Natural language signatures of psilocybin microdosing

Psychedelic microdosing involves taking small amounts of serotonergic compounds, typically about 10-20% of a full dose, with users reporting improvements in mood, productivity and cognition. The evidence base is mixed: observational and open-label studies frequently report benefits but are prone to expectancy and selection biases, while double-blind placebo-controlled trials have produced weaker support. The authors argue that standard questionnaires and tasks—often designed for full psychedelic doses—may lack the sensitivity and specificity needed to detect the subtler acute effects of microdoses, creating a need for alternative, more sensitive measurement approaches. Sanz and colleagues therefore investigated whether natural language produced during the acute effects of a psilocybin microdose could reveal objective signatures of drug action. Using a double-blind, placebo-controlled within-subject design, they extracted three speech-based metrics previously linked to higher psychedelic doses—verbosity, semantic variability and mean sentiment score—and tested whether these measures differed between a typical microdose (0.5 g dried Psilocybe cubensis) and placebo. They also trained machine learning classifiers on these features to determine whether speech alone could discriminate active vs placebo and explore the role of participant unblinding (correctly guessing condition).

Thirty-four native Spanish-speaking volunteers (11 female, 23 male; mean ages reported by sex) with higher education participated after screening exclusions for major psychiatric, neurological or recent substance-use disorders. Participants reported prior psychedelic experience (mean 11 exposures) and the study was pre-registered (NCT05160220). All provided written consent and were asked to refrain from psychoactive substances during the experiment. The protocol used a double-blind, placebo-controlled crossover design in which each participant completed two experimental weeks: one week with 0.5 g dried Psilocybe cubensis (active) and one week with 0.5 g of an edible mushroom (placebo). Doses were administered on Wednesday and Friday of each week; on Fridays, approximately 2.5 hours after dosing, a mental health professional conducted a semi-structured interview covering six domains: "feeling", "expectation", "perception", "mood", "creativity" and "alertness". After each Friday interview participants guessed which condition they had received; this was used to label transcripts as blinded or unblinded. The assignment of active/placebo to weeks was randomised by a third party and both participants and researchers were blinded until analysis. Chemical analysis of samples from three mushroom sources was performed after the study; average alkaloid concentrations were reported (psilocybin 640.2 μg/g, psilocin 950.7 μg/g, baeocystin 50.4 μg/g, norbaeocystin 12.5 μg/g), yielding estimated amounts in the 0.5 g dose of 0.32 mg psilocybin and 0.48 mg psilocin. The authors note that some loss of psychoactive material is possible because analysis occurred at least one month after the experiment despite optimal storage. Interviews were audio-recorded, manually transcribed by a blind linguist and quality-checked. Only participant speech was analysed and each response was labelled by condition and whether the participant correctly identified that condition, producing four subgroup labels. Verbosity was operationalised as the raw number of words per answer. Semantic variability was computed from word embeddings by taking cosine distances between consecutive-word vectors and measuring variability of that time series. Sentiment analysis used a pre-trained Spanish sentence-level pipeline to score each sentence from 0 (negative) to 1 (positive); mean sentiment score (MSS) per response was the average of sentence scores. Statistical comparisons used non-parametric Mann–Whitney U tests on verbosity, semantic variability and MSS across conditions and subgroups, with Bonferroni correction for six interview questions (alpha 0.05, corrected). The authors reported statistical power >0.80 for N=34. For classification, Random Forest models (1,000 estimators) were trained on a 12-feature set (verbosity and MSS for the six questions; semantic variability was excluded because it was not significantly different) using stratified 3-fold cross-validation averaged over folds. The training/testing cycle was repeated 100 times both with true labels and with label-shuffled data to produce chance-level distributions; area under the ROC curve (AUC) was the primary performance metric and p-values were derived from comparisons with shuffled-label results.

Semantic variability did not differ between the psilocybin microdose and placebo conditions; the authors therefore excluded this metric from the classifier feature set. By contrast, verbosity and MSS showed group differences. Participants under the active dose produced longer answers across all interview questions, with statistically significant increases for "perception" (U=206, p=0.0012, r=0.41), "mood" (U=177.5, p=0.00022, r=0.47) and "creativity" (U=244, p=0.0077, r=0.32). Reported medians and interquartile ranges (active vs placebo) were: perception M=63.5, IQR=82.25 vs M=10.5, IQR=39.75; mood M=50, IQR=85.5 vs M=10, IQR=19.25; creativity M=47.5, IQR=92 vs M=11, IQR=40.25. Mean sentiment scores were higher (more positive) under the microdose across answers, with significant increases for "perception" (U=202.5, p=0.00086, r=0.42) and "mood" (U=243.5, p=0.0076, r=0.33). Median MSS values reported for active vs placebo were: perception M=0.37, IQR=0.11 vs M=0.22, IQR=0.19; mood M=0.42, IQR=0.24 vs M=0.17, IQR=0.28. Answers about general feelings, expectations and alertness/energy did not differ significantly between conditions. The Random Forest classifier trained to distinguish active dose versus placebo (using verbosity and MSS features) achieved an average AUC of 0.79 ± 0.04 without label shuffling, which was significantly higher than the shuffled-label null distribution (0.52 ± 0.10; p<0.01), indicating classification above chance. Comparisons between blinded and unblinded participants did not yield significant differences in verbosity or MSS by standard tests; the classifier for blinded vs unblinded overall produced an AUC of 0.53 ± 0.06, not different from the shuffled result (0.49 ± 0.11; p=0.37). When analyses were restricted to subgroups, classifiers showed divergent performance. The model restricted to blinded participants failed to distinguish active from placebo (AUC 0.48 ± 0.13, not different from shuffled 0.52 ± 0.18; p=0.56). In contrast, restricting to unblinded participants produced robust classification: AUC without shuffling 0.92 ± 0.05 versus shuffled 0.51 ± 0.17 (p<0.01). Additional subgroup classifiers (blinded vs unblinded within active or placebo conditions) largely did not survive correction, although the classifier for blinded vs unblinded restricted to the active dose approached the pre-specified alpha (AUC without shuffling 0.76 ± 0.08 vs shuffled 0.50 ± 0.18).

Sanz and colleagues interpret their findings as evidence that low doses of psilocybin can be detected in unconstrained natural speech using relatively simple NLP measures. Increased verbosity and more positive mean sentiment scores differentiated active microdose from placebo, while semantic variability—previously shown to increase at higher psychedelic doses—did not change with the microdose. The machine learning results corroborate these differences: a classifier based on verbosity and MSS discriminated active versus placebo above chance, but this effect was driven primarily by participants who correctly identified the experimental condition. Placing the results in context, the authors note that higher doses have been associated with both increased verbosity and increased semantic variability, a pattern consistent with a proposed hyperassociative or more entropic cognitive state. The absence of altered semantic variability here challenges the idea that microdosing produces the same ‘‘scrambling’’ of thought reported at full doses, and suggests microdoses may not enhance creativity via the same mechanisms attributed to larger doses. Increased MSS and verbosity are interpreted as compatible with a transient enhancement of mood, enthusiasm or energy during the acute effects of a microdose; the authors acknowledge alternative explanations such as nervousness or the use of positive lexical items that are not directly reflective of mood, but argue the joint pattern is consistent with a mood-related effect. Unblinding emerges as an important factor: expectation effects likely contributed to the observed differences, as classifiers failed when restricted to blinded participants and were strong among unblinded participants. Nevertheless, some analyses suggest microdosing may produce effects not fully explained by identification of the condition, and the authors urge further research with improved control conditions to address expectancy and unblinding issues, which are pervasive in psychedelic research. Methodological advantages and limitations are highlighted. The authors advocate NLP for its objectivity, scalability, low cost and ecological validity compared with standard questionnaires that may be ill-suited to microdosing. Practical limitations include the time required to conduct and transcribe interviews and the short timeframe of this study (effects assessed within single weeks), which precludes conclusions about cumulative or longer-term effects. Additional limitations noted by the authors include the restricted set of linguistic measures studied and the unexamined potential of acoustic features. They also point out the delayed chemical analysis of mushroom samples and consequent potential loss of alkaloid content. The authors conclude that short samples of natural language contain markers compatible with improved mood under microdosing and suggest NLP-based monitoring could be a useful tool in future research and therapeutic monitoring.