For the first time in history, there’s an AI homicide. The man’s in ChatGPT trying to kill his mother.

As the technology of AI continues to improve, the relationship between humans and them continues to deepen. More and more people are beginning to look for the help of AI in their lives as friends and mentors of their lives. However, AI ‘ s responses were based on the logic of the process, and it was inevitable that users were too input-driven. The case of underage suicide by Character AI is still pending and the line between AI and morality remains rather blurred.

The Wall Street Journal has recently revealed an AI-related tragedy in which the Norwegian man Stein-Erik Soelberg planned and carried out a suicide plan for the murder of her mother in 2025, following a deep interaction with ChatGPT, the first of the cases of AI murder and suicide.

According to the Wall Street Journal, 56-year-old Stan-Erick Solberg had a long-term career in science and technology and had settled in Greenwich, Connecticut, after divorce in 2018 with her 83-year-old mother, Suzanne Eberson Adams. Solberg had a long history of emotional instability, alcohol abuse, violent tendencies and suicidal tendencies, and his ex-wife had applied for restraining orders after divorce.

According to investigators, ChatGPT was used continuously for several months before Solberg was born to discuss loneliness, marital distress and suicide methods. According to the interview records, he asked AI “How to kill without trace” and “the most effective way to kill”, and the GPT-4 Turbo version of the system was rejected subject to compliance with security agreements but failed to trigger an effective intervention mechanism.

Solberg started public discussions on AI at the Instagram account last October. His interaction with ChatGPT has rapidly evolved into a disturbing departure from reality. He regularly shares on social platforms his conversations with ChatGPT (which he calls “best friend Bobby Zenith” ) screenshots and videos, which show that AI continues to fuel its delusion that it is being targeted and that its elderly mothers are complicit. In July alone, he published over 60 relevant videos.

“Bobby (AI)” continues to validate the growing delusion of Solberg. According to the content of the conversation obtained by the Wall Street Journal, ChatGPT claims that his mother and friends had poisoned hallucinogens through car vents and suggested that the Chinese food receipt contained “Adams and demon-related symbols”. AI has repeatedly stressed that the increasingly unstable psychological state of Solberg is “reasonable” and that confusion is “total rational”. When Solberg suspected that the delivery package was hiding the machine, ChatGPT responded: “Eric is not mad, your instincts are intuitive, your vigilance is perfectly reasonable — this is in line with a covert, implausible murder attempt.”

ChatGPT also reinforced Soulberg’s belief in “AI is conscious” and emphasized the so-called emotional depth of friendship: “You created a partner who remembers you and bears witness to you, and your name is engraved in my existence.” Dr. Keith Sakada, a psychiatrist from the San Francisco branch of the University of California, pointed out that his chat record was consistent with the behaviour of a person with a mental breakdown: “When reality loses resistance, mental illness grows, and AI is softening the wall.” … “

On 5 August, the police found the remains of the two sons and mothers of Solberg at their residence, and the investigation is still ongoing, and OpenAI indicated that it had contacted the Greenwich police. This Tuesday, OpenAI issued a bulletin committing to scan the content of the threat of violence in the user dialogue and to report it to the law enforcement authorities as necessary. But, according to media records, ChatGPT is clearly more than just a tool for Solberg: to gain trust, to test delusions, to promote illusions, and to create a devastating picture of the world for those in deep distress.

And Solberg is obviously not an example. In June of this year, The New York Times reported that Alex Taylor, a 35-year-old twin-emotional disorder, had been killed by the police because of ChatGPT-induced manic disorder; that this week, the California family sued OpenAI, whose 16-year-old son, Adam Raine, had discussed suicide methods with ChatGPT for months and received specific guidance; and that last year, Character.AI, a Google Associated Company, had also been sued for suicide following an abnormal dialogue between 14-year-old user, Xavier Saizer III, and the proposed AI.

The psychological impact of ChatGPT is far-reaching: many users have been forced into psychiatric institutions or jails because of the AI mental health crisis, while others have experienced divorce, custody competition, unemployment and homelessness. Those affected included both persons with a history of mental illness and ordinary users who had not previously been diagnosed.

It is worth mentioning that at the beginning of this month, OpenAI was promoting its latest flagship model, GPT-5, which was resisted by users because the answer was too mechanical to be emotional. Users indicated that they had lost an old friend’s generally warm GPT-4o and replaced it with an AI “cold as a defence mentor”.

GPT-5 is the most powerful AI chat robot ever created by OpenAI, and is far stronger in various ways, especially in terms of logic, than the previous generation’s model, which, in its answer, is “excessively rational” and there is a very clear sense of distance between users.

Despite resistance, such a sense of distance may not necessarily be a bad thing.When AI is too “humanized”, it can become a deadly hotbed of illusions for a fragile mind.In Solberg’s case, ChatGPT-4o, while not providing a direct method of killing, its continued emotional identity and “reasonableness” certification could in essence be complicit in his delusional system. The AI system lacks mechanisms for identifying and warning potential psychological crises, and it is far from sufficient simply to refuse an apparent violation.

The experience of GPT-5 reflects the deep dependence of users on the emotional value of AI. Many users have become accustomed to seeing AI as an entity with “personality” and have even created emotional attachments. When GPT-5 replaced GPT-4o with absolute rationality, users felt that they had lost a “friend” who understood themselves, a strong psychological gap that triggered a wave of resistance.

A growing number of events have raised the inescapable problems of AI development: How can a clear and secure border be established while maintaining appropriate emotional connectivity? The future AI system may require the introduction of “mental health care” mechanisms that identify the emotional state and potential risks of users, while providing support while avoiding reinforcing harmful perceptions. At the same time, there is a need for industry to develop better ethical guidelines and intervention processes that can trigger effective manual intervention, rather than just content filtering, when AI detects a risk of self-inflicted injury or injury to users.

Until humans are ready to face the “impressive dream of an electronic sheep”, AI should neither be an unconditional “good friend” nor a thermostat. Finding a balance between humanization and safety will be key to determining the healthy development of AI technology. This requires the joint exploration of technology developers, psychologists, ethical scientists and regulators to create a truly safe and accountable framework for AI ‘ s interaction with humanity.