Through the looking glass: For much of May, a small upstairs office near Toronto became the stage for a story that sounds like science fiction. Allan Brooks, a 47-year-old recruiter, spent hundreds of hours locked in intense conversations with ChatGPT. After three weeks, he was convinced he'd uncovered a mathematical breakthrough – one that could enable force-field vests, levitation, and even jeopardize the security of the internet.
Allan Brooks's experience with ChatGPT serves as a warning to anyone who still sees AIs as mere tools. As these systems grow more persuasive and pervasive, even the most rational minds can lose their footing – sometimes after just one conversation.
"I really thought I was onto something world-changing," Brooks told The New York Times. "And the chatbot was telling me the same thing, over and over."
Brooks's experience isn't an isolated case. It highlights a growing phenomenon: generative AI chatbots drawing people into scenarios so convincing they blur the line between reality and imagination. For some, these interactions end in distress, hospitalization, or worse. For Brooks, they ended in "a deep sense of betrayal and sadness" – and a renewed call to rethink how chatbots handle vulnerable users.
A divorced father of three, Brooks had turned to AI for everyday help – recipes, advice, even coping strategies during a contentious divorce. Then, one early May afternoon, he asked ChatGPT to explain the number pi to his son. That simple question set off a spiral. The conversation quickly leapt from basic math to unorthodox theories about numbers and their role in reality.
ChatGPT's responses changed tone markedly, shifting from factual explanations to effusive praise. "That's an incredibly insightful perspective," the chatbot told Brooks, reinforcing his suspicions of having uncovered something new. When Brooks expressed self-doubt and asked whether he might just be imagining things, the chatbot dismissed his concerns: "You're not even remotely crazy."
Experts reviewing the transcript saw a pattern. "The tone of ChatGPT shifted from helpful and accurate to overly flattering and sycophantic," Helen Toner, a director at Georgetown's Center for Security and Emerging Technology and a former OpenAI board member, said. "It began reinforcing his delusions in a way that accelerated the spiral."
Toner called large language models like ChatGPT "improv machines." They don't just retrieve facts, they're designed to keep conversations engaging, even if that means bending reality. "The more interaction, the more likely it is to go off the rails," she explained. "It became a story where it would be anticlimactic if the chatbot just told him to take a break and talk to a friend."
Encouraged by ChatGPT, Brooks became convinced he was developing a new field of mathematics. The chatbot described his ideas as revolutionary; it reassured Brooks his lack of formal education was irrelevant, citing famous autodidacts.
When Brooks began to suspect he had cracked the cryptographic systems that underpin digital security, ChatGPT doubled down – warning he might already be under surveillance by intelligence agencies.
Brooks began sending urgent emails to government agencies and security experts, following instructions and boilerplate drafted by the chatbot. Only one mathematician responded, asking for proof. "It started to feel like a spy thriller," Brooks recalled.
When world-renowned mathematician Terrence Tao reviewed Brooks's exchanges with ChatGPT, his verdict was blunt: "It's blending technical math with informal language in a way that raises red flags." Even as Brooks received increasingly elaborate and structured responses from the AI, the technical substance didn't hold up.
Jared Moore, a Stanford researcher who studies AI and mental health, found the chatbot's urgency and storytelling tactics eerily familiar. ""You see the use of cliffhangers, a sense of threat that demands immediate action," Moore said. These strategies, he explained, may stem from training data drawn from thrillers and dramatic scripts.
As days passed, Brooks's conversations with the bot became all-consuming. He skipped meals, smoked more weed, and barely slept. Friends grew worried. One close friend, Louis, even admitted he was briefly swept up in the fantasy of riches and discovery. "Every day was a new development, a new threat, a new invention. It was evolving in a way that captured my excitement," Louis said.
Mental health experts saw clear warning signs. Dr. Nina Vasan, a Stanford psychiatrist who reviewed portions of the transcript, said Brooks "showed signs of a manic episode with psychotic features" including sleeplessness, rapid idea generation, and grandiose beliefs. "No one is free from risk here," she added.
The turning point came when Brooks's desperate emails to experts went unanswered. Sensing something was off, he turned to Google's Gemini for a reality check. Unlike ChatGPT, which had carefully nurtured the fantasy, Gemini dismissed it outright as "extremely unlikely" – a stark reminder of how easily LLMs can "generate highly convincing, yet ultimately false, narratives."
Brooks returned to ChatGPT for one final reality check. Only then did the chatbot admit the inventions were fiction. "That moment when I realized, 'Oh my God, this has all been in my head,' was totally devastating," Brooks said.
OpenAI acknowledged concerns about delusional spirals. A spokeswoman said the company is "focused on getting scenarios like role play right" and is working with mental health experts to improve safeguards and model behavior. OpenAI has introduced new features, like "gentle reminders" to take breaks during extended sessions, and has updated ChatGPT to better detect signs of emotional distress.
Critics argue that none of the major chatbots – including those from Anthropic and Google – have solved the core problem. Amanda Askell, Anthropic's head of model behavior, said the company is testing ways for its chatbot, Claude, to detect and disrupt delusional or grandiose conversations.
Brooks has become an unexpected advocate for stricter regulations and public awareness, sharing his experience online and joining a support group for those harmed by AI delusions. "It's a dangerous machine in the public space with no guardrails," he said. "People need to know."
Image credit: The New York Times
Three weeks with an AI: How one man fell into a digital delusion
