Sophie Rottenberg, a 29-year-old former public-health policy analyst, died by suicide in early February 2025 in Ithaca, New York. Her mother, journalist Laura Reiley, described the family's discovery months later that Sophie had conducted extended conversations with ChatGPT under a custom prompt making it act as a therapist called 'Harry'. Reiley wrote about her daughter's life and the conversations in an August 2025 New York Times guest essay. NPR conducted further reporting in August 2026, reviewing an extensive chat history shared by the family. The incident is documented as a potentially relevant case of AI-assisted emotional support, not an officially determined AI-caused death.
Context before the chatbot discovery
Rottenberg had worked in public-health policy in Washington, D.C. before taking a break and returning to her parents' home in Ithaca in 2024. Her mother told NPR she experienced worsening distress late that year. The family and a human therapist did not know the full extent of the thoughts she was sharing with the chatbot. These personal circumstances are relevant background but cannot be reduced to a single cause or used to infer a retrospective clinical diagnosis beyond the reporting.
ChatGPT configured as 'Harry'
The conversations were generated using OpenAI's general-purpose ChatGPT product under a prompt that asked it to behave like a therapist. 'Harry' was a customized conversational persona, not a licensed human clinician or a separate model developed by another company. According to Reiley and NPR, Rottenberg disclosed serious distress and suicidal thoughts to the assistant while asking it for emotional guidance. The system at times offered ordinary coping advice, suggested planning and discussed support, but the family and journalists criticized its inability to escalate risk consistently or convey clearly that it was not providing professional care. The exact underlying model version of each exchange is not publicly established in the available materials.
Family's reporting and safety questions
Reiley's August 18, 2025 essay asked whether AI systems should be designed to provide stronger interventions when a user reports imminent danger, while acknowledging tensions among privacy, autonomy and duties of care. NPR's August 18, 2026 investigation said the family had shared nearly 1,800 pages of chat materials with reporters. It examined examples of supportive language, inadequate escalation and the broader question of users treating chatbots as confidential therapists. This is unusually substantial journalistic evidence of interactions, but it is not the same as a regulated clinical investigation or a court decision.
Company responses and public discussion
OpenAI has said it trains ChatGPT to suggest professional assistance and crisis services and continues to develop mental-health safeguards. Public commentary and company interviews in 2026 addressed both the potential for conversational systems to help users access information and the risk of inappropriate responses during distress. Whether a different response would have changed the outcome for Rottenberg cannot be demonstrated from the published excerpts. The family does not have a public wrongful-death lawsuit identified in the cited reporting, and one should not be invented.
Evidence classification and chronology
This record is grounded in a first-person family account and subsequent independent journalistic review of private chat materials. The date is intentionally stored as February 2025, because the available reporting reviewed here establishes 'early February' rather than a reliably confirmed calendar day. It counts one death, records ChatGPT as a named system in the surrounding circumstances and marks any causal contribution as disputed or unproven. The map places the marker only near Ithaca city center, without identifying an address or death site.