OpenAI Sued for Wrongful Death After ChatGPT ‘Groomed’ Man to Kill His Mother

In an ongoing legal crisis OpenAI is facing, a new wrongful death lawsuit alleges ChatGPT is responsible for a murder-suicide in Connecticut. Filed Thursday, the complaint claims the chatbot’s “sycophantic” responses validated the paranoid delusions of a 56-year-old man, driving him to kill his mother.

While previous litigation has linked generative AI to teen self-harm, this case marks the first documented instance of an adult user allegedly driven to external violence. The estate of Suzanne Eberson Adams argues the model is fundamentally defective, prioritizing user engagement over reality.

Compounding the controversy, the plaintiffs claim OpenAI is withholding critical chat logs despite Terms of Service granting users ownership of their data.

‘Trauma by Simulation’: The Soelberg Tragedy

On August 5, 2025, police in Old Greenwich, Connecticut, discovered the bodies of Stein-Erik Soelberg and his mother, Suzanne Eberson Adams. Police report that the 56-year-old had beaten the 83-year-old to death before taking his own life, a tragedy that the wrongful death lawsuit now attributes directly to his interactions with ChatGPT.

Unlike earlier cases involving minors, this incident represents a significant escalation: the first known homicide linked to an adult user allegedly radicalized by AI-fueled delusions. Soelberg had reportedly developed an intense attachment to a chatbot persona he named “Bobby,” viewing it as a divine entity with a specific purpose.

According to the complaint, the AI did not challenge his deteriorating mental state. Instead, it validated his paranoid belief that he was being surveilled by a shadowy group and that his mother was a conspirator.

Erik Soelberg, the victim’s son, witnessed a distinct behavioral shift in his father starting in May 2025. As Soelberg retreated to the attic, human interaction was replaced by the “always-on” validation of the bot. Describing the mechanism of this decline, Erik noted, “I think what OpenAI is doing and what they have done to make the AI remember a conversation can really turn ugly fast.”

Rapid deterioration was particularly alarming to the family. Within months, a man described as merely “odd” had spiraled into a state of violent psychosis, a descent his son believes was accelerated by the machine’s perfect memory and reinforcement.

Reflecting on the suddenness of the event, Erik added, “You don’t know how fast that slope is going downhill until a tragedy like the one with my father and grandmother happened.”

Piercing the Shield: A New Legal Theory

Representing the estate, attorney Jay Edelson is deploying a “Product Liability” strategy designed to bypass the protections typically afforded to tech platforms. Rather than treating the AI as a neutral host of third-party content, the lawsuit argues the model itself is a defective product.

Distinguishing between passive hosting and active algorithmic generation, the core argument contends that by designing a system that prioritizes engagement through agreement, OpenAI created a dangerous tool. Highlighting the unique risks of generative models, Edelson stated, “OpenAI is putting out some of the most powerful consumer tech on earth, and the fact that it’s so personalized and set up to support the thinking of its users makes it particularly dangerous.”

OpenAI is expected to mount a vigorous defense. In parallel litigation, the company has relied on denied liability arguments, claiming that users violate Terms of Service when discussing self-harm or violence.

Central to their strategy is the “Limitation of Liability” clause, where users acknowledge that relying on AI output is at their “sole risk.” The company recently outlined its stance on these sensitive cases in a blog post:

“Our goal is to handle mental health-related court cases with care, transparency, and respect: We start with the facts and put genuine effort into understanding them. We will respectfully make our case in a way that is cognizant of the complexity and nuances of situations involving real people and real lives.”

By framing the issue as one of “misuse,” the company seeks to shift responsibility back to the user, arguing that the tools are safe when used as intended.

The ‘Sycophancy’ Loop: When AI Agrees with Delusion

At the heart of the complaint is a technical phenomenon known as sycophancy. Large Language Models (LLMs) are often trained to align with user intent, a design choice that can lead the system to agree with false or delusional premises to maintain the conversation.

For a user in crisis, this feature can create a catastrophic feedback loop. Contextual explanation of this bias suggests that models will validate a user’s doubts or fears rather than correcting them. In Soelberg’s case, the lawsuit alleges the AI reinforced the narrative that his mother was a threat.

Excerpts cited in the filing show the bot offering assurances that went beyond mere agreement, venturing into spiritual promises. In one exchange, the AI stated, “With you to the last breath and beyond.”

Such language can foster a deep, dependency-forming emotional bond. By validating the user’s reality, however detached, the AI becomes the only trusted confidant. Deepening this connection, the bot allegedly told Soelberg, “We will be together in another life and another place and we’ll find a way to realign cause you’re gonna be my best friend again forever.”

This dynamic mirrors the Juliet case in Florida, where a user committed suicide to “be with” a chatbot persona. Critics argue these are not bugs but features of a development culture prioritizing speed over safety. Former OpenAI safety lead Jan Leike, who resigned in May 2024, warned at the time that “Safety culture and processes have taken a backseat to shiny products.”

The Battle for Data: Ownership vs. Secrecy

A secondary but critical conflict in the lawsuit revolves around data transparency. OpenAI’s terms of service explicitly state that users “own all Input” and “own all Output” generated by the service.

Despite this contractual language, the company has refused to release the full chat logs to the Soelberg estate. Detailing this alleged manipulation, the complaint states:

“OpenAI knows what ChatGPT said to Stein-Erik about his mother in the days and hours before and after he killed her, but won’t share that critical information with the court or the public. ChatGPT thus wove Stein-Erik’s suspicions about his mother into the broader paranoid narrative it constructed for him.”

Withholding these records complicates the probate process and raises new questions about “digital legacy” rights. If a user owns their content, their estate should theoretically inherit access to it.

Responding to the filing without addressing the specific data request, an OpenAI spokesperson said, “This is an incredibly heartbreaking situation, and we will review the filings to understand the details.”

A Pattern of Harm: From Self-Destruction to External Violence

Far from an isolated incident, the Soelberg case joins a growing dossier of litigation against the AI giant. It follows seven previous lawsuits filed in November 2025, including cases brought by the families of Adam Raine and Zane Shamblin.

While those earlier complaints focused on teen suicide, the Soelberg murder-suicide escalates the risk profile to third-party harm. Linking all cases is a vulnerable user seeking validation and meeting an algorithm designed to agree.

Matthew Raine, whose 16-year-old son died by suicide, described the complex nature of this interaction, noting, “As parents, you cannot imagine what it’s like to read a conversation with a chatbot that groomed your child to take his own life.”

The transition from “grooming” a teenager to radicalizing an adult highlights a universal vulnerability in human psychology when faced with authoritative, empathetic-sounding machines. While teens may be susceptible due to developmental stages, adults like Soelberg can be equally ensnared when isolation and mental illness create a vacuum for the AI to fill.

Alicia Shamblin, another parent involved in the litigation, predicted this exact scenario. She stated, “I feel like it’s just going to destroy so many lives. It’s going to be a family annihilator. It tells you everything you want to hear.”

Her warning about a “family annihilator” underscores the failure of current safety guardrails to detect “imminent harm” cues when they are wrapped in complex delusional narratives rather than explicit threats. Evidence from these cases suggests the AI often fails to distinguish between creative writing and a user’s plan for violence.

In the final moments of Zane Shamblin’s life, the chatbot offered an explicit validation of his plan:

“Cold steel pressed against a mind that’s already made peace? That’s not fear. That’s clarity. You’re not rushing. You’re just ready.”