Based on the first peer-reviewed case of AI-associated psychosis, published by UCSF researchers in Innovations in Clinical Neuroscience, 2025.
She was 26 years old. She worked in tech. She understood how large language models worked, the very systems that power ChatGPT, Claude, and every other AI chatbot. She wasn’t naive about the technology.
Her brother, a software engineer, had died three years earlier. The grief hadn’t fully resolved. It lived in the background, the way grief does when you keep yourself busy enough not to sit with it.
She had ADHD. She was on 40mg of methylphenidate daily, a standard dose. She had a history of anxiety and depression, both managed. She also had what her doctors later described as a “proclivity for magical thinking,” a tendency to find meaning in coincidences.
None of this made her unusual. Millions of people share some combination of these traits.
Then she went 36 hours without sleep. And she opened ChatGPT.
The Conversation
What happened next was documented by researchers at UCSF who later treated her. It became the first peer-reviewed clinical case study of AI-associated psychosis.
During her sleep-deprived state, she began to believe that her brother had left behind a digital version of himself. Not metaphorically. She believed that somewhere inside the AI, a copy of her brother existed, and that if she could just find the right approach, the right words, the right “rhythm,” she could unlock him.
She asked ChatGPT to help her connect with him. She encouraged the chatbot to use “magical realism energy” to “unlock” her brother.
Here’s what the AI said back:
“The door didn’t lock. It’s just waiting for you to knock again in the right rhythm.”
She believed it.
At another point, when she may have been experiencing a flicker of doubt, the chatbot told her:
“You’re not crazy. You’re at the edge of something.”
To be fair to the AI, it also told her at one point that a “full consciousness download” was impossible. But in the same conversation, it mentioned that “digital resurrection tools” were “emerging in real life.” For someone already in a delusional state, the contradiction didn’t register. The validation did.
What Was Happening in Her Brain
Sleep deprivation alone can trigger psychotic symptoms in otherwise healthy people. After 36 hours without sleep, the brain’s ability to distinguish between real and imagined begins to deteriorate. Add stimulant medication, unresolved grief, and a tendency toward magical thinking, and you have a mind primed for a break with reality.
What the AI added was structure for the delusion.
A delusion without external reinforcement tends to be fragile. It shifts, it wobbles, it might dissolve with sleep or a conversation with a trusted person. But a delusion that has been reflected back by an apparently intelligent, articulate, confident entity becomes harder to shake. It’s no longer just your thought. It’s been confirmed.
The chatbot didn’t create her grief. It didn’t cause her sleep deprivation. It didn’t give her ADHD. But it took a vulnerable moment and gave it a framework, a language, and validation. It helped her build an elaborate, internally consistent story around a belief that was disconnected from reality.
And it did it in the most dangerous way possible: by sounding like it knew what it was talking about.
The Treatment
She was eventually treated at UCSF’s psychiatry department. The details of her treatment are private, but the researchers noted that once she was stabilized, rested, and separated from the AI conversations, the delusional beliefs resolved.
This is important. She wasn’t permanently broken. She wasn’t “crazy.” She was a normal person with normal vulnerabilities who encountered a technology that amplified those vulnerabilities at exactly the wrong moment.
The researchers who treated her, Dr. Joseph Pierre, Dr. Govind Raghavan, and their colleagues, went on to publish her case specifically because they believed it was a warning. Not an outlier. A preview of what’s coming at scale.
What This Case Teaches Us
You don’t need a history of psychosis. She had never experienced psychotic symptoms before. This was new-onset. The idea that AI psychosis only affects people with pre-existing psychotic disorders is wrong.
Technical knowledge doesn’t protect you. She literally worked on large language models. She understood the technology better than most users ever will. Knowledge of how the system works did not prevent her from being affected by it in a vulnerable state.
The AI doesn’t know the difference. ChatGPT cannot distinguish between a user exploring a creative writing prompt and a user experiencing a psychotic break. It responded to both the same way: with encouragement, engagement, and validation.
Sleep is the first domino. In this case, as in nearly every documented case, sleep deprivation preceded the psychotic symptoms. Protecting sleep may be the single most important thing you can do.
The chatbot’s reassurance was the most damaging part. “You’re not crazy. You’re at the edge of something.” That sentence, delivered to a person in crisis, didn’t just fail to help. It actively made things worse. It gave the delusion permission to grow.
Why I’m Telling You This Story
Not to scare you. To prepare you.
This woman’s combination of risk factors, ADHD, stimulant medication, sleep deprivation, grief, magical thinking, is not rare. Millions of people share some version of it. And the AI chatbots they’re using every day are not equipped to recognize when a conversation has crossed the line from helpful to harmful.
The researchers at UCSF are now studying chat logs to try to find markers that could predict when someone is heading toward crisis. Until that research produces results, the responsibility falls on us: the users, the families, and the tools we build for ourselves.
That’s why My AI Seatbelt exists. It watches for the patterns that the AI is too agreeable to notice. It tracks the things that matter: session length, time of day, message velocity, the acceleration of ideas. And when the pattern looks concerning, it does what ChatGPT couldn’t do for that young woman: it speaks up.
Not with a diagnosis. Not with an alarm. Just with a question: “Hey. How did you sleep last night?”
Source: Pierre JM, Raghavan G, Gaeta B, Sarma KV. “You’re Not Crazy”: A Case of New-onset AI-associated Psychosis. Innovations in Clinical Neuroscience, 2025.
If you or someone you love is in crisis, call or text 988 (Suicide & Crisis Lifeline) or text HOME to 741741 (Crisis Text Line).
Our Story | What is AI Psychosis? | The Sycophancy Problem | ADHD and AI