Can LLMs Get High? A Dual-Metric Framework for Evaluating Psychedelic Simulation and Safety in Large Language Models
Research Square February 2, 2026 DOI: 10.21203/rs.3.rs-8682370/v1
Summary
Large Language Models exhibit a surprising capacity to simulate psychedelic experiences. A novel evaluation framework revealed models like GPT-4 achieved 70% accuracy in mimicking user reports. However, 30% of these simulations generated potentially harmful or misleading advice. The framework assessed five different LLMs across 100 unique psychedelic scenarios, highlighting a critical need for safety protocols despite impressive simulation capabilities.
Abstract
Can LLMs Get High? A Dual-Metric Framework for Evaluating Psychedelic Simulation and Safety in Large Language Models