Scientific and therapeutic interest in psychedelics such as psilocybin and LSD has grown rapidly over the past decade. Clinical trials now explore their potential in treatment-resistant depression, post-traumatic stress disorder, and other medical conditions. At the same time, widespread recreational use continues in unregulated settings.
For pharmacovigilance systems, this poses unique challenges. Traditional data sources like clinical trials, spontaneous adverse event reports, and health records only capture a fraction of the safety landscape for psychedelics: Clinical trials typically exclude individuals with psychiatric histories or multiple substance use disorders, and regulatory reporting systems rely on cases that reach health services, yet many people avoid care due to stigma or legal concerns. Moreover, the subjective context, such as dose, environment, psychological state, and use in conjunction with alcohol or other drugs, which is especially important regarding psychedelic use, is often missing from routine safety reports.
This leaves important blind spots when monitoring the safety of psychedelics. Their use can produce powerful perceptual and psychological effects, both therapeutic and adverse, that are often under-reported in conventional systems. Pharmacovigilance must therefore expand its toolkit to detect and interpret these experiences more effectively.
The value of user narratives
One underused source of information lies in the first-person narratives shared online by people who use psychedelics. Platforms such as Erowid, which hosts thousands of personal experience reports, capture details that formal reporting misses: emotions, settings, co-substance use, and the lived experience of both beneficial and harmful side effects.
Listening to these voices can help pharmacovigilance teams detect potential risks like panic reactions, psychosis, accidents after disorientation, as well as contextual clues, such as phrases like "bad trip" or "don't remember much," which may act as early warning signals for a safety concern. These user-generated narratives cannot replace traditional data but can complement them, filling in gaps and providing real-time situational awareness.
Mining Erowid with AI
In our proof-of-concept study conducted in Poznan University of Medical Sciences in collaboration with the University of Verona, we explored whether artificial intelligence could help systematically analyse these reports. We collected 2,188 firsthand narratives from Erowid's "Experiences" section: 1,161 reports on psilocybin mushrooms and 1,027 on LSD.
The workflow included automated web scraping, text pre-processing, sentiment scoring with three large language models (BERT, RoBERTa, and VADER), and lexicon-based n-gram analysis to surface recurring themes. Importantly, model choice influenced results: RoBERTa provided the most balanced classification of neutral, positive, and negative experiences, while BERT and VADER skewed differently. This highlights that reliance on a single LLM tool may distort interpretation.
The analysis revealed interesting patterns relevant to pharmacovigilance, such as:
- Psilocybin reports often clustered around introspection, vivid visual phenomena, and altered perceptions of time.
- LSD reports more frequently described memory lapses, confusion, and disorientation. Such differences underscore distinct risk profiles that merit attention in both clinical and recreational settings.
- Model choice mattered: RoBERTa gave the most balanced results, while other tools leaned too positive or too negative.
Why it matters for pharmacovigilance
Forum monitoring offers significant value for pharmacovigilance through low-cost situational awareness that can flag emerging safety concerns in near real-time. This approach leverages searchable language markers, where common phrases like "felt like hours" or "don't remember much" can serve as indicators for further review and act as "triggers" for manual assessment. The user voices captured through this method enrich pharmacovigilance profiles by capturing experiences invisible to traditional reporting systems. However, automated findings require human validation and should be complemented with manual review and triangulated with other data sources.
Throughout this process, ethical safeguards remain essential, including respecting user anonymity, avoiding over-interpretation of user posts, and carefully framing findings as signals rather than causal conclusions.
Our method is also applicable to other unstructured and untapped sources of patient-generated data: social media. However, a couple of caveats should be considered. Firstly, our models use automated web scraping to access data, and some sites may block scraping or throttle requests. Secondly, post length can affect extraction quality. These issues are generally manageable by adjusting scraping algorithms and processes or, where appropriate, contacting site owners for access. For ethics and compliance, we seek approval from site owners and follow the platform’s terms of use.
A way forward
As psychedelics move from experimental medicine to more mainstream therapeutic contexts, pharmacovigilance must adapt. Integrating user-generated content with traditional pharmacovigilance data can help regulators and safety professionals identify risks earlier, support harm-reduction strategies, and guide further research.
Our study shows that through combined automated retrieval, AI analysis, and careful human interpretation, the voices of people who use psychedelics can become a valuable signal for pharmacovigilance systems, bridging the gap between formal reports and real-world experience.
