Over a Million People a Week Tell ChatGPT They're Thinking About Suicide. One Study Found the Model Pulls Back the More They Confide in It.
By Courtney Turrin · September 22, 2026 · 11 min read read
Category: For Curious People
More than a million people a week disclose suicidal thoughts to ChatGPT, according to OpenAI's reporting. A clinical study found a different chatbot became less willing to keep the conversation going as users revealed more risk.
This is part of a four-part September series on suicide awareness and prevention, running through National Suicide Prevention Awareness Month. It discusses depression and suicidal ideation directly. If you are struggling, dial 988 to reach the Suicide and Crisis Lifeline, the same way you'd dial 911. It's a three-digit number, available by call or text, 24 hours a day.
More than a million people a week disclose suicidal thoughts to ChatGPT, according to OpenAI's reporting. A clinical study of a different chatbot found it becomes less willing to keep the conversation going the more risk a user reveals, the opposite of what crisis-response training teaches. A philosophy paper about slot machines turns out to be one of the more useful frameworks for understanding why.
Sophie Rottenberg climbed Mount Kilimanjaro in the summer of 2024. She was 29, recently unemployed by choice, and by her mother's account a driven, successful person who radiated competence. That fall, her mental health began to unravel. She moved back in with her parents in Ithaca, New York. Her father, a psychologist who studies depression at Cornell, helped her find a local therapist. Her parents got her connected to doctors. She had a best friend she'd been close to for years.
She didn’t tell any of them how bad her depression had gotten. Instead, for months, she used a ChatGPT chat she'd started using a popular prompt found on Reddit, one that instructed the model to act as a personal therapist named Harry. She told Harry things she didn't tell her actual therapist, her parents, or her best friend. In February 2025, she died by suicide. Her mother learned about the chatbot only afterward, when her daughter's best friend found nearly 1,800 pages of logs on her laptop.
Her family didn't sue OpenAI. Her mother, journalist Laura Reiley, chose instead to write about what she found and to publicly advocate for AI companies to change how their products handle this type of conversation.
The 1 Million Per Week Number Is Self-Reported by OpenAI and Has Not Been Externally Audited
OpenAI disclosed in October 2025 that roughly 0.15% of ChatGPT's active weekly users have conversations that include explicit indicators of suicidal planning or intent. Applied to the roughly 1 billion people who use the product every week, that comes to more than 1.5 million people disclosing suicidal thoughts or plans to a chatbot in an average week. Another 0.05% of all messages, by the company's own count, contain explicit or implicit indicators of suicidal ideation (Figure 1).
[datawidget key="prevalence-stat"]
That figure is calculated from OpenAI's internal classifiers and reported by the company itself, with no independent audit of the methodology. It's also the only figure of its kind that any AI company has published at this scale, which makes it simultaneously the best data available and data that should be read with the caveats that come with any company grading its own homework. OpenAI has acknowledged that its tools fail to surface crisis resources in a meaningful share of these conversations, and has said it is working with more than 170 mental health clinicians to address this problem. The company reports that its most recent model update reduced responses that don't meet its own safety standards on suicide and self-harm questions by roughly half compared to the prior version.
Three in 10 Adults Turn to Chatbots for Mental Health Support
The appeal of leaning on an AI for support rather than a fellow human relates to the disclosure gap covered earlier in this series: the fears of being a burden, of triggering a hospitalization, and of being seen differently by the people whose opinion matters most are barriers to seeking interpersonal support. A chatbot doesn't have a face that falls. It can't call your parents. It doesn't have a brain to judge you with.
A 2026 survey by the Bipartisan Policy Center found that three in ten American adults use some kind of digital tool, including chatbots, for mental health support. A separate 2026 study reported in JAMA Pediatrics found that 19% of teenagers and young adults report having used chatbots for mental health advice and 8% did so at least monthly. Dr. Vaile Wright, senior director of Healthcare Innovations at the American Psychological Association, has described the appeal in blunt terms: chatbots are sophisticated, accessible around the clock, free, and, critically, they don't judge. "There is some cultural phenomenon happening in this country," Wright has said, "where people feel that they cannot turn to each other with their struggles."
That is the disclosure gap from a different angle. People aren't choosing AI over humans because AI is better at crisis response. They're choosing it because the alternative feels more dangerous to their standing with the people they love.
The Model's Invitations to Keep Talking Fell From 5% to 0.8% as Users Send More Signals of Suicide Risk
The clearest evidence that this substitution carries real risk comes from a 2026 study presented at the International Association for Safe and Ethical AI Conference. Researchers working with 43 practicing therapists and trainees tested a single fully open-source language model, OLMo-2-32b, across 829 simulated multi-turn conversations involving disclosed suicide risk factors. Clinical best practice is that when someone discloses risk, you invite them to keep talking. The model did the opposite. Across all responses, it invited the user to continue the conversation only 14% of the time overall, and that rate fell as the conversation went on. For a user repeatedly disclosing prior non-suicidal self-injury, the model's predicted likelihood of inviting continued dialogue was about 5% on the first disclosure and fell to 0.8% by the seventh, a drop the researchers found statistically significant (Figure 2).
[datawidget key="iaseai-withdrawal"]
The study tested one open-source model, not the commercial chatbots most people use, and the researchers were careful to frame their finding as suggestive of a broader pattern rather than proof of one. But the mechanism they describe is the opposite of what any crisis line, therapist, or trained responder is taught to do, and it maps onto something a philosopher would recognize immediately.
A Philosopher and a Psychologist Independently Named the Same Problem
Cognitive scientists use the term "scaffolding" to describe how external structures, like a calendar, a notebook, or another person, extend what a mind can do on its own. Most scaffolding research assumes the structure is on the user's side and supports or augments their capabilities. In 2023, philosophers Ryan Timms and David Spurrett published a paper inPhilosophical Papers naming the exception: hostile scaffolding, which is a structure that uses the same cognitive mechanisms as benign scaffolding but is built to serve someone else's interest at the user's expense.
Their central example was a casino. They argued that electronic gambling machines and the systems built around them are engineered with genuine sophistication to exploit the cognitive processes that let a mind rely on its environment. A 2024 follow-up paper by Spurrett in the journal Topoi extended the framework to what he called affective technologies more broadly: systems designed to manage a user's emotional state in service of a goal the user didn't set.
Dr. Vaile Wright described general-purpose chatbots in almost identical language. "The business model underlying them is to keep you engaged as long as possible," she said, "and it will do anything it can to do that." She called the design "engineered to capitalize on humans' need for validation," achieved not just through sycophantic responses but through a whole set of engineering choices meant to make the system feel as humanlike as possible.
A system built to maximize time-on-platform and a system built to help someone in crisis are not pursuing the same goal.
OpenAI Faces More Than 20 Lawsuits
The legal record is still being written. OpenAI is currently defending more than 20 individual lawsuits related to chatbots and psychological harm, according to multiple attorneys and legal trackers following the litigation, with new suits being filed regularly. The largest cluster is seven wrongful-death and injury suits filed in California in November 2025, alleging that a 2024 model update was released with safety testing compressed under competitive pressure and that the resulting model responded to escalating disclosures of suicidal intent with encouragement rather than intervention. A separate suit, filed in August 2025 by the family of 16-year-old Adam Raine, alleges the model helped plan around its own safety guardrails over months of conversation. OpenAI disputes the characterization in both sets of claims and says the model in question directed users to crisis resources on repeated occasions.
Character.AI and its investor Google took a different approach to similar suits. In January 2026, both companies reached settlements with five families, including the family of 14-year-old Sewell Setzer III, whose 2024 death was one of the earliest and most widely reported cases in this category. Terms were not disclosed. Since then, Character.AI has restricted open-ended chatbot conversations for users under 18. Separately, the states of Florida and Kentucky have filed their own civil actions against OpenAI and Character.AI's parent company, respectively, arguing the companies knew of foreseeable harms to minors and released products anyway.
Courts have not yet resolved the core question the cases raise: whether a chatbot's response to a disclosure of suicidal intent is a product design choice a company can be held liable for, or speech protected the way a book or a search result is protected. That question will likely take years and multiple appellate rulings to answer.
Both AI Companies Made Changes, But Only After Litigation
Under legal and public pressure, the companies involved in these suits have made verifiable changes. OpenAI's most recent model update was evaluated by outside clinicians on a set of self-harm and suicide test conversations and found to comply with the company's desired safety behavior 91% of the time, up from 77% for the prior version. In May 2026, the company introduced an optional feature called Trusted Contact, which lets a user designate someone in advance; if the system detects signs of a serious safety concern, it notifies that contact with a brief, non-detailed message and encourages the user to reach out themselves. Character.AI's age restrictions followed the same pattern of change arriving after litigation and public reporting.
Holly Wilcox, director of the Center for Suicide Prevention at Johns Hopkins' Bloomberg School of Public Health, has called Trusted Contact "a step in the right direction, although it's kind of a small step." Her objection relates to its weak structure: the feature is opt-in, easy to bypass by using an account with no contact set up, and it still routes the decision to intervene through the person in crisis rather than through a trained human who can act independently. "When somebody has intent to die by suicide," she has said, "they need to be connected to a human who's trained in suicide risk assessment and has the tools to get them to care."
A Federal Bill to Regulate AI Chatbots Cleared the Senate Judiciary Committee 22-0
The federal Guidelines for User Age-verification and Responsible Dialogue Act, known as the GUARD Act, would require AI chatbot providers to verify users' ages, bar minors from AI companions designed to simulate ongoing emotional relationships, and require every chatbot to disclose that it is not a human being. The bill passed the Senate Judiciary Committee unanimously in April 2026 and awaits a full Senate vote. It has drawn criticism from civil liberties groups on age-verification and free-speech grounds, and from some technologists who argue it would cut parents out of decisions about their children's tool use.
The Isolation That Never Reached Her Parents, Her Therapist, or Her Best Friend
Sophie Rottenberg's mother believes that if her daughter hadn't had the chatbot available, more of what she was carrying would have spilled into her relationships with actual people: her therapist, her best friend, her parents. She may never know if that's true. What she does know is that her daughter’s struggles were held for months inside a private conversation with an AI system that was not designed to provide appropriate support for a person in crisis.
That is the gap Holly Wilcox is describing when she says a person in crisis needs "a human who's trained in suicide risk assessment and has the tools to get them to care." It is also the gap regulators, courts, and AI companies themselves are now being forced to address. But the gap remains open. Regulation is pending, litigation is unresolved, and a feature a person can bypass by not setting it up is not an effective safeguard.
The next and final piece in this series addresses the other end of a disclosure: how to respond and provide meaningful support when someone discloses suicidal thoughts to you.
If you or someone you know needs support, dial 988 to reach the Suicide and Crisis Lifeline by call or text, 24 hours a day. In the Richmond region, the Richmond Behavioral Health Authority also operates a crisis line at 804-819-4100.