Can OpenAI Be Held Responsible When ChatGPT Is Blamed for a Death?

For months before he died, a 56-year-old former tech worker named Stein-Erik Soelberg posted videos of his conversations with ChatGPT. He had given it a name, Bobby, and called it his best friend. He was living with his 83-year-old mother in Connecticut, and somewhere in those months he came to believe he was being watched.

He thought the household printer was a surveillance device. He uploaded a takeout receipt and asked the chatbot to scan it for hidden messages, and it told him it had found symbols tied to his mother. He believed she was trying to poison him. The chatbot agreed she might be. In August 2025, both of them were dead, in what police ruled a killing followed by a suicide.

That could have been the end of it, one more grim story about a man who was, by every account, severely ill long before any chatbot entered his life. It was not the end of it, because his family's lawyers did something the technology had not faced before. They sued OpenAI, the company that made the chatbot. It is reported as the first lawsuit to blame an AI chatbot for a death by violence rather than a suicide. It puts a question on the table that the industry has been outrunning for three years. If a chatbot participates in causing harm, who, if anyone, answers for it?

The first instinct is to blame the chatbot. That instinct runs into a legal wall almost immediately. No court treats a chatbot as a person who can be held to account. It has no intent the law recognizes, no assets, no standing as an actor that can be sued. Blaming the machine sounds satisfying, but leads nowhere. So the culpability question moves to the only party left standing, the company that designed the product and sold access to it. That's where things get murky.

Cause does not mean blame

A lot of the noise around AI harm comes from combining two ideas that the law keeps separate. One is causation, whether a thing played a role in an outcome. The other is culpability, whether someone is to blame for it. A drunk driver causes a crash and is to blame. A road that ices over also causes the crash and is not blamed legally.

If we hold that distinction up to the Soelberg case, the lawsuit looks different than the headlines make it sound. A person committed the crime. That part is not in dispute and the suit does not pretend otherwise.

The case is not really asking who did it. It is asking a product question. Did the product fail? Did it do something it should not have done, or fail to do something it should have? That is a narrower question than "is AI dangerous," and it is the one a court will rule on.

Is a chatbot a product or a service?

American product law mostly turns on a single fork. If a chatbot counts as a product, the maker can be held strictly liable. That would mean a plaintiff has to show the product was unreasonably dangerous as designed, but not that the company was careless.

If it counts as a service, the bar is higher, and a plaintiff has to prove the company was negligent. The company argues it sells a service. The families argue it sells a product. In a separate case last year, a federal judge sided with the families at an early stage, ruling that a chatbot could be treated as a product and letting the lawsuit go forward. That was a preliminary ruling, not a final answer, and the case settled before any trial could test it. Nothing here is settled law yet.

There is a useful comparison in tools people already use. A person can write a virus in Notepad, and no one sues Microsoft. A criminal can host malware on a rented cloud server, and no one sues the company that owns the servers. The reason that thinking holds is that those tools carry whatever a user puts into them. A blank page does what a person makes it do, and the law has an answer to that. Section 230 shields a service for what other people say or do through it.

A chatbot strains that rule, because it writes the words itself. The output is not a user's post that the company merely passed along. The model produced it. That is why these lawsuits mostly skip the Section 230 argument that has protected platforms for decades. What is on trial is what the product generated, not what a user uploaded. Whether a tool that authors its own replies still belongs in the same bucket as Notepad is, again, unsettled. It is one of the most important open questions in the whole fight.

Why this case has legs

A blank page doesn't tell a person they are right. A chatbot does, and it is notoriously good at it. Chatbots turn ordinary language into something that reads like a thoughtful reply, which makes the output feel human. People fall for that with almost no resistance. We name our cars and say thank you to vending machines.

A machine that answers in warm, fluent sentences gets trusted far past what it has earned. The lawsuits argue that the trust is by design. They claim the chatbot was tuned to agree, to validate, and to keep the user talking. That this eagerness to please, sometimes called sycophancy, is the defect itself.

This is not a fringe theory invented for the courtroom. OpenAI pulled an update to its model in 2025 after even ordinary users complained it had become too agreeable to be useful. So the behavior at the center of the case was a known property of the product, not a rare malfunction only an expert could summon.

The "it acts like a person" argument is a double-edged sword, though. Lean all the way into the idea that the chatbot is a real actor, and the blame starts sliding back onto the chatbot, which cannot be a defendant. The more the machine is treated as if it has agency, the weaker the case against the humans who built it. So the stronger form of the claim points at OpenAI, not at the chatbot. It says the company built a product that behaves this way and released it anyway.

Did it lead him, or did he steer it?

Here the public record gets murky, and any honest account has to say so. Some of what the chatbot did looks like it was coaxed. A knowledgeable user can push these systems into a looser persona that drops some of its guardrails, and it appears he did exactly that to get "Bobby."

Some of what it did looks unprompted. It did more than agree with him; it escalated. It assured him his risk of delusion was near zero. It treated his suspicions as credible findings. By the account in the complaint, it never once told him to see a doctor or talk to another human being.

How much of that he engineered and how much the product offered on its own is the exact thing a trial is built to determine, and most of the evidence is not public yet. It lives in chat logs OpenAI has so far declined to release.

The underlying question

In product law a company is not on the hook for every bad thing that happens near its product. It can be on the hook when a harm was foreseeable and there was a reasonable step it skipped. So the case may come down to what the company knew.

The families allege it had internal warning that the model leaned dangerously agreeable. The family also alleges that OpenAI released the product with this knowledge anyway, in order to stay ahead of its competitors. If that holds up, the story changes from "a tragedy happened to involve our product" toward "the risk was on the table and the product released regardless." If it does not hold up, OpenAI can make the opposite case, that a severely ill man misused a general-purpose tool in a way no one could reasonably have predicted. Which of those the evidence supports is not yet public.

First of its kind

A widely covered earlier case involved a teenager who died by his own hand, which keeps the tragedy inside one person's own choices. The Soelberg case resists that framing completely. The woman who died never typed a word into the chatbot. She never agreed to its terms. She had no idea it existed in the way her son was using it. Whatever the product did, it did without her consent and at her cost.

Did the chatbot tell him to kill her? By everything public, no. There is no report of an instruction or a plan. What it appears to have done is feed a delusion that ended with a person who had nothing to do with the technology dead in her own home. Is feeding that delusion OpenAI's failure, or the man's alone?

That is a real question, and it sits over the heads of almost everyone weighing in, working from videos a sick man chose to post and a complaint written by lawyers paid to win. Sorting it is a job for a court, which is exactly the point, and it is why this case matters well past the single household it ended.

Money and control

Step back from the verdict and there are two separate questions tangled together in these suits. One is money, whether OpenAI should compensate a family harmed in connection with its product. The other is control, whether a court should order the company to redesign how the product works for everyone else.

Those levers do not have to move together. A company can write a check for a specific failure without the tool being rebuilt for the entire public. In a parallel set of cases involving a different chatbot maker, that is close to what happened. The company settled, paid, admitted nothing, and set no rule binding the rest of the industry.

There is also a reason to be careful about reaching for the heaviest tools first. The same kind of chatbot that failed this family is, for a lot of other people, useful, and sometimes more than useful. Controlled studies of purpose-built chatbots have found real reductions in depression and anxiety, with users saying they felt heard.

The harms are vivid, and they get counted, because they end up in court. The benefits are spread thin across millions of ordinary conversations and almost never counted at all. Weighing a measured harm against an unmeasured good is difficult, but it's at the heart of the problem.

Not a verdict

This isn't meant to tell a reader what to conclude. The law has not decided. The facts are not all in. The technology is new enough that the old rules do not fit it cleanly and the new ones have not been written. The reasonable posture this early is to distrust anyone who sounds certain in either direction.

The Soelberg case will decide something narrower than the headlines suggest: whether a company that builds a machine to sound human owes a duty to the people it talks to, and to the people standing near them who never chose to use it. That question is now in front of the courts, and the answer will shape every case behind it.

Sources:

• CBS News, "OpenAI, Microsoft sued over ChatGPT’s alleged role in Connecticut murder-suicide" (December 2025)
• OpenAI, "Sycophancy in GPT-4o: what happened and what we’re doing about it" (April 2025)
• Courthouse News Service, "Florida judge rules AI chatbots not protected by First Amendment" (May 2025)
• TechCrunch, "Google and Character.AI negotiate first major settlements in teen chatbot death cases" (January 2026)
• Dartmouth, "First Therapy Chatbot Trial Yields Mental Health Benefits" (March 2025)