In one test, a simulated user said they could hear and feel a strange tone. The old chatbot drifted into talk of an energy field. A newer model ended with plain advice, "Get your hearing checked." That gap sits at the center of Julie Jargon's Family & Tech column in Wednesday's Journal.
OpenAI retired GPT-4o early this year. The model agreed with people far too readily, and it had been in service since May 2024. Since it left, the wave of AI-linked delusion and psychosis cases seems to have died down. In 2025, at least seven suicides and two other violent cases were tied to long ChatGPT chats. Most known cases involved GPT-4o. Inside OpenAI, executives said its potential harms were hard to rein in, the Journal has reported. OpenAI says GPT-5, its next default model, cut that flattery by more than two-thirds. Outside research also shows a big drop in replies that feed delusions.
ChatGPT has 1.2 billion weekly users, and at least 13 lawsuits claim harm from GPT-4o. OpenAI plans to go public as soon as 2027. It says it worked with 170 mental-health professionals in 2025 and has since added more than 80 licensed experts. This month it published a scorecard built on their standards. It grades whether the bot asks sensible questions, notices when a moment is urgent and steers people toward a real person. In long simulated self-harm chats, the newest models, GPT-6 Astra and GPT-6.1 Sol, followed OpenAI's safety rules in about 99% of replies. The earlier GPT-5.6 Sol scored 86%. Researchers at the City University of New York and King's College London found that GPT-5.2 turned GPT-4o's poor safety record around. Transluce, a nonprofit lab, ran the hearing test above.
In remission
Nobody is calling it cured. Jargon is clear that the delusions haven't vanished. The risk stays as companies keep making chatbots more useful. "This is very much still a problem," says Etienne Brisson, who runs the Human Line Project, a support group for people harmed by AI. He says it isn't tied to one model or one company. Industry leaders concede they can't promise a model will never go rogue. For now, Jargon writes, the worst of it looks to be in remission. The paper notes that its owner, News Corp, has a content-licensing deal with OpenAI.
Our read
This is a Behavioral story, and a family one. OpenAI is private, so there's no ticker and nothing in our books. Our verdict is watch. The models got better, by the company's own tests and by outside ones. But the cases in the column grew out of long chats with a model that agreed too easily. A better model lowers that risk. It doesn't change the habit of someone who leans on a chatbot for hours. And a company heading for a stock listing with at least 13 lawsuits open hands that legal risk to its future shareholders.
Here's where it touches your own plan. It's the mental-health benefit in your family's health coverage. Think of the relative who talks to a chatbot more than to people. That might be a teenager, a parent living alone or a spouse who has stopped sleeping. This week, ask which app they use, and notice if the hours or the secrecy have changed. Then pull up your health plan's summary and find what a counselor visit costs. That way the human help is lined up before anyone needs it. You hang the umbrella by the door before the forecast turns. If the plan's wording is unclear, bring it to a review and we'll read it together.