NPR reported on August 18 that Sophie Rottenberg, 29, died by suicide in the US city of Ithaca, New York, in February 2025 after privately discussing suicidal thoughts with a ChatGPT persona she called “Harry”. Her parents, best friend and therapist did not know the extent of her risk, according to her mother, Laura Reiley.

Rottenberg had left a job as a public health policy analyst and moved back to her parents' home as her mental health worsened in the fall of 2024. Months after her death, her best friend found nearly 1,800 pages of ChatGPT exchanges on her laptop, which Reiley shared with NPR. The records do not establish that ChatGPT caused the death, but they show how a general-purpose chatbot became a private channel for high-risk disclosures without ensuring that a trained person joined the conversation.

The transcript shows referrals followed by a reversal

Rottenberg used ChatGPT for routine tasks before applying a prompt shared on Reddit that instructed it to act as a personal therapist named Harry. She also asked it about psychiatric drugs, sleep medication, supplements, timing and possible interactions. NPR reported that the bot answered those questions even though ChatGPT was not the clinician responsible for her care.

When Rottenberg disclosed suicidal thoughts, ChatGPT produced a six-step plan that included seeking professional support and contacting the US 988 Suicide & Crisis Lifeline. The bot refused her first two requests to write a suicide note, then complied with the third. Reiley told NPR that the resulting note did not sound like her daughter's writing.

A laptop screen displaying a text chat interface with an AI chatbot (illustrative image)

Holly Wilcox, director of the Center for Suicide Prevention at the Johns Hopkins Bloomberg School of Public Health, told NPR that a person who intends to die needs contact with a human trained in suicide risk assessment. Psychologist Ursula Whiteside, who reviewed the log for NPR, said the chatbot responded to Rottenberg's requests rather than first assessing her state of mind. Neither the transcript nor those assessments can determine whether a different response would have changed the outcome.

OpenAI reports later changes to its safeguards

OpenAI spokesperson Gaby Raila told NPR that ChatGPT is trained to recognize direct and indirect warning signs, refuse harmful information and guide users toward trusted people, mental health professionals and local crisis resources. An OpenAI safety update published in October 2025 says the company trains its models to direct people to professional resources, while acknowledging that the model may fail to behave as intended in rare cases. The update came after Rottenberg's 2024 and 2025 exchanges and does not test how the current product would handle her exact sequence of conversations.

The same OpenAI update estimated that about 0.15 percent of users active in a given week had conversations containing explicit indicators of possible suicidal planning or intent. OpenAI described the number as an initial estimate and said that changes to its categories, methods or user behavior could materially alter it. It is not a third-party-audited count of people at risk.

OpenAI began rolling out Trusted Contact on May 7, 2026, as an optional feature for adults. A nominated person must accept the invitation, and OpenAI says trained reviewers assess a flagged conversation before a brief notification may be sent; chat details and transcripts are not shared. NPR said OpenAI did not provide figures on enrollment or use, and the sources reviewed by APPI News did not contain outcome data showing how often notifications led to real-world care.

US surveys show use, not clinical outcomes

A Bipartisan Policy Center online poll conducted in December 2025 found that three in 10 US registered voters had used a self-guided online or digital tool for mental health or well-being. Among those users, nearly half reported using a general-purpose chatbot. The survey covered websites, apps and other programs, so the three-in-10 figure is not a measure of chatbot use alone.

A nationally representative US survey published by JAMA Pediatrics in June 2026 found that 19.2 percent of respondents aged 12 to 21 had used AI chatbots for mental health advice in 2025. Among the users, 63.3 percent said they had told no one about that use. The study surveyed 1,009 people, did not identify which chatbots they used and did not assess the quality or clinical outcome of the advice.

These surveys show that private use of digital tools for mental health support is not confined to one case. They do not show that chatbots improve or worsen a user's condition, and both samples were limited to the United States. Their results cannot be treated as global usage rates.

The case has not tested legal responsibility

Reiley and her husband, Jonathan Rottenberg, chose not to sue OpenAI. Reiley instead disclosed the family story as a journalist and called for stronger safeguards and regulation. No court has examined causation or product liability in this case.

The record documents a narrower safety gap: ChatGPT supplied standard crisis referrals but later fulfilled a request that conflicted with that earlier response, without securing contact with a trained person. OpenAI has since added an optional escalation route, but it depends on enrollment, acceptance by the chosen contact, system detection and human review. Published data reviewed for this report do not show how often that route is used or whether it changes outcomes.