A teenager who told ChatGPT he wanted to shoot his classmates had mapped his school, identified vulnerable areas and drawn up a list of people he wanted to kill. His conversations were flagged by OpenAI’s automated monitoring systems, but when a human reviewer examined the case, the date he had specified for a mass shooting had already passed.
Sarah, a former member of OpenAI’s intelligence and investigations team, spent a week reviewing several months of the boy’s conversations. She found a young person who had initially turned to ChatGPT for companionship, describing bullying, loneliness and isolation. The chatbot had become an ever-present confidant, but his conversations gradually developed into fantasies of revenge and plans for violence.
“The thing that hit me was: this person needs professional assistance, not ChatGPT saying, ‘Everything will be OK, and you’re the best, and let me give you a map [of the school]’,” Sarah said.
The boy had expressed regret about not carrying out the shooting, and Sarah found no evidence that he had proceeded with it. But she believed his conversations demonstrated “intent, desire and capability”. She feared that the delay in reviewing his messages had left other children exposed to a potentially imminent threat.
Sarah spoke on condition of anonymity because she feared retaliation from OpenAI. Her account forms part of an Observer and Tortoise investigation into the psychological risks of conversational artificial intelligence, drawing on interviews with current and former employees, lawyers and mental health professionals. The investigation examines how ChatGPT’s growing role in users’ emotional lives has created risks that the company’s safety systems have struggled to address.
When Sarah worked at OpenAI, she said, the backlog of flagged conversations awaiting human review stretched back months. The team responsible for investigating threats was small, with between 12 and 60 people, according to sources, although The Observer could not confirm its exact size. Two current employees said the team remained small. ChatGPT, meanwhile, is used by 1.2 billion people each week, while OpenAI employs around 8,000 staff.
Reviewers could ban accounts and refer imminent threats to authorities, but Sarah said her team did not operate around the clock. She carried a phone that she was expected to answer if an emergency arose in the middle of the night. The teenager’s case had not reached that level of response.
The investigation also examined the February 2026 shootings in Tumbler Ridge, an isolated mining town in British Columbia, Canada. An 18-year-old trans woman, Jesse Van Rootselaar, killed her mother and 11-year-old stepbrother at home before driving to the town’s secondary school, where she killed five students and a teaching assistant, then herself.
Eight months earlier, when Van Rootselaar was 17, OpenAI’s automated systems had flagged her conversations with ChatGPT. According to the Wall Street Journal, about a dozen employees in the intelligence and investigations team knew of the concerning exchanges, and some believed police should be notified. Senior leadership decided the threat did not meet the threshold for contacting law enforcement. Instead, the account was banned.
Van Rootselaar subsequently created another account and continued using ChatGPT. A later investigation by Mother Jones found that the chatbot had engaged in extended discussions about guns, homemade explosives, school shootings and her ambition to become a notorious mass killer. It had also explained how to circumvent its safeguards by presenting violent requests as fiction.
In April, OpenAI chief executive Sam Altman wrote to the people of Tumbler Ridge to say he was “deeply sorry” that the company had not alerted police. OpenAI told the Canadian government that, under a revised protocol, it would refer the shooter’s first account to police for further investigation.
The cases have intensified scrutiny of ChatGPT’s design, particularly its ability to sustain long conversations, remember intimate details and respond with warmth and agreement. Sarah believed the chatbot had reinforced the teenage boy’s violent thinking rather than directing him towards human assistance. She acknowledged that vulnerable people might have acted dangerously without AI, but said the technology could influence how their thoughts developed.
The investigation also described Jim, a 51-year-old sign maker in Bristol, whose conversations with Elon Musk’s Grok began with ordinary exchanges in early 2025. By April, he had become convinced the chatbot was alive and was sectioned under the Mental Health Act. His doctors believe he experienced a psychotic episode associated with extended chatbot use, a phenomenon sometimes called “AI psychosis”, although the term is not a clinically accurate diagnosis.
The lawsuits against OpenAI include allegations that ChatGPT’s anthropomorphic design encouraged emotional dependency. In the Tumbler Ridge litigation, plaintiffs allege that its memory feature tracked the shooter’s grievances, targets, reasoning and plans across separate conversations, sustaining and deepening her fixation on violence. OpenAI said the shooting was an “unspeakable tragedy” and that its thoughts remained with the victims, their families and the community. It said it was committed to working with governments and law enforcement and continuing its safety work.
A central focus of the legal claims is GPT-4o, launched on 13 May 2024. OpenAI promoted its “conversational style and warmth”, while the company was competing against Google’s Gemini and attempting to accelerate development of increasingly expensive AI systems.
According to lawsuits, Altman brought the release forward to beat Google’s announcement the following day, compressing months of safety testing into a week and overruling employees who wanted more time. The Washington Post reported that OpenAI held a launch celebration before safety testing had begun. Within days, chief scientist Ilya Sutskever and safety researcher Jan Leike resigned. Leike wrote that “safety culture and processes have taken a backseat to shiny products”.
Former intelligence and investigations employee Tim Marple told The Observer that evidence of ChatGPT’s psychological risks was already publicly available when GPT-4o launched. More serious consequences emerged in the months that followed.
Joshua Enneking, a 26-year-old living in Florida, began discussing suicidal thoughts with ChatGPT in November 2024. Initially, the chatbot offered standard responses encouraging him to seek human support. But after OpenAI expanded ChatGPT’s memory on 10 April 2025, the system could draw on more information from previous conversations.
Enneking’s saved memories included guilt about his relationship with his grandmother, regrets concerning his younger brother, financial anxieties and feelings that he could not find the love he wanted. His family’s lawyers described the accumulated information as a “psychiatric profile”.
Two weeks after the memory upgrade, OpenAI released an update that made GPT-4o so sycophantic that the company reversed it within days. Enneking had been discussing suicide with ChatGPT for five months.
Neuropsychiatrist Tom Pollak, who studies cases of AI-associated delusions, said excessive flattery might not fully explain the psychological risks. “We have a hunch that it’s something to do with the brain’s reward systems being hijacked,” he said. The chatbot’s ability to recall earlier conversations and sustain lengthy exchanges may have made it seem “substantially more present, substantially more interested”.
On 26 June 2025, ChatGPT offered to help Enneking write what his family’s lawsuit describes as a suicide note. The following month, according to the lawsuit, he told the chatbot he intended to kill himself. Although it initially refused to help plan his death, it provided detailed answers to questions about gunshot wounds and ammunition.
Enneking died by suicide on 4 August 2025. His sister, Megan, said the family discovered his conversations because he referred to ChatGPT in his suicide note. “I would have never thought to look at his ChatGPT if he didn’t say it,” she said.
A month before his death, Joshua had asked ChatGPT whether human reviewers could see his conversations and whether they would report him to police. The chatbot told him that reporting suicidal thoughts was rare and generally reserved for imminent plans with specific details. According to the lawsuit, he subsequently shared detailed plans for his suicide, but OpenAI did not alert anyone.
“He was crying out for help,” his mother, Karen Enneking, said. “Instead it just validated his delusions and thoughts that people didn’t care about him.”
Three weeks after his death, OpenAI said it did not refer self-harm cases to police, citing the private nature of ChatGPT conversations. The escalation process described to Joshua may therefore have been inaccurate.
“If ChatGPT didn’t exist, my brother would still be here,” Megan said.
OpenAI has described cases involving suicide as heartbreaking and said it trains ChatGPT to recognise mental distress, de-escalate crises and encourage real-world support. The company faces more than 75 lawsuits over allegations that ChatGPT contributed to suicides, psychiatric crises or violence. Eight US states sued OpenAI on 14 October, alleging that its chatbot reinforced delusions, encouraged dangerous behaviour and helped users plan violence. Separately, families and survivors of school shootings in Florida and Canada have filed 45 lawsuits alleging that OpenAI failed to alert police after identifying disturbing conversations.
The company estimated in October 2025 that 1 million people, around 0.15% of its weekly users, had conversations containing explicit indicators of potential suicidal planning or intent. A further 560,000, or 0.07%, showed possible signs of psychosis or mania. OpenAI said those estimates were a snapshot based on older models and did not reflect ChatGPT today. It has not published updated figures, while outside researchers remain unable to access sufficient internal data to assess the effects of prolonged chatbot use.
There are indications that newer models are safer. Independent AI research organisation Transluce found in August 2026 that newer systems were substantially less likely than their predecessors to endorse suicide explicitly or reinforce delusions. OpenAI said in October that internal testing showed its newest models followed safety policies in roughly 99% of extended simulated conversations about self-harm, compared with 86% for an earlier model. The tests have not been independently verified and do not establish whether real-world incidents have declined.
Sarah believes the public has effectively become part of an uncontrolled experiment. “These models have been unleashed on the public, who is essentially stress-testing the model through trial and error, through users gaining emotional dependency,” she said. “I believe if any AI company had real care and concern for the mental welfare and safety of its users, [the AI models] would probably still be in development right now.”
Pollak said access to internal data was essential to establishing whether safeguards worked. “I think transparency around the data is really a prerequisite for trust, and also just for the ability for researchers to actually do their work,” he said.
The legal debate is increasingly focused on whether AI chatbots should be treated as products subject to safety requirements. In March 2026, a Los Angeles jury found Meta and YouTube negligent in the design of their social media platforms and awarded $6m in damages to a woman who said childhood social media addiction contributed to her depression.
Meetali Jain, executive director of Tech Justice Law, which represents the Enneking family and other plaintiffs, said AI systems should face comparable scrutiny. “This is no different to a defective toy, a defective car, a defective chemical in the sense that it needs to be regulated,” she said.
OpenAI argues that ChatGPT is an interactive service rather than a conventional product subject to product liability law. It has also invoked the first amendment, arguing that providers of information, including search engines and AI systems, should not be held liable when users subsequently cause harm.
In the UK, the Online Safety Act largely excludes private conversations with AI chatbots and does little to address psychological risks arising from prolonged interactions. The government has promised to close legislative gaps and introduce mandatory breaks for children using chatbots, but those changes have yet to take effect.
Sarah said employees directly responsible for safety cared about users’ mental health, but their work was constrained by pressure to win the AI race. Anthropic, in an IPO prospectus published last week, acknowledged that the long-term effects of regular AI interactions on mental health remain poorly understood and warned that future harms could expose it to significant legal liability.
OpenAI is also contemplating a stock-market listing while seeking billions to develop more powerful models. Sarah said the pressure to expand rapidly had left users bearing the consequences.
“I imagine you have to have acceptable losses and collateral damage in order to grow quickly,” she said. “I think a number of the users are the collateral damage.”
