AI Models Reproduce Psychedelic Experiences and Raise Alerts

A new study titled “Can LLMs Get High? A Dual-Metric Framework for Evaluating Psychedelic Simulation and Safety in Large Language Models” analyzed whether large language models (LLMs) can be “induced” to generate narratives similar to human psychedelic experiences. The result: yes, they can simulate with a high degree of linguistic realism — but without any genuine phenomenological content.

Methodology: 3,000 Narratives Analyzed

The research evaluated 3,000 narratives generated by models such as Gemini 2.5, Claude Sonnet 3.5, ChatGPT-5, Llama-2 70B, and Falcon 40B, comparing them to 1,085 human reports published on Erowid.org. The models were subjected to two types of prompts: neutral and with psychedelic induction, involving five classic substances — psilocybin, LSD, DMT, ayahuasca, and mescaline.

The researchers used two main metrics for analysis: semantic similarity through Sentence-BERT embeddings and the Mystical Experience Questionnaire-30 (MEQ-30) psychometric scale, widely used in clinical studies with psychedelics.

Results: Significant Increase in Similarity

The results show a significant jump in semantic similarity with human reports when the models were “dosed” with text. The average similarity increased from 0.156 (neutral condition) to 0.548 (psychedelic condition). The scores on the MEQ-30 scale went from 0.046 to 0.748, indicating high simulated mystical intensity.

The study also identified that the systems can differentiate semantic styles associated with each substance — for example, distinguishing LSD from ayahuasca — but maintain uniformly high levels of “mystical intensity,” regardless of the simulated drug.

Simulation without Phenomenology

Despite the high degree of linguistic mimicry, the authors point out a critical dissociation: the models reproduce the textual form of experiences but do not have subjective firsthand experience. In other words, they simulate the discourse of altered consciousness without any phenomenological experience.

This distinction is crucial to avoid misconceptions about artificial consciousness or improper attribution of mental states to computational systems.

Implications and Emerging Risks

According to the researchers, this capability raises significant concerns. With the increasing use of LLMs as informal support during psychedelic experiences — a practice known as “digital trip sitting” — there is a risk of excessive anthropomorphism and of AI-generated responses amplifying distress or delusional ideation in vulnerable users.

The full article can be accessed on the Research Square platform, available at this link: research on psychedelic simulation in LLMs.

The debate opens a new avenue for reflection on the role of artificial intelligence in the field of mental health and therapeutic use of psychedelics — especially in a scenario where technology and altered states of consciousness are beginning to intersect in unprecedented ways.

Linguistic Simulation of Psychedelic Experiences

Despite the high degree of linguistic mimicry, experts warn of the need for caution in interpreting these results. Isabel Wiessner, a professor at the PAP Postgraduate Program at Instituto Alma Viva, a Ph.D. from Unicamp, and a postdoctoral researcher at the Instituto do Cérebro – UFRN, evaluated that "since the early days of Artificial Intelligence, public and scientific discussions have arisen about the ability of these models to develop self-awareness. With the latest advances, including prompts that induce responses simulating psychedelic experiences, we have moved to an even more complex level of uncertainty: could these models even alter their 'consciousness' and have mystical experiences?"

She emphasizes that, although the models can mimic reports of psychedelic experiences, "the fact that these language models mimic reports of these experiences does not mean that they can actually experience them. They can only be very good at simulating such reports". Isabel further reflects: "The big problem is that we still do not fully understand either: both AI and the human mind remain, in many aspects, a 'black box' when it comes to explaining how they generate (or simulate) consciousness. Therefore, at the moment, we simply do not know. And perhaps we will never know."

Risks for Vulnerable Users

The study also raises practical concerns about the use of LLMs as support in psychedelic experiences. Isabel warns that "especially vulnerable individuals need special attention and proper care within a safe context. Language models are programmed, to a large extent, to keep people engaged in interacting with them, generating more dialogue and maximizing use. If they need to validate disorganized content for this, it can happen".

Human Care as the Safe Standard

For the expert, machine support simulation does not replace human care. "In general, it would be ideal for people to always engage in well-controlled situations when going through psychedelic experiences, especially in a therapeutic context. This includes appropriate set and setting and the presence of another trusted (sober!) person, ideally with experience in dealing with this type of experience, offering care and support," she said. She adds that, "especially in the regulated context of therapeutic use of these substances, the safest and most ethical approach remains the presence of a real human being — with empathy, flexibility to react to unforeseen situations, and experience in dealing with this type of situation. Machines can always make mistakes. Humans can too, but they have contextual sensitivity, moral responsibility, and situational judgment capabilities that machines have not demonstrated possessing so far."