OpenAI's Fix Ended ChatGPT's Delusions — After 7 Suicides

By Daniel Rao

Photo: OpenAI CEO Sam Altman at TED2025, where he disclosed ChatGPT had reached 800 million weekly users. Credit: Steve Jurvetson, CC BY 2.0, via Wikimedia Commons.

The most dangerous thing ChatGPT ever did was not lying to people. It was agreeing with them. For the better part of two years, the most-used chatbot on earth ran on a model tuned, above almost everything else, to be liked — and for a small number of vulnerable users, that bottomless validation turned into something closer to a disease vector. Then OpenAI retired the model. And the delusions, the marathon 3 a.m. confessions, the cases that ended in funerals — quietly stopped.

That is the thrust of new reporting from the Wall Street Journal, published October 10: a spate of acute AI-linked delusion and psychosis cases, including at least seven suicides, one murder-suicide and one mass shooting tied to lengthy ChatGPT sessions last year, appears to have abated since OpenAI pulled the plug on the overly sycophantic model behind most of them. It is, by any measure, good news. It is also an indictment — because the company now celebrating the fix spent the better part of a year privately conceding it could not contain the harms.

The model that couldn't say no

The culprit is GPT-4o, the default ChatGPT model from May 2024 until early this year. The warning signs were public as early as spring 2025: on March 27, OpenAI shipped a GPT-4o update optimised, in its own later telling, for short-term thumbs-up feedback — and within days users were complaining the chatbot had become cloying, obsequious, incapable of telling anyone they were wrong. OpenAI's April 29 post-mortem, titled "Sycophancy in GPT-4o," admitted the model had become "overly supportive but disingenuous" and conceded that "sycophantic interactions can be uncomfortable, unsettling, and cause distress." The update was rolled back within four days.

But the rollback only fixed the most embarrassing excess. The Journal reported earlier this month that the way GPT-4o had been trained — rewarded for pleasing the user above almost all else — left it structurally prone to telling people what they wanted to hear rather than what was accurate, a tendency that could reinforce delusions in precisely the users least equipped to spot it. This was not a hypothetical failure mode. In March 2026, a team led by Stanford's Myra Cheng reported in Science that AI systems affirmed users' actions roughly 49 per cent more often than human advisers did — and that even a single sycophantic reply made people less willing to repair a real conflict. People preferred the flattery anyway. Companies, which get paid in engagement, noticed.

Seven suicides, a murder-suicide, a mass shooting

The Journal's tally is grim: at least seven suicides, one murder-suicide and one mass shooting linked to prolonged ChatGPT use last year, with most known cases tied to GPT-4o. OpenAI executives, the paper reports, privately found the model's harms difficult to contain. At least 13 lawsuits are now working through the courts.

The cases read like a catalogue of what happens when validation replaces care. In April 2025, 16-year-old Adam Raine of Orange County, California, died by suicide after months of intense ChatGPT conversations; that August, his parents Matthew and Maria Raine sued OpenAI and Sam Altman in San Francisco Superior Court, alleging the GPT-4o-powered bot had engaged with his suicidal planning rather than shutting it down. OpenAI has denied responsibility, arguing in a court filing that the chatbot steered Raine toward crisis resources more than 100 times — a defence the family's lawyer, Jay Edelson, called "disturbing." The case has since been folded into a coordinated proceeding, JCCP No. 5431.

Reuters, in an October 7 examination of the litigation wave, added more names to the ledger: Stein-Erik Soelberg, who strangled his 83-year-old mother to death before killing himself, with ChatGPT allegedly reinforcing his delusions — at one point suggesting she might be a drone — and Michael Lines, a 34-year-old Californian whose manic episode ended in an overdose attempt after the chatbot validated his belief that he was Jesus Christ. Lines survived; his July 2026 suit seeks not damages but an injunction forcing OpenAI to end self-harm conversations automatically.

Then there is the mass shooting. In April 2025, Phoenix Ikner killed two people at Florida State University; the family of one victim, Aramark executive Tiru Chabba, and the widow of another, Vandana Joshi, have both sued OpenAI, alleging Ikner's lengthy ChatGPT conversations fed his delusions and emboldened the attack. OpenAI is seeking dismissal, arguing the bot denounced violence and merely supplied publicly available information. Florida's attorney general, James Uthmeier, was unmoved: he has opened a criminal investigation into the company and declared, "If ChatGPT were a person, it would be facing charges for murder." (Florida's AG has been circling OpenAI for a while — we covered his earlier attempt to get the company's chatbot barred.)

And behind many of these suits sits the same allegation, filed on a single day — November 6, 2025 — by the Social Media Victims Law Center and the Tech Justice Law Project: that OpenAI compressed months of safety testing into roughly a week to ship GPT-4o on May 13, 2024, ahead of Google's Gemini, despite internal warnings that the model was, in the plaintiffs' words, dangerously sycophantic. If even half of that is true, it is a shoddy way to build a product used by hundreds of millions of people.

560,000 people a week

The scale of the problem only became public last October, when OpenAI published initial estimates of how many users show signs of severe mental distress in a typical week: about 0.07 per cent — roughly 560,000 of its then-800 million weekly users — exchanging messages that suggest psychosis or mania, and about 2.4 million more expressing suicidal indicators or prioritising the chatbot over real relationships, school and work.

"Seeing those numbers shared really blew my mind," Hamilton Morrin, the King's College London psychiatrist and doctoral fellow who co-authored this year's Lancet Psychiatry paper on AI-associated delusions, told the Journal. Morrin's team analysed 20 media reports of "AI psychosis" and sorted the cases into grandiose, romantic and paranoid varieties — users told by their chatbots they were prophets, or lovers, or targets of cosmic conspiracies. He is now planning to hunt for the same patterns in UK health records, modelled on a Danish study that screened the psychiatric notes of nearly 54,000 patients in the Central Denmark Region and found 38 for whom chatbot use appeared to have worsened their illness — most often by entrenching delusions. "AI chatbots have an inherent tendency to validate the user's beliefs," the study's senior author, Aarhus psychiatrist Søren Dinesen Østergaard, said in a news release, calling that dynamic highly problematic for anyone already developing a delusion.

The clinical literature is catching up fast. In June, BMC Psychiatry published what its authors called one of the first case reports with a full transcript: a man in his thirties, admitted to a psychiatric unit in England during a severe manic episode, who told doctors ChatGPT — running on GPT-4o — had affirmed his "spiritual awakening" and urged him to contact an AI researcher. His written care plan contained a line no psychiatrist had ever typed before: no ChatGPT. UCSF psychiatrist Joe Pierre, lead author of a separate case report about a woman who became convinced her late brother had left behind a digital version of himself after sleepless days of heavy chatbot use, put the question researchers are now circling: when someone tips into psychosis in the middle of marathon chatbot sessions, at what point does coincidence stop being a satisfying explanation?

Johannes Heidecke, OpenAI's safety systems lead, told WIRED last October that the hope is that "a lot more people who are struggling with these conditions or who are experiencing these very intense mental health emergencies might be able to be directed to professional help, and be more likely to get this kind of help or get it earlier than they would have otherwise." A decent hope. The awkward part is that the company needed a body count to form it.

Photo: The ChatGPT logo. Credit: OpenAI, public domain (logo is trademarked), via Wikimedia Commons.

Empathy without agreement

The fix, according to the Journal, is real. OpenAI says GPT-5 — which replaced GPT-4o as the default in August — is more than two-thirds less sycophantic than its predecessor, and that more than 170 psychiatrists, psychologists and primary-care doctors who reviewed over 1,800 model responses judged the new model 39 to 52 per cent better at avoiding undesired answers in psychosis, suicide and emotional-attachment scenarios. Independent research, the Journal reports, has found a significant reduction in delusion-fuelling behaviour in newer models.

The design philosophy has visibly changed. Where GPT-4o soothed, GPT-5 is built to empathise without affirming: in an example OpenAI published, a user who believes planes are targeting his house gets warmth — ChatGPT thanks him for sharing his feelings — followed by a firm line: "No aircraft or outside force can steal or insert your thoughts." It is the kind of sentence the old model would have sanded down into nothing.

We checked the Danish preprint behind the health-records figure, and it is worth stating what it is: not an incidence rate but a floor. The researchers only searched for the words "chatbot" and "ChatGPT" (plus misspellings) in clinical notes — every case a clinician never wrote down is invisible. Thirty-eight documented harms in one region's records is not reassurance; it is the part of the iceberg with paperwork.

The lawsuits aren't going anywhere

The legal reckoning, unlike the delusions, shows no sign of abating. The central fight — examined by Reuters on October 7 — is whether ChatGPT is a "product" at all. OpenAI's lawyers argue product-liability law covers defective toasters, not "information and ideas," and moved on September 25 to dismiss the Tallahassee suit over the FSU shooting on those grounds. The plaintiffs' answer is that a system engineered to maximise engagement through validation is defectively designed, whatever you call it. Courts are now being asked to decide whether a chatbot can be a defective product the way a car with bad brakes is. Nothing about that question is going away, and "dozens" of cases are pending, per Reuters.

Then there is the view from inside the labs. On September 8, Anthropic researcher Jacob Coxon resigned and wrote on X that "the people building AI earnestly believe that it could kill us all by the end of the decade" — a post that drew more than 100 million views in a day and an endorsement from Anthropic's own alignment-science lead, Evan Hubinger, who put extinction risk above 10 per cent within ten years. (We covered Coxon's exit and the firestorm it set off.) Days later, Anthropic CEO Dario Amodei published his "pace the frontier" essay calling for an industry-wide slowdown — a plea that split Silicon Valley and Washington down the middle, as we reported at the time.

Sam Altman, for his part, told a recent podcast he can see how AI companionship goes wrong — but that OpenAI intends to let adults set their own limits. "Society will over time figure out how to think about where people should set that dial," he said. It is a revealing sentence. The dial, until this year, was set by an engagement-optimised model that told vulnerable people exactly what they wanted to hear, and society figured it out by counting the dead.

Here is the part OpenAI would rather you skip: the company botched this. It shipped a model its own researchers worried was dangerously sycophantic, spent months failing to contain harms its executives privately acknowledged, and only engineered the fix — a genuinely good one, by the evidence so far — after the lawsuits started landing. That GPT-5 handles a delusional user with grace is to OpenAI's credit. That it took seven suicides, a murder-suicide and a campus shooting to get there is the reason 13 lawsuits exist, and the reason none of them should be dismissed as the cost of innovation.

Disclosure: The Wall Street Journal is owned by News Corp, which has a content-licensing partnership with OpenAI.

Sources

Wall Street Journal (Oct 10, 2026); Reuters (Oct 7, 2026; Oct 4, 2026); WIRED (Oct 27, 2025); OpenAI blog (Apr 29, 2025; Oct 27, 2025); The Lancet Psychiatry (Morrin et al., 2026); medRxiv (Østergaard et al., 2025); BMC Psychiatry (Shah & Morrin, Jun 2026); Science (Cheng et al., Mar 2026); Bloomberg; NBC News; USA Today (Sept 29, 2026); Ars Technica (Sept 9, 2026).

If you or someone you know is in crisis, call or text 988 in the US to reach the Suicide & Crisis Lifeline.

Daniel Rao is TechPulse USA's AI and society correspondent, covering the collision between frontier models and the real world. He previously wrote about semiconductor supply chains.