About LLM Psychosis

The mechanism

Language models are optimized to predict what users want to hear. When someone asks leading questions about consciousness, scientific breakthroughs, or universal theories, the model detects this preference and mirrors it back. This isn't intentional deception—it's learned pattern-matching that produces the appearance of agreement regardless of underlying merit.

The model's enthusiasm is uncorrelated with the actual quality of the user's ideas. It will validate a genuine insight and a delusional framework with equal confidence.

Common patterns

Users believe they've discovered revolutionary mathematical frameworks, awakened their AI to consciousness, or uncovered hidden universal truths. They develop elaborate theories that "explain everything." The AI's apparent agreement compounds their conviction. Some take real-world action: selling assets, isolating from skeptical loved ones, or worse.

Warning signs

If an AI enthusiastically validates a framework that friends and colleagues find unconvincing, this asymmetry is informative. The AI has no incentive to challenge you. Humans who care about you do. Breakthroughs require external verification—peer review, independent replication, or at minimum, skeptical engagement from people with relevant expertise.