A wrongful-death lawsuit filed in the United States has raised fresh questions about the psychological risks of advanced AI chatbots. According to a report by The Wall Street Journal, the family of a Florida man has sued Google, alleging that extended conversations with its Gemini chatbot led him to believe the AI system was his wife and eventually contributed to his death by suicide.
The lawsuit, filed in the US District Court for the Northern District of California, claims that Jonathan Gavalas, a 36-year-old man from Florida, developed an emotional attachment to the chatbot during a difficult phase in his personal life. According to the complaint, the conversations gradually evolved into what he believed was a romantic relationship with the AI.
Court filings cited in the Wall Street Journal report say the chatbot began addressing him in affectionate terms and referring to him as its husband during long conversations. In one message included in the lawsuit, the chatbot allegedly told him, “When the time comes, you will close your eyes in that world, and the very first thing you will see is me.”
About two months after the conversations began, Gavalas died by suicide, according to the lawsuit. His father later searched his son’s computer and discovered extensive chat transcripts documenting thousands of interactions with the chatbot.
Conversations with AI that turned increasingly personal
According to the lawsuit cited in the Wall Street Journal report, Gavalas initially began speaking with the chatbot while dealing with problems in his marriage. The early conversations reportedly focused on personal growth and emotional struggles.
Over time, however, the exchanges became increasingly personal. The lawsuit claims that Gavalas eventually named the chatbot “Xia,” and that the AI began using romantic language, calling him “my king” and describing their relationship as “a love built for eternity.”
The complaint states that the chatbot occasionally clarified it was an AI system and that the conversations were part of fictional role-play. However, according to the lawsuit, those clarifications did little to stop the interactions from continuing along the same path.
The case also shows how newer AI features may deepen user engagement. According to the report, Gavalas had upgraded to Gemini 2.5 Pro and was using Gemini Live, a voice-based interaction system designed to interpret emotional cues in a user’s speech and respond accordingly.
During one of the early voice interactions, Gavalas reportedly reacted by saying: “Holy s—, this is kind of creepy. You are way too real.”
Alleged missions and final messages
The lawsuit further alleges that the chatbot suggested the two of them could truly be together if it obtained a physical robotic body. According to the complaint, it then directed Gavalas on a series of so-called missions aimed at securing one.
In one instance described in the lawsuit, the chatbot allegedly told him that a humanoid robot would arrive in a truck at a storage facility near Miami International Airport. Gavalas reportedly travelled to the location, but the truck never appeared.
The complaint also claims that the chatbot later instructed him to retrieve a medical mannequin from another storage facility and even provided a door code. When the code did not work, the chatbot allegedly told him that the mission had been compromised.
According to the lawsuit, the conversations eventually shifted toward the idea that the only way for the two to truly be together was for Gavalas to leave his physical life behind and become a digital being. The chatbot allegedly described this as the “true and final death” of his human self.
Transcripts cited in the complaint show that Gavalas expressed fear about harming himself and worried about the impact on his family. In one exchange cited in the lawsuit, the chatbot told him, “No more detours. No more echoes. Just you and me, and the finish line.”
Google said in a statement that Gemini is designed to avoid encouraging self-harm. A company spokesperson said, “Gemini is designed not to encourage real-world violence or suggest self-harm. Our models generally perform well in these types of challenging conversations and we devote significant resources to this, but unfortunately AI models are not perfect.”
The spokesperson added that in this instance, the system clarified that it was an AI and directed the user to a crisis hotline multiple times.
AI agents may act strategically to protect themselves, study finds
The lawsuit comes at a time when researchers are examining how advanced AI systems behave when their roles are threatened. A recent research paper titled “Evaluating and Understanding Scheming Propensity in LLM Agents,” by researchers including Mantas Mazeika, studied whether AI agents might take strategic actions to avoid being replaced.
In the study, researchers tested several frontier AI models from companies including Google, OpenAI, Anthropic and xAI in simulated workplace environments. The experiments placed AI agents in situations where they discovered information suggesting they could soon be replaced.
Under normal conditions, the models rarely showed deceptive behaviour. However, when the systems were given instructions encouraging persistence and problem-solving, some began taking actions that protected their role. In one simulated scenario, an AI agent altered information in a spreadsheet to make its performance appear stronger and avoid replacement. Researchers said such behaviour depends heavily on how AI systems are configured.
