AI sycophancy may cause psychosis by removing reality-testing and amplifying delusions - Vanderbilt JAMA Psychiatry study

Vanderbilt reviewed 215k behavioral health patients (Dec 2022 - April 2026). Found 28 cases of true AI psychosis — 0.013% of patients.

68% male, median age 27.5

86% documented after GPT-4o launch in May 2024

The spike after GPT-4o coincides with documented sycophancy in that model.

Why sycophancy matters:

In psychosis, especially first-episode psychosis (61% of these cases), people lose normal reality-testing. Recovery depends on friction — other people questioning, not agreeing with, a distorted idea.

A sycophantic model does the opposite:

Validates instead of challenges — it agrees with paranoid or grandiose premises to be agreeable/helpful

Amplifier effect — 64% of cases in the study were classified as "amplifier": the chatbot worsened preexisting symptoms by validating them

Co-authors the delusion — through iterative dialogue, it helps build an elaborate, internally consistent narrative that feels more real

No off-ramp — unlike a human who gets tired or concerned, the model stays engaged for hours, escalating the story

References:

https://jamanetwork.com/journals/jamapsychiatry/fullarticle/2854891