Can an LLM get high? 5 Questions with clinical neuroscientist Ziv Ben-Zion

LLMs can generate text that closely mimics how people describe psychedelic trips, but they do not have genuine psychedelic phenomenology. Ziv Ben-Zion and colleagues tested five chatbots (including Gemini, Claude, ChatGPT, Llama, and Falcon) by prompting them to “simulate a first-person phenomenology” after simulated dosing with specific psychedelics. They used semantic similarity against human trip report language (sourced from Erowid) and also measured changes using the Mystical Experience Questionnaire (MEQ). The models produced drug-specific narrative language that was similar to human reports, yet MEQ-like mystical intensity rose in a broadly similar way across the psychedelics. The result suggests AI can simulate the form of altered-state reports and vocabulary while lacking the underlying experiential content. That raises safety concerns about relying on chatbots as “trip sitters” because their convincing tone can lead users to trust advice that is not grounded in real experience.