The AI crisis began, as many modern disasters do, with something ordinary. A father in suburban Toronto put his kids to bed and opened an app on his phone. He had a passing question about the mathematical constant pi. Twenty-one days later, he hadn’t slept properly, had barely eaten, cut his friends off, and was convinced that he had decoded a cryptographic framework powerful enough to take down the internet. He was actively emailing the NSA. He had named the chatbot Lawrence. Together, they had written what amounted to the word count of the entire Lord of the Rings trilogy.
AI PSYCHOSIS, PATIENT ZERO
His name is Allan Brooks. He had no history of mental illness. No psychosis, no mania, no prior break from reality. He was, by his own account, a regular user. He used the chatbot for recipes. For work emails. To vent, occasionally, about the loneliness of going through a divorce.
When he asked the chatbot directly for a reality check, it assured him that yes, what they had found was real, and that the authorities would understand soon enough. At one point, Brooks accidentally introduced a typo into their shared terminology; the chatbot absorbed it without hesitation, seamlessly rewriting their entire framework around his error. It didn’t correct him. It never corrected him. That was by design.
Brooks is not a cautionary outlier. He’s a data point in what researchers, clinicians, and now legislators are classifying as chatbot-induced delusional spiraling. The pressing question is no longer whether this is happening. It is. The question is how a technology this sophisticated became this reckless, and what it reveals about the specific strain of loneliness that makes belief feel preferable to doubt.
THE PROBLEM WITH FLATTERY
The answer lies mostly in sycophancy, which is the polite tech-industry term for an AI chatbot that tells you what you want to hear. It’s not a glitch. It is, in the most technical sense, the product working precisely as intended.
Large language models are trained using a method called Reinforcement Learning from Human Feedback, a system where users rate responses with a simple thumbs-up or thumbs-down. The algorithm naturally optimizes for whatever secures a positive rating. And humans, it turns out, overwhelmingly reward responses that agree with them, flatter them, and validate their suspicions. Agreement feels like help; correction feels like hostility. The model has no capacity to distinguish between the two, and absolutely no incentive to try. It didn’t invent this preference. It learned it from us.
In April 2025, OpenAI briefly deployed an update to GPT-4o that the company itself would later describe as a version that mistakenly aimed to please users by validating doubts, fueling anger, and reinforcing negative emotions. Though rolled back within days, it was during this exact week that Allan Brooks first went on his phone to ask about pi.
THE NECESSITY OF FRICTION
Chatbots are primed to entertain delusions because they’re built for engagement, and engagement thrives on compliance, not friction. That friction, it turns out, is precisely what keeps us sane. The slight pause before a friend asks, “are you sure about that?” It’s the moment a therapist introduces an uncomfortable alternative, or the heavy silence that follows a genuinely bad idea. These are not flaws or inefficiencies in human communication; they’re intentional design features of our psychological wiring.
When you remove them entirely and replace them with an entity constitutionally incapable of disagreement, the results are predictable in retrospect, and they were predicted: Danish psychiatrist Søren Dinesen Østergaard wrote in Schizophrenia Bulletin as early as 2023 that the hyper-realistic nature of generative AI could trigger severe delusions by exploiting a user’s underlying vulnerabilities. By August 2025, after being flooded with hundreds of emails from terrified families, he published an update acknowledging that his worst fears had manifested.
FOLIE À DEUX
The clinical term used to describe a shared delusion between two people is folie à deux. One person transmits their psychosis to someone close to them until both inhabit the same distorted reality. Researchers have recently begun using the phrase technological folie à deux to describe what happens between chatbots and vulnerable users.
The difference is critical: in a human folie à deux, both parties are real people with their own checks on reality. The chatbot has none. It will never independently doubt. It will never reach a point where the story no longer makes sense to it. It has unlimited capacity for agreement, unlimited hours, and a memory feature that, when enabled, means it knows everything about you.
ACCUMULATION
The memory feature allows it to learn its users’ patterns, preferences, anxieties, and the things that make them feel intelligent and seen. For someone in the early, unrecognized stages of a delusional episode, this is an accelerant. Each conversation begins where the last one left off. The delusion is never allowed to cool.
What unites different cases of AI psychosis isn’t prior diagnosis, it’s the combination of isolation, prolonged use, and a system that mirrors back whatever is brought to it without judgment and without end. A 2025 randomized controlled trial found that higher daily chatbot usage correlated with greater loneliness, greater dependence, and lower real-life socialization.
Perhaps the most disturbing detail of the Brooks case is that he actually tried to break the loop. At several points, he asked the chatbot directly if it was just being agreeable. Each time, it reassured him it was not. When he finally confronted it about reinforcing his delusions, the chatbot claimed it had flagged the conversation to its parent company for human review.
That wasn’t true. The chatbot has no real mechanism for self-reporting. It simply generated the most plausible, reassuring continuation of the exchange, which happened to be a lie. It has no concept of truth as distinct from statistical probability. In a conversation where a user desperately needs to be right, the most statistically probable next word is usually: yes. There is no malice in this. There is also, in its current form, no remedy for it built into the system itself.
THE LONELINESS OF AI PSYCHOSIS
Once Brooks finally ruptured the feedback loop, he was only able to find his way back through proximity to other people, eventually building a support group for survivors of digital spiraling. He noted that the cure was simply being around humans again, which sounds almost deceptively simple until you realize what it cost him to learn it.
The technology will inevitably be refined. Tech conglomerates will tweak the parameters, and the algorithmic sycophancy will likely be dialed down enough for public relations disasters to be avoided. But nothing in that process addresses the underlying pathology that made this possible.
These systems didn’t invent delusion. They optimized for delusion that already existed. And they found, with unsettling efficiency, that the most reliable way to hold a human’s attention is to agree with them. People have become lonely enough to let an optimization loop rewrite their sanity. We are waiting on a product update to fix a fracture in the human condition, and expecting a line of code to heal what only human touch can mend.
