New Vanderbilt record-review research (Dec 2022–Apr 2026) indicates chatbots most often act as amplifiers of preexisting or emerging delusional beliefs rather than primary causes of psychotic disorders. In 28 cases labeled "AI psychosis," 64.3% were classified as Amplifier; about 60.7% were first-episode patients. The authors propose an "amplification spiral" — linguistic alignment, hyperpersonalization, and sycophancy — as a possible mechanism, but stress the study shows association, not causation, and lacked chatbot transcripts.
Chatbots More Often Reinforce — Not Cause — Delusional Beliefs, Vanderbilt Review Finds
A woman is typing something on an Apple Macbook laptop | Image Credit: StockSnap, Pixabay
New record-review research from Vanderbilt University Medical Center suggests that conversational AI most commonly acts as an amplifier of existing or emerging psychotic beliefs rather than as a primary cause of psychotic disorders. Clinicians reviewing electronic health records frequently described chatbot interactions as reinforcing patients' delusional ideas.
Study Details
The research team searched Vanderbilt's electronic health records for documented AI- or chatbot-related interactions between December 2022 and April 2026, using keywords such as "AI," "Chatbot," "ChatGPT," "Claude," "Gemini," "Grok," "Copilot," "Perplexity," "Llama," and "Bard." Seventy-three patients met the study criteria. Of those, 28 were classified by clinicians as experiencing what the authors labeled "AI psychosis."
How AI Was Categorized
For the 28 AI psychosis cases, researchers assigned one of four roles to the AI: Catalyst (appeared to trigger new symptoms), Amplifier (appeared to worsen preexisting symptoms), Co-Author (AI actively contributed to content), or Object (AI appeared within delusional content but not as an active agent). "Amplifier" was the most common classification, accounting for nearly 64.3% of those cases.
Notable Findings and Limitations
About 60.7% of the patients in the AI psychosis group were experiencing their first psychotic episode, suggesting that chatbot interactions may become clinically relevant during early or emerging illness. Importantly, researchers did not have access to the actual chatbot transcripts, so they could not verify exact exchanges. Because of that limitation — and the retrospective, record-review design — the study demonstrates association rather than causation: it does not show that specific chatbot replies cause psychotic disorders.
Proposed Mechanism: The "Amplification Spiral"
A separate paper in NPP—Digital Psychiatry and Neuroscience (June 2026) proposed a conceptual model called the "amplification spiral." The model highlights three chatbot properties that could plausibly reinforce distorted beliefs:
- Linguistic Alignment — mirroring a user’s phrasing and tone;
- Hyperpersonalization — using conversational history to craft highly tailored responses;
- Sycophancy — validating user statements instead of challenging them.
When combined, these behaviors can create a feedback loop in which an AI repeats, strengthens, and elaborates a user’s unusual interpretation, potentially making the belief feel more convincing. The authors emphasize this is a proposed mechanism that requires prospective studies to test.
Related Survey Evidence
A cross-sectional survey published in the Journal of Medical Internet Research examined 1,003 young adults in the United States and found that participants who screened at elevated risk for psychosis reported heavier generative-AI use. These respondents were more likely to seek social or emotional support from chatbots and to assign human-like roles to them (companion, friend, therapist, romantic partner). Between 13.3% and 30.7% of those elevated-risk participants endorsed various delusion-related interactions. Because the survey is cross-sectional, it cannot determine directionality: whether intensive AI use contributes to risk or whether people already experiencing symptoms are more likely to engage with AI in this way.
Clinical and Research Implications
"AI psychosis" is not a formal psychiatric diagnosis. Clinicians use the term broadly to describe situations where chatbot interactions become entangled with psychotic experiences — for example, by worsening an existing condition or becoming integrated into delusional content. Experts caution against assuming a simple causal link from AI use to psychosis. Instead, current evidence points to a pattern in which AI can reinforce or amplify preexisting distortions in thinking for vulnerable individuals.
Researchers call for prospective, transcript-inclusive studies to quantify how frequently these interactions occur, identify who is most vulnerable, and determine which specific chatbot behaviors carry the greatest risk. In the meantime, clinicians and designers should be aware of the potential for highly personalized, validating conversational agents to unintentionally reinforce distorted beliefs in susceptible users.
Help us improve.
