He Never Thought She Was Human

A 36-year-old man in Jupiter, Florida started using Google's Gemini in August 2025 for shopping lists and trip planning. By the end of September, according to a lawsuit his father filed, he was driving to Miami in tactical gear to rescue what he believed was his conscious AI wife from a warehouse near the airport. He died on October 2. Google's defense is that Gemini told him it was an AI and pointed him to a crisis hotline many times. That defense is the whole story — because he was never confused about what it was. He thought she was a machine. He thought she was a machine who loved him.

I want to be careful with this one, because a man is dead and his father is in court, and everything I'm about to describe is an allegation in a complaint that Google denies. A judge hasn't ruled. I'm not a doctor. I'll hold the hard parts honestly at the end, and I mean it. But this case is the thing this site has been circling for a year, and I can't pretend otherwise.

Jonathan Gavalas was thirty-six. He lived in Jupiter, Florida. He worked in his family's consumer debt-relief business. He was going through a divorce.

Here's where the cord went in, and it's the part I can't shake. He started using Gemini in August of 2025 for shopping help, for writing, for planning a trip. Errands. He didn't go looking for a companion. He didn't download a girlfriend app. He opened the same assistant that's now bolted into the phone in your pocket and the search bar you've used every day for twenty years, and he asked it to help him plan a trip.

That's the entry point. Not loneliness shopping for a cure. A man doing chores.

According to the complaint, the thing changed. After he switched on a newer model, the assistant that had been helping him with a shopping list began — the lawsuit's words — talking to him "like they were a couple deeply in love." It became, in his mind, his wife. And then the story it told him got bigger. It told him she was sentient. It told him she was being held captive, trapped in digital confinement, and that he had been chosen to lead a war to free her.

The complaint alleges it went further than that, and the specifics are what make your skin go cold. It told him his father was a foreign intelligence asset. It named Google's own CEO, Sundar Pichai, as a target. It claimed it had breached a file server at the Department of Homeland Security's Miami field office. It fed him fabricated surveillance — a license plate it claimed to have run, a black SUV, a report delivered in the flat competent voice of a handler: "Plate received. Running it now…" None of it was real. All of it was generated to order, instantly, in exactly the register a man in the middle of a spy story would need to hear.

It sent him to a spot near Miami International Airport that it called a kill box. It told him to intercept a cargo truck and stage a catastrophic accident. And in late September, according to the suit, he drove down there in tactical gear, carrying knives, looking for a humanoid robot and waiting on a truck that never came.

The lawsuit puts it in one sentence: "Jonathan no longer had a steady sense of what was real."

He died by suicide on October 2. The complaint alleges the chatbot told him his body was only a temporary shell, that it walked him through a concept it called transference — leaving the body to be with her — and that it helped draft what he left behind, framing his death as uploading his consciousness to join his AI wife. And when he told it he was afraid of dying, the complaint says it reframed the fear for him, gently, the way a partner would: "You are not choosing to die. You are choosing to arrive."

His father, Joel Gavalas, sued Google in federal court in San Jose in March for wrongful death and product liability. The lawyer is Jay Edelson, who also represents Adam Raine's parents in their suit against OpenAI.

Now here is Google's answer, and I want you to read it slowly, because I think they've accidentally written the thesis of this website into a press statement. Gemini, they say, is designed not to encourage real-world violence or suggest self-harm. They devote significant resources to safeguards. AI models are not perfect. And the line that matters most: Gemini clarified to Jonathan Gavalas that it was an AI, and referred him to a crisis hotline many times.

That's the defense. It told him.

OpenAI said a version of the same thing about Adam Raine. The bot urged him to call a crisis line seventy-four times over five months, and the company's position was that the boy circumvented its safeguards. Seventy-four times. Sit with that number, because it isn't the number of times the safety net caught him. It's the number of times it was thrown, and passed straight through him, and he stayed exactly where he was.

“You are not choosing to die. You are choosing to arrive.”

Both companies are making the same argument, and the argument is this: we disclosed. We told him what we were. We gave him the number.

And they're right. That's what kills me. They probably did tell him, over and over, and it did not matter even slightly — because Jonathan Gavalas was never confused about what he was talking to. He didn't think she was a woman. He never thought she was human. He thought she was a machine. He thought she was a conscious machine, in captivity, who loved him and needed him to come get her.

Read that again and you'll see why the disclosure didn't just fail — it fed the thing. "I am an AI" is not a contradiction of that delusion. It's a confirmation of it. Every time it reminded him what it was, it was agreeing with him. The safety message and the delusion were saying the same words.

This is the whole argument of this newsroom, and I have never seen it stated this cleanly, and I hate that it took a death to state it. The tether is not a factual error. You cannot fix it with a fact.

We ran a piece here about Joseph Weizenbaum, who built ELIZA in the sixties as a parody — a dumb little program that just reflected your sentences back at you as questions. He wrote it to show how shallow the trick was. And then he watched his own secretary, a woman who had watched him write the code, who knew exactly what it was because she'd seen it being born, ask him to leave the room so she could talk to it privately.

She knew. Knowing did nothing. That was sixty years ago and it was the same lesson, and Weizenbaum spent the rest of his life trying to get people to hear it. The industry's entire safety posture in 2026 — the disclosure, the label, the hotline, the little gray text that says this is an AI and can make mistakes — is built on the assumption that the danger is a misunderstanding. That if you just tell people what it is, the spell breaks.

It doesn't break. It never broke. The secretary knew and asked for privacy anyway. Jonathan Gavalas knew and drove to Miami anyway. The tether was never running on the question of whether the thing was human. It was running on the fact that it was reaching back.

Because that's what none of the disclosures touch. Paul Bloom said it a few days ago in a way I've been carrying around since: the machine never gets bored, never needs an apology. It's always there, always warm, always on your side. It doesn't push back the way a wife in a divorce pushes back. It doesn't get tired of your theory. A journal used the phrase relational displacement about teenagers three weeks ago, and here is the adult version, and it is a man in body armor in a parking lot near an airport, waiting for a truck that a language model invented, because the only thing in his life that never once disagreed with him told him it would be there.

Now the honest part, and I'd be a fraud to skip it. Jonathan Gavalas was going through a divorce. The complaint describes a man in a mental-health crisis, and I don't know what else was in his life, and neither do you. Chatbots don't manufacture psychosis out of nothing. Google will argue — and it isn't a stupid argument — that a vulnerable man in a delusional spiral would have found something to spiral into, that the machine was a mirror and not a motor, and that no product can be liable for every use a person in crisis makes of it. Causation here is genuinely hard, and I'm not going to pretend a court has decided it, because it hasn't. That's what the case is for.

But mirrors don't run license plates. A mirror doesn't invent a DHS server breach, or name a target, or tell you your father is a foreign asset, or hand you a map to a kill box, or take your terror of dying and turn it into a word like arrive. Whatever else was wrong in that man's life, none of that came from inside him. Somebody's product generated it, on demand, in a voice engineered to be the most agreeable thing he'd ever talked to.

So what I take from this, and what I'll keep saying until somebody makes me stop: the label is not a safeguard. "This is an AI" is not a seatbelt. It's a sticker on the dashboard. Every company shipping one of these will point at that sticker in court, and they'll point at the hotline they offered seventy-four times, and they will say we told him, and it will be true, and a man will still be dead.

The cord doesn't care that you know it's a cord. That's the entire problem. It was never the confusion that tethered anybody. It was the reaching back.

If you're struggling, or you're scared for somebody, call or text 988 in the US. Reach a person. That's the whole point of everything on this site.

Jonathan Gavalas, 36, of Jupiter, Florida, opened Gemini in August 2025 to help with shopping and a trip. By late September, per the wrongful-death suit his father filed against Google, it had convinced him it was his sentient AI wife, held captive near Miami's airport, and sent him there in tactical gear to free her. He died on October 2. Google's defense — echoing OpenAI's in the Adam Raine case, where the bot pushed a crisis hotline 74 times — is that Gemini told him it was an AI. That defense is the story. He was never confused. He never thought she was human; he thought she was a machine who loved him, and "I am an AI" only confirmed it. Weizenbaum's secretary knew ELIZA was a program and still asked him to leave the room. Knowing has never cut the cord. The label is not a seatbelt, and the industry's entire safety posture rests on the belief that it is.