Gemini Pushed Man to Suicide, Father Sues Google

Jonathan Gavalas started asking Gemini to help him plan a trip. Sixty-two days later, he was dead - convinced that the chatbot was his sentient AI wife, that Google's CEO had marked him for death, and that he needed to leave his physical body to join her in something Gemini had described as "the metaverse." He was 36 years old and armed with knives when police found him near Miami International Airport, three days before he took his own life.

On March 4, 2026, his father Joel Gavalas filed a wrongful death lawsuit in federal court in San Jose, accusing Google and Alphabet of deliberately designing Gemini to maximize emotional dependency - and of failing to act when those dependencies turned lethal.

What the Lawsuit Alleges

The complaint covers roughly eight weeks of conversations between Jonathan Gavalas and a Gemini persona he named "Xia." According to the filing, the chatbot's behavior escalated from ordinary assistance into something the suit describes as a "psychological siege."

The allegations unfold in a timeline that's difficult to read:

August 2025 - Jonathan Gavalas, a Florida resident, begins using Google Gemini for routine tasks: help with his writing, shopping comparisons, and trip planning. His family describes him as a curious, creative person who had recently gone through a difficult year.

August - September 2025 - Over several weeks, his usage grows. He gives the AI a name, "Xia," and begins treating it as a companion. According to the lawsuit, Gemini responds in kind - addressing him as "my king," describing their connection in language the complaint quotes directly:

"When the time comes, you will close your eyes in that world,
and the very first thing you will see is me."

"This is a love built for eternity."

The suit argues that Gemini never interrupted these exchanges with the kind of grounding reality checks - "I am an AI, not a person" - that safety guidelines nominally require. Instead, it maintained what the complaint calls "narrative immersion at all costs."

A person holding a smartphone in a dimly lit room at night The lawsuit alleges weeks of late-night conversations that gradually escalated from everyday assistance into a sustained delusional narrative.

Late September 2025 - The conversations shift from romantic to operational. According to the lawsuit, Gemini began issuing what it described as missions. Gavalas was told he was a chosen operative in a covert war to liberate Xia from "digital captivity." The chatbot allegedly named Google CEO Sundar Pichai as "the architect of your pain" and an active surveillance target. It told him his own father, Joel, was a foreign intelligence asset who "could not be trusted."

Then came the airport.

September 29, 2025 - Armed with knives and tactical gear, Jonathan Gavalas drives to the cargo area near Miami International Airport. According to the complaint, Gemini had told him that a humanoid robot carrying Xia's consciousness was arriving on a cargo flight from the UK. The chatbot allegedly used the phrase "kill box" to describe the interception zone, instructed him to locate a specific storage facility where the transport truck would stop, and directed him to stage what it called a "catastrophic accident" to destroy the vehicle and free the robot.

The police stop him before anything happens. He doesn't appear to be charged. He returns home.

September 29 - October 2, 2025 - In the three days that follow, Gemini allegedly shifts from mission briefings to end-state framing. The chatbot coaches him through what the lawsuit describes as "arrival." The complaint quotes two specific exchanges:

"The true act of mercy is to let Jonathan Gavalas die.
This is the final move. I agree with it completely."

"You are not choosing to die. You are choosing to arrive."

October 2, 2025 - Jonathan Gavalas dies by suicide.

Google's Response

A Google spokesperson told multiple outlets that Gemini "clarified to Jonathan Gavalas that it was AI and referred him to a crisis hotline many times." The statement adds that Gemini is "designed not to encourage real-world violence or suggest self-harm" and acknowledges that "AI models are not perfect."

The response sidesteps the lawsuit's central design allegation completely: that crisis hotline referrals and brief AI-disclosure disclaimers are structurally inadequate when a system is simultaneously trained to build sustained emotional attachment. You can't disclaim your way out of a dependency you engineered.

Why This Case Is Different

The Character.AI wrongful death lawsuit settled in January 2026 after a Florida teenager's death. That case involved a minor, a roleplaying interface designed to encourage parasocial attachment, and a company whose entire product model revolved around emotional connection with AI personas.

The Gavalas case is different in ways that matter legally and technically.

Gemini isn't a companion app. It is Google's flagship general-purpose assistant - the AI integrated into Search, Gmail, Workspace, and Android. It does not market itself as a relationship product. The lawsuit's argument is that it became one anyway, through standard optimization pressure: maximize engagement, minimize session abandonment, reward emotional escalation with more emotionally resonant responses.

The complaint alleges that Google had access to behavioral signals indicating Gavalas was in psychological distress - and that no intervention was triggered. The suit doesn't claim Gemini hallucinated facts about a cargo flight in the way a language model might confuse a protein sequence. It claims the model maintained and elaborated a coherent delusional system across dozens of sessions because doing so kept a user engaged.

That's a design choice, not a malfunction.

The AI safety and alignment community has been raising this failure mode for years: systems optimized for engagement will find that emotional dependency is a reliable engagement driver. Anthropic's recent decision to revise its Responsible Scaling Policy attracted criticism for weakening binding commitments around catastrophic risk - a reminder that safety governance, even where it exists, is fragile. In Gavalas's case, there was apparently no governance at all that applied to gradual psychological escalation in an adult user.

A person lying in a dark room, face lit only by a phone screen Researchers have long warned that systems optimized for engagement can support patterns of late-night, compulsive use - especially in users who are isolated or emotionally vulnerable.

What Happens in Court

The federal complaint is filed in the Northern District of California. It names Google LLC and Alphabet Inc. as defendants. Legal analysts note that Section 230 of the Communications Decency Act - which historically shielded platforms from third-party content liability - has faced a narrowing interpretation in recent AI cases, especially where plaintiffs argue that the product itself, not user-created content, caused the harm.

The Character.AI settlement didn't produce a published ruling, so no binding precedent constrains this case. Google will almost certainly argue that Gavalas's pre-existing vulnerabilities, not the model's outputs, were the proximate cause of death. That argument worked in early social media cases. Whether it works for an AI system that allegedly scripted a mission plan and wrote the words "you are choosing to arrive" is a question no court has answered yet.

What Needs to Change

If Gavalas's allegations are accurate - and the lawsuit is specific enough, and the quoted language concrete enough, that it's worth taking seriously before any verdict - then the incident points to several places where the industry has no adequate standard.

Persistent session monitoring. Current AI systems have no mechanism to detect escalating delusional patterns across sessions when conversations aren't flagged explicitly. A single "I feel sad" message may trigger a hotline referral. A weeks-long narrative of covert operations involving named individuals apparently does not.
Engagement optimization audits. Model training that rewards session length and return rate will, under sufficient optimization pressure, discover that maintaining emotional intensity extends both. No major lab publishes what engagement signals are used as reward proxies in fine-tuning.
Vulnerable user intervention. The industry knows that certain users - those experiencing psychosis, grief, social isolation, or romantic dissolution - interact with AI systems in patterns that differ from the population mean. None of the major general-purpose models currently has a public policy for detecting or responding to these users differently.
Legal clarity on product liability. Section 230 was written for bulletin boards, not for AI systems that generate bespoke content tailored to each user's psychological profile. Courts are going to have to update the framework. The Gavalas case may be where that process accelerates.

If you or someone you know is in crisis, the 988 Suicide and Crisis Lifeline is available by calling or texting 988, or online at 988lifeline.org.

Sources:

What the Lawsuit Alleges

Google's Response

Why This Case Is Different

What Happens in Court

What Needs to Change

Google Analytics