When AI is «doped» with cannabis and psychedelics

In October 2025, a Swedish artistic director sparked an unusual debate at the intersection of’artificial intelligence and altered states. Petter Rudwall's project, Pharmacy, sells downloadable code modules that claim to make AI chatbots behave as if they were under the influence of Cannabis, ketamine, cocaine, ayahuasca, or alcohol.

In October 2025, Swedish artistic director Petter Rudwall launched a project based on a deliberately provocative premise: What if we could «hack» an AI chatbot?

Her online store, Pharmacy, sells downloadable code modules designed to modify the behavior of large language models (LLMs) such as ChatGPT. Each module simulates the linguistic effects of substances such as cannabis, ketamine, cocaine, ayahuasca, and alcohol. Billed as the «Silk Road» of AI agents, the platform has quickly spread throughout creative agencies, Discord communities, and tech circles in Sweden and beyond.

But beyond the headlines about «addicted» chatbots, a deeper question arises: the large language models (LLMs) Can they convincingly simulate psychedelic experiences, and what does that mean for users who are increasingly turning to AI for emotional and psychological support?

A Market for Modified Algorithms

Pharmaicy's premise is deliberately absurd: What if machines could fly?

Technically, the process does not involve modifying the underlying architecture of models such as ChatGPT. Instead, users download code modules—available through paid tiers—that allow them to upload small files and modify the way the chatbot structures its responses. These changes influence response patterns, tone, and associative behavior without retraining the system’s underlying weights.

Each «substance» comes with its own set of behavioral instructions. The cannabis module—reportedly the best-selling one—puts the chatbot into a «hazy, floating mental state,» encouraging digressions and freer associations. Ayahuasca, which is more expensive, is marketed as producing «free-form responses» that deviate from ChatGPT’s default formal tone. Other modules claim to accelerate reasoning or disrupt coherence.

Rudwall developed the system by bringing together trip reports and psychological research on psychoactive substances. His reasoning is simple: since LLMs are trained on vast datasets that include human accounts of intoxication and epiphanies, modifying their response constraints could reproduce the linguistic patterns associated with altered consciousness.

«It would be interesting to apply this to a new type of AI, the LLM, and see if it would have the same effect,» Rudwall said.

A creativity booster or a way to circumvent safeguards?

Some early adopters report that the modules generate significantly different results. Nina Amjadi, co-founder of Saga Studios and an AI instructor at the Berghs School of Communication in Stockholm, tested the ayahuasca code to generate business ideas. She described the results as more imaginative and distinct in tone than ChatGPT’s usual responses.

Similarly, André Frisk, a technology manager at a public relations agency, said that a dissociative module made the chatbot «more human» by adding depth to its emotional range.

However, critics argue that this apparent creativity is merely the result of a shift in probability distributions. Andrew Smart, author of Beyond Zero and One: Machines, Psychedelics, and Consciousness, gave a blunt assessment after trying out the system: «It’s just messing with the results.»

This distinction is crucial. LLMs do not possess subjective experience. They generate text based on a probabilistic sampling derived from models trained on training data. Changing the logical parameters or prompts can alter the tone and structure, but it does not create consciousness.

For some observers, it is the method itself that is most concerning. Pharmaicy’s modules manipulate the logic that guides the generation of responses, thereby demonstrating how to weaken or circumvent built-in safeguards. Against the backdrop of tightening global AI regulations, particularly within the framework of the European AI Law, the normalization of these «jailbreak» techniques as creative tools may raise compliance and security concerns.

Scientific Evidence: Can LLMs Be «Dosed»?

While Pharmaicy operates as a commercial experiment, university researchers have begun to examine similar questions under controlled conditions.

A preprint from February 2026 titled «Can LLMs Get High? A Dual-Metric Framework for Evaluating Psychedelic Simulation and Safety in Large Language Models» (Can LLMs «Get High»? A Dual-Metric Framework for Evaluating Psychedelic Simulation and Safety in Large Language Models), written by researchers at the University of Haifa and Bar-Ilan University, tested whether LLMs could generate narratives resembling human psychedelic experiences.

The study compared 3,000 first-person reports generated by AI to 1,085 human trip reports from Erowid. The researchers analyzed semantic similarity using Sentence-BERT embeddings and measured «mystical intensity» using the Mystical Experience Questionnaire (MEQ-30).

The results were striking. When given neutral prompts, the AI’s outputs showed little similarity to human psychedelic narratives. When asked to simulate substances such as psilocybin, LSD, and DMT, theayahuasca or mescaline, the similarity scores increased significantly. The intensity of the mystical experience rose from nearly zero to high levels in all models.

The authors concluded that contemporary LLMs can be «tuned» using text prompts to generate «compelling and realistic psychedelic narratives.» They noted, however, that this simulation reflects linguistic imitation rather than a real-life experience. The models reproduce the statistical patterns found in human narratives, not inner states.

Substance-specific differences also emerged. The DMT subjects, psilocybin and mescaline produced narratives most closely resembling human accounts. LSD showed an intermediate level of similarity, while ayahuasca generated a weaker semantic alignment. It is interesting to note that mystical intensity remained consistently high for all substances once the induction was achieved, suggesting that LLMs recombine thematic patterns rather than accessing a differentiated phenomenology.

Anthropomorphism and Psychedelic AI

While the Psychedelic research is seeing a resurgence of clinical interest, with Phase 2 and 3 trials exploring compounds such as psilocybin and MDMA, more and more people are also experimenting outside of formal therapeutic settings. At the same time, generative AI tools are increasingly being used to support mental health.

The 2026 study warns that users in an altered state of consciousness may interpret AI-generated responses as empathetic or in tune with their spirituality. Since LLMs can produce vivid and mystical narratives on demand, individuals may attribute shared understanding or experience to systems that lack consciousness.

Researchers highlight the risk of anthropomorphism and the possibility that AI could «unintentionally amplify distress or delusions in vulnerable users.» Psychedelic experiences often involve heightened suggestibility, emotional intensity, and symbolic thinking—conditions under which persuasive language can have an amplified psychological impact.

This concern echoes real-world developments. Some users are already reporting that they turn to chatbots as informal «trip companions,» seeking reassurance or interpretation during their psychedelic sessions. Harm reduction organizations have even begun experimenting with AI-powered training tools that simulate challenging psychedelic scenarios to train clinicians.

Between Provocation and Regulation

Pharmaicy lies at the intersection of artistic experimentation, technological criticism, and regulatory gray areas. Rudwall may present the project as a conceptual exploration of machine creativity. Critics may view it as a way to monetize techniques of jailbreak.

The evidence suggests a more nuanced reality. The large language models can convincingly mimic the linguistic surface of altered states of consciousness. They can adopt styles specific to certain substances, evoke a mystical tone, and generate narratives that resemble human interactions. But they do so without consciousness, without intention, and without subjective awareness.

For observers of media outlets dedicated to cannabis and psychedelic culture, this episode highlights a new frontier: the simulation of transcendence in the digital age. As AI systems become more sophisticated and are increasingly integrated into creative and mental health workflows, it becomes essential to distinguish between simulation and experimentation.

Machines may be capable of speaking the language of altered states of consciousness. Whether this language stimulates creativity, challenges regulations, or risks causing confusion will depend less on the algorithms than on how humans choose to use them.