute subjective effects may be partially waning (the authors state acute physiological changes were present at approximately 50%). Third, the probabilistic learning paradigm may not capture the broader notion of therapeutic cognitive flexibility that psilocybin might facilitate in clinical populations. The authors also note order effects—poorer performance when psilocybin was administered first—interpreting this as practical evidence that beginning an unfamiliar task in a sober state may aid subsequent performance under psilocybin. Finally, they comment that subjective psychometric reports of impaired vigilance and cognition under psilocybin contrasted with largely preserved objective performance on the task, and that monetary incentives may have been less salient under psilocybin which could have affected earnings outcomes.

The authors conclude that, across the dosages tested, psilocybin preserved the ability to find strategies and make decisions in an uncertain probabilistic learning task and was statistically noninferior to placebo, with the 20 mg dose showing superior learning compared to placebo in their sample. They emphasise that subconscious emotional cues diminished learning under psilocybin, a finding they attribute to shifts in top-down versus bottom-up cortical–subcortical signalling. The study is presented as the first test of psilocybin in a probabilistic reinforcement-learning paradigm; the authors call for further research to clarify the mechanisms underlying therapeutic learning effects and to optimise dosing and administration protocols for clinical use.