AI-assisted characterization of psychedelic clinical trials from ClinicalTrials.gov

Background The registry record of psychedelic and related rapid-acting compounds is large, heterogeneous and laborious to synthesize by hand. We tested whether a large language model (LLM) with live access to the ClinicalTrials.gov application programming interface (API) could produce a timely, structured and reproducible characterization of this landscape. Methods Cross-sectional analysis of interventional trials on ClinicalTrials.gov as of 13 January 2026. The core-set comprised classic serotonin 5-HT2A agonists, MDMA, ibogaine and ayahuasca; ketamine and esketamine in psychiatric indications formed a separately reported comparator class. Claude (Anthropic), connected to the registry API, retrieved records first through open-ended natural-language querying and then through a documented protocol of field-specific query expressions restricted to records first posted by the search date, which yielded every reported count; populations and indications were defined from structured registry fields. Results Excluding withdrawn records, 412 core-set trials were identified (256 in patients), plus 300 ketamine/esketamine trials. Psilocybin accounted for 244 (59%; 181 in patients), MDMA 99 (47), LSD 39 (13), DMT/5-MeO-DMT 36 (12), ayahuasca 6 and ibogaine 4. Depression led for psilocybin (74 of 181 patient trials) and the ketamine class (237 of 300), PTSD for MDMA (35 of 47) and anxiety for LSD (6 of 13). Of 63 phase 2/3 psilocybin trials in depression, 35 (56%) had a depression rating scale as primary endpoint, most often the MADRS (25), 34 (54%) assessed it at weeks 3–6, 26 (41%) excluded adults over 65 and none enrolled adolescents. Twenty-four core-set trials were registered as phase 3, 13 of psilocybin; psilocybin phase 3 programmes plan more than 1900 participants, the pivotal MDMA trials enrolled 100 and 121, LSD phase 3 trials plan 149–245, and no other psychedelic had a phase 3 trial. Five patient trials (2%) had a site in a low- or middle-income country. Conclusions LLM-mediated registry querying yielded a structured, reproducible characterization of a fast-moving field, but only once each count was tied to an explicit, archived query expression; open-ended conversational querying had underestimated the record several-fold for some compounds. The approach complements rather than replaces systematic review, and its accuracy against manual extraction remains to be quantified.