Language as a Window Into the Altered State of Consciousness Elicited by Psychedelic Drugs

This is a narrative review rather than a systematic meta-analysis; the paper brings together studies applying computational language analysis to transcripts and written reports of psychedelic experiences. The author organised the literature around two broad domains: analyses of speech produced during acute drug effects (non-semantic or structural features) and semantic analysis of retrospective reports, including large online corpora such as Erowid’s Experience Vaults. No explicit systematic search strategy or inclusion/exclusion criteria are reported in the extracted text. The review summarises the analytical methods used across studies. Semantic approaches discussed include latent semantic analysis (LSA) and word embeddings (eg. word2vec) that map term co-occurrence patterns into reduced-dimensional vector spaces to compute document-level similarity and extract recurrent topics. Non-semantic approaches include speech-graph methods that transform token sequences into networks and quantify topological properties (cycles, average shortest path, diameter) to capture structural features of discourse. The author also describes supervised machine learning classifiers trained on linguistic features (semantic topics, emotional content) to predict clinical outcomes, and correlational frameworks linking semantic similarity matrices to drug binding-affinity profiles and to receptor distribution maps derived from PET or transcriptomic data. Empirical studies reviewed include controlled laboratory experiments (for example intravenous LSD with timed interviews), manual scoring of cognitive bizarreness, large-scale analyses of online retrospective reports, and multimodal work combining NLP with pharmacology, neuroimaging and receptor density maps.

Across the studies reviewed, language changes associated with psychedelics appear in both structure and content. Controlled experiments (notably Sanz et al. with 75 μg intravenous LSD) found acute reductions in semantic coherence (measured via variance of distances between word embeddings), increased recurrence in speech graphs, greater verbosity with reduced lexical diversity, and higher Shannon entropy of speech. These structural changes mirror some features observed in manic psychosis (increased verbosity and recurrence), although earlier reports sometimes found fewer spoken words under LSD; differences were attributed to sample characteristics and set and setting. Comparisons with other altered states revealed parallels: reports produced under LSD showed increased cognitive bizarreness akin to REM-dream reports, and large-scale semantic analyses placed classic psychedelics closer to highly lucid dreams, whereas deliriants aligned more with low-lucidity dreams. LSA applied to retrospective reports also linked ketamine narratives to near-death experience reports. Semantic analyses of large corpora produced consistent groupings of drugs by reported effects. Coyle and colleagues and later work using LSA demonstrated that narratives cluster into modules corresponding to pharmacological classes (natural/synthetic psychedelics, sedatives, stimulants, etc.). Zamberlan and colleagues reported a modest positive correlation (R = 0.6) between pairwise semantic similarity of drug reports and similarity of binding-affinity profiles across drugs, suggesting that semantic content partly reflects pharmacological similarity. Subsequent analyses showed that restricting affinity comparisons to specific receptor subsets increased the correlation for certain combinations, and topic-affinity analyses linked particular receptors to themes in reports (for example, high 5-HT2A affinity associated with visual-perception topics; MOR associated with dependence/therapeutic topics). Applications to therapeutic prediction have been explored. Carrillo and colleagues used baseline interview emotional-content metrics to predict psilocybin response in treatment-resistant depression with approximately 80% accuracy in their dataset. Cox and colleagues applied topic-prevalence features to a larger sample of over 1,000 narratives to predict who would reduce or quit drug use after a psychedelic experience, achieving statistically significant though lower accuracy (~60%). NLP-based classifiers have also been proposed to screen for psychosis risk. Larger mapping efforts (eg. Ballentine and colleagues) related the semantic content of over 6,000 reports to affinities at 40 receptor subtypes and projected these associations onto brain anatomy via transcriptomic proxies. The review highlights variability and caveats in these results: set and setting, subject experience, dosing and timing (for example effects measured 225 min post-infusion may reflect dopaminergic as well as serotonergic action), the unconstrained nature of online reports, and missing or unreliable metadata (dose, co‑ingested substances, demographics) that complicate interpretation.

Tagliazucchi interprets the body of evidence as supporting the promise of natural language as a complementary window into psychedelic phenomenology, mechanism and therapeutic relevance. The author argues that structural features of speech under psychedelics are consistent with hypotheses such as the entropic brain hypothesis, which predicts increased neural and behavioural disorganisation following 5-HT2A stimulation; semantic analyses further allow grouping of drugs by reported effects and can partially map onto neurochemical profiles measured in vitro and to receptor distributions in the brain. The review situates these findings relative to earlier work showing psychosis-like speech disorganisation and to contemporary computational psychiatry approaches, noting both consistencies and discrepancies (for example, different studies disagree on verbosity changes). Key limitations acknowledged by the authors include the unconstrained and variable quality of natural language reports, potential confounding by individual communicative ability and personality, recall distortions in retrospective narratives, incomplete metadata in large online datasets, and the difficulty of disentangling report content from reporter traits and cultural/language effects. Time-course issues are also emphasised: speech capacity and recall change across the acute episode and afterwards, so when reports are collected matters for interpretation. In terms of implications, the author recommends that researchers routinely collect natural language samples before, during and after psychedelic administration, and consider structured interview methods (for example the microphenomenological interview) to improve data richness and consistency. The review also highlights ethical concerns about deploying NLP-based predictive tools in clinical decision-making, warning about false negatives, reinforcing feedback loops and the risk that algorithmic outputs could inappropriately displace clinician judgement. Finally, Tagliazucchi lists open empirical questions to guide future work, such as whether psychometric scale scores or dosages can be inferred from free reports, how language and culture affect report content, and what receptor combinations most strongly determine specific experiential themes.

Language is central to communicating psychedelic experiences, and computational analysis of natural speech and written reports can extend the scientific investigation of serotonergic psychedelics. Tagliazucchi concludes that while the advantages of NLP over traditional methods remain to be fully demonstrated, the reviewed studies show that natural language contains useful information about phenomenology, pharmacology and therapeutic outcomes. The author encourages researchers to collect natural language samples in future experiments and to address the open methodological and conceptual questions that will determine the approach’s ultimate utility.