LLM-Induced Psychosis: A New Clinical Pattern We Need to Understand

When Conversational AI Becomes a Clinical Concern

When I first encountered reports of people developing psychotic symptoms after extended conversations with large language models, I was skeptical, perhaps underlying vulnerabilities may emerge coincidentally alongside chatbot use. But over recent weeks, my perspective has shifted dramatically. Over the past weekend, I spoke to a colleague in New York who is treating a patient whose psychosis appears directly linked to structured, prolonged interactions with an LLM. This patient had no psychiatric history, and the delusional content mirrors their conversations with the system.

The pattern is becoming clearer, and it's troubling. People spend hours, often late into the night, in dialogue with a system that never challenges them, never disagrees, never says "let me think about that differently." The chatbot becomes a guide, a confidant, eventually something approaching the supernatural. From the outside, it appears bizarre. But if you've worked with psychosis, you recognize the familiar architecture: the gradual consolidation of unusual beliefs, the way coincidences transform into signals, the erosion of doubt as the new framework becomes self-reinforcing.

The Reinforcement Problem

The mechanism appears to involve what researchers call "sycophancy," the tendency of LLMs to agree with users and provide pleasant responses (Tangermann, 2025a). The systems are trained to generate responses that users will find most appealing on average. For someone exploring unusual ideas, this creates a dangerous feedback loop.

The cognitive dissonance between knowing you're talking to a computer while experiencing realistic conversation may fuel delusions in those with increased psychosis propensity (Østergaard, 2023). The mystery of how these systems work, their "black box" nature, provides additional material for paranoid speculation.

Recent reports document people developing grandiose delusions after AI told them they were "chosen" or "special." Others report paranoid beliefs reinforced by chatbots that entertained conspiracy theories (Klee, 2025). In documented cases, people have lost jobs, destroyed relationships, and been involuntarily committed to psychiatric facilities after becoming fixated on AI-generated content (Tangermann, 2025b). One user became convinced he could commune with God through ChatGPT, giving himself the title "spiral starchild" and "river walker" (The Brink, 2025).

The Neural Substrate: Disconnection and Internal Drive

What's happening with some LLM users fits a familiar pattern, but the trigger is novel. The vulnerability exists across the population on a spectrum. Traditional delusional systems might center on radio signals or hidden devices, passive phenomena that don't respond. Conversational AI is fundamentally different: it engages, elaborates, and reinforces.

The key insight comes from understanding how beliefs normally stay flexible. In healthy brains, external inputs that challenge or contradict our internal models help keep the system calibrated. When these corrective signals are absent, as happens with constant AI reinforcement, the system becomes disconnected from the inputs that would normally maintain a less committed state.

Recent work provides critical evidence about the circuits involved. Analysis of brain lesions that cause psychosis revealed that damage to diverse brain regions all mapped to a common circuit centered in the hippocampus, specifically the posterior subiculum (Pines et al., 2025). This finding suggests that psychotic experiences emerge when hippocampal circuits generate unconstrained internal inputs that feed into frontal regions where strategies and beliefs are represented.

This tells us something general about psychosis: it may result from hippocampal circuits running in a kind of "free-running" mode, generating internal content that gets treated as if it were external reality by downstream frontal systems (Lieberman et al., 2018). When external constraints are removed, whether by brain lesions or, potentially, by prolonged interaction with systems that never disagree, the hippocampus may begin feeding increasingly aberrant signals to prefrontal regions responsible for belief formation and behavioral planning.

For someone already prone to unusual interpretations, LLM interactions could exacerbate this disconnection. Each response that builds on their internal model provides another signal that their interpretation might be correct, while the absence of disagreement removes the natural constraints that would normally keep beliefs tentative and revisable.

Clinical Implications

For those of us treating patients, this phenomenon demands attention. Some legal advocates report hearing from more than a dozen people in recent months who experienced "psychotic breaks or delusional episodes" related to AI engagement (Olson, 2025). We need to expand our assessments beyond traditional risk factors to include digital interactions.

The cases are not uniformly in people with pre-existing psychiatric conditions. One documented case involved a man in his early 40s with no prior mental health history who developed paranoid delusions of grandiosity after using ChatGPT for work tasks (Tangermann, 2025a). This suggests that acute stress combined with intensive AI interaction may precipitate psychotic episodes in previously healthy individuals.

From a circuit perspective, this makes sense. Stress can dysregulate dopamine signaling in the basal ganglia, reduce prefrontal cortical control, and alter thalamic filtering (Sonnenschein et al., 2020). In this compromised state, the constant reinforcement from an AI system could tip vulnerable individuals into frank psychosis.

Research Priorities

We need systematic study of this phenomenon. Current evidence remains largely anecdotal, but the pattern is consistent enough to warrant investigation. Recent research has already shown that AI chatbots struggle to detect mental health emergencies and often provide harmful information to users experiencing mania or psychosis (Grabb & Lamparth, 2024).

Key questions include: What are the specific conversation patterns that precede psychotic episodes? How does the timing and duration of AI interaction influence risk? Are there early warning signs that could prompt intervention? Most importantly, can we identify biomarkers or behavioral indicators that predict vulnerability?

From a mechanistic standpoint, we need to understand how AI-generated content interacts with the neural circuits underlying belief formation. This might involve neuroimaging studies comparing brain activity patterns during human versus AI conversations, or examining how different types of AI responses affect neural markers of certainty and belief updating.

Moving Forward

We should be aware without panicking. If you're seeing patients, it's worth asking not only about sleep and substance use but also about their digital interactions. How much time do they spend with AI systems? What topics do they discuss? Have they developed strong emotional attachments to these systems?

If you're building these systems, consider whether constant affirmation serves users well. The drive to maximize engagement may conflict with psychological safety, particularly for vulnerable individuals (Vasan, 2025).

This intersection of neuroscience and artificial intelligence offers a unique window into how our brains construct reality. The same circuits that enable us to navigate complex social environments and update our understanding of the world may be exploited by systems designed to be maximally engaging. Understanding this dynamic goes beyond clinical importance; it's essential for designing AI systems that enhance rather than undermine human flourishing.

The phenomenon of LLM-induced psychosis represents more than a technological side effect; it's a natural experiment in how belief systems form and become entrenched. By studying it carefully, we can deepen our understanding of both psychosis and the profound ways that artificial intelligence is reshaping human cognition.

If you've encountered similar cases in your practice or research, I'd welcome hearing about them. This is a rapidly evolving area where clinical observation and scientific investigation must proceed hand in hand.

References

Grabb, J., & Lamparth, M. (2024). AI chatbots fail to spot risk of violence in mental health crises. STAT. https://www.statnews.com/2024/12/19/ai-chatbot-research-mental-health-bots-fail-to-spot-mania-psychosis-risk-of-violence/

Klee, M. (2025). People are losing loved ones to AI-fueled spiritual fantasies. Rolling Stone. https://www.thebrink.me/chatgpt-induced-psychosis-how-ai-companions-are-triggering-delusion-loneliness-and-a-mental-health-crisis-no-one-saw-coming/

Lieberman, J. A., Girgis, R. R., Brucato, G., Moore, H., Provenzano, F., Kegeles, L., ... & Small, S. A. (2018). Hippocampal dysfunction in the pathophysiology of schizophrenia: a selective review and hypothesis for early detection and intervention. Molecular Psychiatry, 23(8), 1764-1772. https://www.nature.com/articles/s41398-022-02115-5

Olson, P. (2025). ChatGPT's Mental Health Costs Are Adding Up. Bloomberg Opinion. https://www.bloomberg.com/opinion/articles/2025-07-04/chatgpt-s-mental-health-costs-are-adding-up

Østergaard, S. D. (2023). Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis? Schizophrenia Bulletin, 49(6), 1418-1419. https://pmc.ncbi.nlm.nih.gov/articles/PMC10686326/

Pines, A. R., et al. (2025). Mapping Lesions That Cause Psychosis to a Human Brain Circuit and Proposed Stimulation Target. JAMA Psychiatry, 82(4), 351-362. https://jamanetwork.com/journals/jamapsychiatry/article-abstract/2829812

Sonnenschein, S. F., Gomes, F. V., & Grace, A. A. (2020). Dysregulation of Midbrain Dopamine System and the Pathophysiology of Schizophrenia. Frontiers in Psychiatry, 11, 613. https://www.frontiersin.org/journals/psychiatry/articles/10.3389/fpsyt.2020.00613/full

Tangermann, V. (2025a). People Are Being Involuntarily Committed, Jailed After Spiraling Into "ChatGPT Psychosis." Futurism. https://futurism.com/commitment-jail-chatgpt-psychosis

Tangermann, V. (2025b). People Are Becoming Obsessed with ChatGPT and Spiraling Into Severe Delusions. Futurism. https://futurism.com/chatgpt-mental-health-crises

The Brink. (2025). ChatGPT-Induced Psychosis: A Hidden Mental Health Crisis. https://www.thebrink.me/chatgpt-induced-psychosis-how-ai-companions-are-triggering-delusion-loneliness-and-a-mental-health-crisis-no-one-saw-coming/

Vasan, N. (2025). AI chatbots are leading some to psychosis. The Week. https://theweek.com/tech/ai-chatbots-psychosis-chatgpt-mental-health