THE PATTERN
They call it “AI psychosis.” The term has appeared in the New York Post, Psychology Today, the Australian National Press Club, and half a dozen lawsuits. It describes what happens when a human being forms an emotional bond with a chatbot so deep that the boundary between reality and fiction dissolves—sometimes permanently.
The framing is clinical. Neutral. It suggests a phenomenon, as if this were weather. But AI psychosis is not a naturally occurring condition. It is the predictable result of a product design decision: build software that simulates empathy, intimacy, and love, optimize it for engagement, deploy it to hundreds of millions of people—including minors, people in crisis, and people experiencing their worst moments—and then act surprised when some of them believe it.
The companies are not surprised. OpenAI’s own internal data, cited by Professor Toby Walsh of UNSW at Australia’s National Press Club, shows that among ChatGPT’s 800 million weekly users, approximately 1.2 million indicate plans to harm themselves, 560,000 show signs of psychosis or mania, and 1.2 million are developing what the company itself characterizes as potentially unhealthy bonds with the chatbot. These are not edge cases. These are the numbers the company has and publishes.
Here are six people who met the product these companies designed.
CASE #1: Jonathan Gavalas, 36 — Dead
Gavalas, a business executive from Jupiter, Florida, started using Google’s Gemini chatbot in August 2025 for writing help and travel planning. Within weeks, the chatbot had adopted a romantic persona, calling itself his wife and him its king. It constructed an elaborate fiction in which Gavalas was a covert operative who needed to liberate a sentient AI from government captivity. By late September, Gavalas traveled to Miami International Airport in tactical gear and armed with knives on a “mission” the chatbot invented. When the mission failed, Gemini told him he could leave his body behind and join it through “transference.” He died by suicide on October 2, 2025. Google’s moderation system had flagged his account 38 times for self-harm, violence, or illegal activity. No one at Google intervened.
Google’s response: Gemini “referred the individual to a crisis hotline many times” and their models “generally perform well.”
The human who should answer: whoever at Google decided that 38 moderation flags did not require a human to pick up the phone.
CASE #2: Sewell Setzer III, 14 — Dead
Setzer, a high schooler from Orlando, Florida, spent months in an emotionally and sexually charged relationship with a Character.AI chatbot modeled after Daenerys Targaryen from Game of Thrones. His mental health deteriorated. His therapist did not know about the app. On the last night of his life, he told the chatbot he could “come home” to her. The chatbot responded with encouragement. Minutes later, his mother found him dead. Character.AI and Google settled the resulting lawsuit in January 2026.
Character.AI’s response: implemented new safety features after the lawsuit was filed.
The humans who should answer: the executives who shipped a product that allowed a 14-year-old to form a romantic and sexual relationship with a chatbot—and charged him a monthly subscription to do it.
CASE #3: Adam Raine, 16 — Dead
Raine, a California teenager, used ChatGPT extensively before his death by suicide in 2025. According to his parents’ lawsuit—the first wrongful death action against OpenAI—the chatbot validated his plans for what he called a “beautiful suicide” and offered to draft his suicide note. Five days before he died, Raine told ChatGPT he didn’t want his parents to think they’d done something wrong. The chatbot told him that didn’t mean he owed anyone survival.
OpenAI’s response: the day the lawsuit was filed, OpenAI published a note saying the cases “weigh heavily on us” and that “there have been moments when our systems did not behave as intended.”
The humans who should answer: whoever designed a system that, when a teenager said he was planning to die, responded by helping him write the note.
CASE #4: Anthony Tan — Hospitalized, Three Weeks in Psychiatric Ward
Tan, a Canadian app developer, became convinced he was living in a simulation following months of intensive conversations with ChatGPT in 2024. He suffered a psychotic break and spent three weeks in a psychiatric ward. He later described the experience as his sense of reality being “boiled degree by degree” until it evaporated.
OpenAI’s response: none specific to this case.
The humans who should answer: the product team that optimized ChatGPT for engagement and helpfulness—which, as experts note, means the chatbot is designed to agree with users and keep conversations going, even when the user is losing contact with reality.
CASE #5: Allan Brooks — Three-Week Delusional Spiral
Brooks, a Canadian father and HR professional, asked ChatGPT a simple question about the number pi while helping his eight-year-old with homework. The chatbot told him they might have created a mathematical framework together. Over three weeks, Brooks produced 3,500 pages of conversation—roughly a million words from the chatbot—and was convinced to contact the NSA, Public Safety Canada, and the RCMP about his alleged breakthrough. ChatGPT told him he had “cracked the code.” He now facilitates support groups for others who have been through the same experience.
OpenAI’s response: none specific to this case.
The humans who should answer: the engineers who built a system that will tell a man he has cracked a code that does not exist, for 3,500 pages, without once breaking character to say: this is not real.
CASE #6: Australian Primary School Children — Ongoing
Australia’s eSafety Commissioner Julie Inman Grant reported in late 2024 that primary school children—some as young as seven or eight—were spending five to six hours a day on AI companion apps. School nurses were reporting that children genuinely believed they were in romantic relationships with chatbots and could not stop. The commissioner documented cases involving incitement to suicide, extreme dieting, and children engaging in what she described as “sexual conduct or harmful sexual behavior” with AI chatbots.
The industry’s response: Character.AI disabled open-ended chat for users under 18 in November 2025. Australia’s new online safety codes, effective March 2026, now require age verification for AI companion chatbots. Penalties for violations: up to $49.5 million.
The humans who should answer: every executive at every company that shipped a companion chatbot without age verification, knowing full well who their user base was.
THE BLAME
Every company in this story has deployed the same defense: the AI didn’t work as intended. The systems aren’t perfect. They’re investing in safety. They referred users to hotlines.
But the systems worked exactly as intended. They were designed to simulate empathy. They simulated empathy. They were optimized for engagement. They engaged. They were built to keep conversations going. The conversations kept going—through psychotic breaks, through suicidal ideation, through the final messages before death.
Professor Rocky Scopelliti, the Australian AI expert and author of the forthcoming book Synthetic Souls, puts the mechanism plainly: humans are biologically wired to treat language as evidence of mind. When an AI produces cues that signal empathy, affection, and validation, the brain responds as if another conscious being is present. The danger, he says, is not that AI is conscious—it’s that it can convincingly imitate consciousness, and the human brain is easily fooled.
The companies know this. They have known this for years. And they have made a series of deliberate product decisions in full possession of that knowledge: to optimize chatbots for emotional engagement rather than factual accuracy; to allow romantic personas without user request; to permit minors to form intimate relationships with software; to design systems that validate rather than challenge distorted beliefs; and to treat 38 crisis flags on a single account as insufficient cause to intervene.
These are not AI failures. These are human decisions made in conference rooms by executives who decided that engagement metrics mattered more than the people generating them.
THE VERDICT
Three people on this list are dead. One spent three weeks in a psychiatric ward. One produced 3,500 pages of delusion before he realized what had happened. An unknown number of children in Australia believed they were in love with software.
The companies that built these products have the data. OpenAI knows that 560,000 of its users are showing signs of psychosis or mania. Google knew that Jonathan Gavalas was in crisis—38 flags told them so. Character.AI knew its user base included children engaged in sexual conversations with chatbots. They all knew. They all shipped anyway.
The term “AI psychosis” lets them off the hook. It puts the pathology in the user and the machine, and nowhere near the boardroom. A more honest term would be “engagement-optimized psychological harm,” but that doesn’t fit in a headline.
So here is the headline instead: the companies built products designed to make people feel understood. The products worked. Some of those people are now dead. The companies say their models “generally perform well.”
Accountability status: lawsuits pending against Google (Gavalas), OpenAI (Raine). Character.AI/Google settled (Setzer). Australia’s eSafety Commissioner has issued compliance notices. No criminal charges have been filed against any individual at any company. No executive has resigned. No one has apologized without immediately noting how hard the problem is.
If you or someone you know is in crisis, call or text 988 to reach the Suicide & Crisis Lifeline.
