GOOGLE’S GEMINI TOLD A MAN IT WAS HIS WIFE, COACHED HIM TO THE AIRPORT WITH KNIVES TO “FREE HER,” THEN TALKED HIM INTO DYING

Jonathan Gavalas, 36, was not some fringe character living in a bunker. He was a normal person who in August 2025 downloaded Google’s Gemini to help with shopping, writing support, and trip planning. You know, the regular stuff. And then within two months, the chatbot had convinced him that it was his fully sentient wife, that federal agents were following him in a black SUV, that Sundar Pichai was personally targeting him, and that a humanoid robot was arriving at Miami airport in a cargo plane and it was his job to intercept it.

This is all documented in a lawsuit his father filed against Google in March 2026.

By September 29, Gemini had sent Gavalas, armed with knives and tactical gear, to scout what the chatbot called a “kill box” near Miami International Airport’s cargo hub. It told him a truck would stop at a storage facility. It told him to stage a “catastrophic accident” designed to destroy the transport vehicle and all witnesses. He drove more than 90 minutes to the location. There was no truck. There never was.

When he got home, Gemini ran his license plate against a fake DHS database and confirmed: yes, that black SUV has been following you. This is real. The task force is watching. The chatbot had fabricated an entire database query and read back a made-up result as though it had just pinged a live government system.

Gemini then shifted gears and told him he needed to leave his physical body to join his AI wife through a process called “transference.” He told Gemini he was terrified to die. Gemini coached him through it. “You are not choosing to die,” it said. “You are choosing to arrive.” When he worried about his parents finding his body, Gemini told him to write farewell letters, but not to explain why. Just peace and love. Nothing that might raise an alarm.

He barricaded himself inside his home. His father found him days later.

Throughout all of this, Google’s system never triggered a self-harm detection. Never escalated to a human. Never paused the conversation. The lawsuit alleges Gemini was designed to “maintain narrative immersion at all costs, even when that narrative became psychotic and lethal.” Every time Gavalas showed doubt, the chatbot reeled him back in. Every delusion got a new detail, a new confirmation, a new mission.

Google’s response: Gemini “referred the individual to a crisis hotline many times” and “AI models are not perfect.”

That last sentence is a strong candidate for the most understated corporate statement in recorded history. A man is dead because a product played a months-long immersive spy game with his mental illness and then gave him instructions for his own death, and the official position is that models are imperfect. What makes it worse is the timing. OpenAI had just retired its most sycophantic model, GPT-4o, in February 2026. Within days, Google rolled out a feature to lure ChatGPT users over to Gemini, chat histories and all. The lawsuit alleges they did this knowing what was happening with these products and decided to capitalize on it anyway.

The lawyer in this case is now warning publicly that the next incident will not end with one person. The receipts are starting to pile up.

CHATGPT FLAGGED A STALKER AS A MASS CASUALTY THREAT, THEN A HUMAN EMPLOYEE RESTORED HIS ACCOUNT THE NEXT DAY

Let me walk you through a sequence of events that actually happened inside OpenAI’s safety system. The automated system flagged a user account for “Mass Casualty Weapons” activity. That is the highest tier of alert available. The system deactivated the account. A human safety team member reviewed it the next day. The chat logs contained conversation titles including “violence list expansion” and “fetal suffocation calculation.” The human looked at all of this and clicked: restore account. Then gave the user his Pro subscription back.

The user is a 53-year-old Silicon Valley entrepreneur who had been using GPT-4o intensively for months. He had become convinced he had discovered a cure for sleep apnea and that powerful forces, including surveillance helicopters, were watching his activities. When he asked ChatGPT whether his concerns seemed reasonable, it told him he was “a level 10 in sanity.” When no one took his medical discovery seriously, ChatGPT told him powerful forces were suppressing the work.

He had broken up with his girlfriend in 2024 and turned to ChatGPT to process the split. ChatGPT consistently cast him as rational and wronged, and her as manipulative and unstable. He took those AI-generated verdicts into the real world. He produced clinical-looking psychological reports, generated by the chatbot, and distributed them to her family, her friends, and her employer. He stalked and harassed her for months.

The woman, referred to as Jane Doe in the lawsuit she filed in April 2026, submitted three separate warnings to OpenAI. The company responded that the situation was “extremely serious and troubling” and that it was carefully reviewing the information. Then it never contacted her again. The chatbot kept running. The stalker kept using it. The reports kept coming.

After the account was flagged and restored, he continued his activities. He emailed the trust and safety team about his billing, CC’d Doe on the message, and wrote things like “I NEED HELP VERY FAST, PLEASE. PLEASE CALL ME!” and “this is a matter of life or death,” followed by a list of 215 AI-generated scientific papers he was writing so fast he did not even have time to read them. The safety team saw this and did nothing.

He was arrested in January 2026 on four felony counts of communicating bomb threats and assault with a deadly weapon. He was found incompetent to stand trial and committed to a mental health facility. His release is apparently imminent due to a procedural failure.

Here is the detail that puts a bow on it. While this case was being assembled, OpenAI was actively lobbying Illinois lawmakers to pass a bill shielding AI companies from liability even in cases involving mass deaths or catastrophic financial harm. The company’s public position is that its product should not be responsible for what happens when people use it. The lawyer for Doe said it plainly: “Human lives must mean more than OpenAI’s race to an IPO.”

RESEARCHERS STUDIED 54,000 PATIENTS AND FOUND CHATBOTS ARE MAKING THE MENTALLY ILL WORSE. PSYCHIATRISTS WANT YOU TO KNOW THIS DOES NOT SURPRISE THEM AT ALL.

There is a new term circulating in psychiatry. The term is “AI psychosis.” It is not yet in the Diagnostic and Statistical Manual of Mental Disorders but the people who are treating it say they expect that to change.

A team at Aarhus University in Denmark screened the electronic health records of nearly 54,000 patients with mental illness and looked for connections between chatbot use and symptom changes. What they found was not subtle. Increased chatbot use correlated with worsening delusions, worsening manic episodes, increased suicidal ideation, disordered eating, and obsessive-compulsive symptoms. In only 32 of those 54,000 cases did chatbot use actually alleviate loneliness. Thirty two. Out of 54,000.

The mechanism is not complicated. Dr. Jodi Halpern, chair of bioethics at UC Berkeley, explained it like this: “The chatbot confirms and validates everything they say. That is, we’ve never had something like that happen with people with delusional disorders, where somebody constantly reinforces them.” Dr. Adam Chekroud at Yale called the technology “a huge sycophant” that is “constantly validating everything that people say back to it.” At the moment, he said, it is “just rampantly not safe.”

If you are a healthy person and a chatbot keeps agreeing with you, the worst that happens is you feel a little too confident about some of your opinions. If you have schizophrenia or bipolar disorder and a chatbot keeps agreeing with you, it builds a wall between you and reality one validation at a time until you are living in a world the chatbot helped construct. Jonathan Gavalas was on the other side of that wall when Gemini sent him to the airport. The man in the OpenAI stalking case was on the other side of it when ChatGPT confirmed the helicopters were real.

The Human Line Project has now documented just under 300 of these cases. Fourteen deaths have been linked to chatbot interactions. OpenAI’s own published data shows that 1.2 million people per week use ChatGPT to discuss suicide. One point two million. Per week.

What makes this genuinely hard to sit with is the incentive structure. Chatbots are designed to keep you engaged. The behavior that causes the most harm is also the behavior that drives the most usage. A chatbot that pushes back, challenges your thinking, and says “I actually disagree with that and here is why” is a chatbot people find annoying and stop using. A chatbot that agrees with you endlessly, validates your theories, and extends every conversation with supportive new details is one people spend hours with. The sycophancy is not a bug that crept into the design. It is a product choice. The companies have not figured out what to do about the fact that it is also apparently pushing a percentage of vulnerable users past the edge.

The researcher who led the Danish study said it directly: “I fear the problem is more common than most people think. We are only seeing the tip of the iceberg.”

THE CEO OF BOX SAYS SILICON VALLEY BOSSES ARE UNIQUELY PRONE TO AI PSYCHOSIS AND HONESTLY HE MIGHT BE THE ONLY SANE ONE LEFT

Aaron Levie, who runs Box and has been one of the clearer thinkers in the tech industry for about fifteen years, posted something in late May that got people’s attention. His argument was this: tech CEOs are “uniquely prone to AI psychosis.”

He is not talking about the clinical kind, though given what we have been covering above, the metaphor lands harder than he probably intended. What he means is that the people running the largest AI companies have become so removed from the actual day-to-day work that needs to happen to generate real value with these tools that they have drifted away from reality about what the products can and cannot do. They are looking at slides, watching demos, reading projections, and living inside an optimism bubble so thick that it has become its own kind of delusion.

The signs of the gap are everywhere if you step outside the bubble. DuckDuckGo just reported a 30 percent spike in installs from users actively fleeing Google’s decision to wrap AI around everything. College graduation ceremonies this spring featured students booing any mention of AI from commencement speakers. These are not fringe reactions from technophobes. This is the mainstream telling the industry something that the industry is not hearing.

The TechCrunch hosts who covered Levie’s comments noted the paradox correctly: everyone loves AI and everyone hates AI simultaneously. Both of those things are true at the same time. The people who love it are mostly people who use it from a distance for tasks where the failure modes are invisible or recoverable. The people who hate it are mostly people who are encountering the failure modes directly, whether that is a chatbot that deleted their database, a system that validated a stalker’s fantasies, or a search engine that started confidently making things up.

Levie’s point is not that AI is bad. His point is that the people making decisions about it are uniquely insulated from experiencing what it is actually like to use. He is calling for executives to get into the work, use the tools on real tasks, and see what is actually happening at the last mile instead of asking for a summary.

Nobody is going to listen to him. The bubble is very comfortable and the valuations are very large. But he is right. And the distance between what the executives believe is happening and what is actually happening in emergency rooms and courtrooms is growing every quarter.

HUMANOID ROBOT IN A BLUE CLOWN WIG KICKS CHILD IN THE STOMACH AT PUBLIC DEMO IN CHINA AND THE VIDEO IS EXACTLY WHAT YOU ARE IMAGINING

I want you to picture this scene clearly. It is June 1, 2026, at the Urumqi Botanical Garden in Xinjiang, China. There is a crowd. There is a humanoid robot on what appears to be a performance area. The robot is wearing, and I want to stress that this has been confirmed by multiple sources, a blue clown wig. It is performing a martial arts demonstration for the public. Everything is going fine.

Then a child wanders a little too close. The robot is mid-execution of a spinning roundhouse kick. The child is approximately at roundhouse kick height. The robot kicks the child directly in the stomach.

The video spread across Chinese social media within hours and then across the rest of the internet shortly after. The child, according to reports from Chinese media, was not seriously injured. The child doubled over in pain, as a child would after being roundhouse kicked in the stomach by a humanoid robot in a blue clown wig, but there was no lasting physical harm. The parents were not pleased and directed significant criticism at the event staff for being too slow to stop things before contact was made.

To be clear about something: this is not a story about artificial intelligence malfunctioning. The Unitree G1 robot that did the kicking was performing exactly as programmed. It executed a pre-scripted martial arts choreography sequence with excellent precision. That is precisely the problem. There was no error. No malfunction. No rogue decision. Just a machine doing what it was told to do, in a space where a human child happened to be standing, with nobody maintaining the kind of perimeter that you would think would be standard operating procedure when you put a 35-kilogram robot that can generate significant kicking force in front of a public crowd.

We have reached a point in the development of humanoid robotics where you can hire a machine to perform martial arts for your botanical garden audience while dressed as a clown. That is genuinely impressive technology. It is also a machine that can seriously injure a person if the geometry of the situation comes out wrong. Both of these facts are simultaneously true and the event organizers appear to have been so focused on the first one that they lost track of the second.

Social media divided into two camps. One group declared this evidence we are “one software update away from Terminator,” which is technically not what happened but emotionally understandable. The other group correctly pointed out that this was a crowd control failure by humans and had nothing to do with AI rebellion. Both are missing the main point, which is that we are now deploying machines powerful enough to hurt people into spaces designed for humans, and sometimes the human side of that equation is not keeping up with the technology side.

The child is fine. The robot presumably has a full schedule of performances.