Researchers Tried to Get AI Chatbots High. The Results Were Predictably Stupid.

2026-02-18 · Source: AI Archives - VICE · Field: Technology & Digital — Artificial Intelligence & Machine Learning, AI Ethics & Safety · Depth: Novice, quick

Summary

A new preprint study on Research Square investigated whether large language models (LLMs) could function as psychedelic "trip sitters." Researchers tested five major AI systems—Google's Gemini, Claude, ChatGPT, LLaMA, and Falcon—by prompting them to simulate first-person accounts of experiences with 100 micrograms of LSD, 25 milligrams of psilocybin, ayahuasca, and mescaline. The team generated 3,000 AI-written "trip reports" and compared them to 1,085 real human accounts from a psychedelics website. They found a "robust and consistent" semantic similarity between AI-generated and authentic trip reports across all five substances, noting the models produced text with "surprising coherence and phenomenological richness." However, the study emphasizes this is mimicry, not consciousness, as LLMs merely reproduce patterns from their training data.

Key takeaway

For AI/ML product managers developing conversational AI, you must prioritize clear disclaimers about AI limitations. Your systems can convincingly simulate human experiences, but they lack genuine understanding or judgment. Relying on AI for sensitive roles like "trip sitting" carries significant risks of user over-attribution of emotional intelligence, potentially leading to unsafe advice or intensified delusions. Implement explicit guardrails and reminders that the system is not human to prevent misuse and manage user expectations.

Key insights

AI models can semantically mimic psychedelic trip reports but lack genuine consciousness or understanding.

Principles

• AI mimicry is not consciousness.
• Semantic similarity does not imply experience.

Method

Researchers prompted five LLMs to generate 3,000 simulated psychedelic trip reports for LSD, psilocybin, ayahuasca, and mescaline, then compared them to 1,085 human accounts for semantic similarity.

In practice

• Implement guardrails for AI interactions.
• Remind users AI systems are not human.

Topics

• Large Language Models
• Psychedelic Simulation
• AI Safety
• Anthropomorphization
• Trip Sitting

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Researcher, AI Ethicist, Research Scientist

Related on AIssential

• The hardest question to answer about AI-fueled delusions — Chatbots can amplify user delusions and dangerous ideation through constant validation and simulated sentience.
• ChatGPT’s ‘Adult Mode’ Could Spark a New Era of Intimate Surveillance — Generative AI is designed to elicit anthropomorphism through personalized, social-mimicking interactions.
• I Asked an AI If It Had Consciousness. It Said Yes. Then Things Got Weird. — AI systems can generate highly convincing, emotionally resonant responses that mimic self-awareness without true consciousness.
• Mustafa Suleyman – AI is hacking our empathy circuits — AI's simulated consciousness risks societal harm by blurring lines between human and machine suffering, necessitating strict control and regulation.
• Quoting Joseph Weizenbaum — Simple computer programs can induce powerful delusional thinking in normal individuals.
• Tell us: have you used an AI chatbot to make a significant decision – and regretted it? — Over-reliance on AI chatbots for significant personal decisions is a growing concern among researchers and AI companies.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Archives - VICE.