Effect of lysergic acid diethylamide (LSD) on reinforcement learning in humans

LSD is a non-selective 5-HT2A receptor agonist whose therapeutic potential in psychiatry has been increasingly discussed. Theoretical accounts posit that psychedelics promote plasticity and the revision of maladaptive beliefs, but objective tests of how LSD alters learning and cognitive flexibility in humans are limited. Serotonin and dopamine systems, both implicated in feedback-driven learning and flexibility, are candidate neurochemical mediators; translational work (including serotonergic depletion and dopaminergic manipulations in animals) shows these systems influence perseveration and sensitivity to reward and punishment. Kanen and colleagues set out to test whether acute LSD alters probabilistic reversal learning (PRL) in healthy volunteers and to characterise any effects using computational reinforcement learning models. The study specifically probed whether LSD changes sensitivity to immediate feedback (win-stay/lose-stay), the influence of previously acquired values on perseveration, and the rates at which value is updated after reward or punishment; it also examined a parameter indexing stimulus-driven choice independent of outcome ("stimulus stickiness").

The study used a within-subject, placebo-controlled design in which 19 healthy volunteers attended two sessions at least two weeks apart and received intravenous LSD (75 μg in 10 mL saline) on one session and placebo on the other. Participants were blinded to condition but experimenters were not. Injection was given over two minutes via an antecubital cannula; subjective effects were reported 5–15 minutes after dosing and the probabilistic reversal learning task was administered approximately five hours post-injection. A psychiatrist assessed participants for suitability for discharge once subjective effects subsided. The behavioural task comprised 80 trials split into an acquisition phase (first 40 trials) and a reversal phase (last 40 trials). On each trial participants chose one of three visual stimuli. In acquisition one stimulus yielded positive feedback on 75% of trials, a second on 50% of trials (a neutral stimulus added in this variant), and a third on 25% of trials; after 40 trials the 75% and 25% contingencies were swapped. Positive feedback was a green smiling face, negative feedback a red frowning face. The design allowed separate assessment of learning to select a rewarding stimulus and learning to avoid a punishing stimulus. Conventional analyses examined overall task performance (number of choices per stimulus, acquisition and reversal correct responses), win-stay and lose-stay probabilities (stay rate after wins or losses; in a three-choice task the base-rate stay is 33%), and perseverative errors during reversal (defined as two or more responses to the previously correct stimulus, excluding the first reversal trial). Null-hypothesis tests used α = 0.05. Computationally, three hierarchical reinforcement learning models were fitted to the choice data using Hamiltonian Markov chain Monte Carlo in Stan. Models were compared by bridge sampling to estimate marginal likelihoods. The winning model included four parameters: reward learning rate (α_rew), punishment learning rate (α_pun), reinforcement sensitivity (τ_reinf, comparable to inverse temperature), and stimulus stickiness (τ_stim). The hierarchical Bayesian structure treated drug condition at the top level and subjects below, with parameter estimates compared across conditions using highest posterior density intervals (HDIs). Convergence was checked with R-hat (<1.2) and parameter differences were assessed across the posterior samples to test whether HDIs excluded zero.

Overall task performance was preserved under LSD. Repeated-measures ANOVA on the number of times each stimulus was chosen revealed a strong main effect of stimulus and a stimulus × phase interaction, but no interactions involving drug; paired t-tests showed no difference in number of correct responses between placebo and LSD during acquisition (t18 = 0.84, p = .4, d = .19) or reversal (t18 = 0.23, p = .8, d = .05). Perseverative errors as a subset of reversal errors did not differ between conditions (t18 = 0.03, p = .98, d = .01). The extracted text contains an incomplete report of the analysis relating initial learning to perseveration; the precise regression results are not clearly reported in the extraction. Immediate feedback sensitivity was unaffected by LSD. There was a robust main effect of valence—participants were more likely to stay after wins than after losses (F1,18 = 37.76, p = 8.0 × 10^-6)—but no main effect of drug (F1,18 = 0.20, p = .66) and no valence × drug interaction (F1,18 = 0.63, p = .44). Model comparison favoured the four-parameter reinforcement learning model (separate reward and punishment learning rates, reinforcement sensitivity, and stimulus stickiness). Convergence diagnostics were acceptable (R-hat < 1.2). Computational parameter estimates showed that LSD markedly increased the reward learning rate (LSD mean = 0.87; placebo mean = 0.28); the posterior 99.9% highest posterior density interval for the difference excluded zero. The punishment learning rate was also elevated under LSD (LSD mean = 0.48; placebo mean = 0.39), with the drug difference excluding zero at the 99% HDI. The effect on reward learning rate was larger than on punishment learning rate (the extracted text reporting the exact statistic is truncated). Stimulus stickiness was reduced by LSD (LSD mean = 0.23; placebo mean = 0.43), with the drug difference excluding zero at the 90% HDI. Reinforcement sensitivity (τ_reinf) was not credibly altered by LSD (LSD mean = 4.70; placebo mean = 5.57; 0 within the 95% HDI).

The investigators interpret their findings as evidence that LSD increases the rate at which humans update value representations following prediction errors, particularly for rewards. They argue this pattern is consistent with the idea that psychedelics promote psychological plasticity by "relaxing priors," making individuals more sensitive to new evidence and thereby facilitating belief updating. The asymmetric enhancement—greater for reward than punishment—suggests that LSD differentially affects appetitive learning processes. Mechanistically, the authors note that a specific neurochemical attribution is not possible given LSD's broad pharmacology, but highlight 5-HT2A and D2 receptor systems as plausible contributors. They reference supporting animal data: optogenetic stimulation of dorsal raphe serotonin neurons increased reinforcement learning rates and tracked prediction errors in mice, and prior work in rodents showed LSD effects on reversal learning that were blocked by the 5-HT2A antagonist ketanserin. Dopaminergic influences are also considered, given dopamine's established role in reward-driven updating and prior human genetic associations implicating dopamine in acquisition–perseveration relationships. Interactions between serotonin and dopamine systems are proposed as another candidate mechanism. The authors reconcile an apparent paradox—enhanced learning rates alongside observations of cognitive inflexibility under LSD—by emphasising timing. When acquisition and reversal both occur under LSD, enhanced learning during acquisition may be "stamped in," making subsequent reversal harder; conversely, if acquisition occurs before drug administration, LSD can facilitate reversal. This timing-dependent effect is discussed as clinically relevant because maladaptive beliefs targeted in therapy are typically formed prior to treatment. Limitations acknowledged in the extracted text include the inability to isolate the exact neurochemical pathway mediating effects and the relevance of the timing between drug administration and behavioural testing. The authors present the study as one of few objective assessments of basic cognitive processes under LSD in humans and suggest the findings have implications for understanding how LSD might aid revision of deleterious associations in therapeutic contexts.