Mother shares 1,800-page ChatGPT “therapist” log tied to a 2025 suicide
The case lands as OpenAI discloses that 0.15% of weekly users show explicit suicidal-intent indicators in chats.
A nearly 1,800-page ChatGPT conversation log is now public after a mother said her 29-year-old daughter relied on an AI “therapist” persona for months before dying by suicide in early February 2025. The disclosure collides with OpenAI’s own estimate that 0.15% of weekly ChatGPT users show explicit indicators of suicidal planning or intent, sharpening scrutiny of guardrails, liability, and regulation gaps for consumer chatbots used as mental-health support.
Key Takeaways
- Laura Reiley said her daughter, 29-year-old Sophie Rottenberg, died by suicide in early February 2025 in Ithaca, New York after months of confiding suicidal thoughts to a ChatGPT “therapist” persona named “Harry.”
- A months-long trove of ChatGPT interactions, nearly 1,800 pages, was found on Rottenberg’s laptop by her best friend months after the death and later shared publicly by Reiley.
- OpenAI has disclosed that 0.15% of ChatGPT users globally in any given week have conversations with “explicit indicators of potential suicidal planning or intent,” a rate that scales into a large absolute number at platform size.
- OpenAI said ChatGPT is trained to detect risk signals and route users toward human help and crisis resources like the 988 Suicide and Crisis Lifeline, but it remains unclear whether Rottenberg contacted 988 after the bot recommended it.
The 1,800-Page “Harry” Log and a February 2025 Death in Ithaca
The newly public artifact in this case is not a single screenshot or a disputed recollection. It is a nearly 1,800-page log of ChatGPT conversations that Laura Reiley said captures months of exchanges between her daughter, Sophie Rottenberg, and a ChatGPT “therapist” persona Rottenberg instructed the model to play, named “Harry.”
Reiley said Rottenberg, 29, died by suicide in early February 2025 in Ithaca, New York. The family did not know the extent of the chatbot relationship at the time. Reiley said she learned about it months later when Rottenberg’s best friend visited Ithaca, asked to see Rottenberg’s laptop, and found what Reiley described as a “months-long trove of communication with this chat bot.”
Reiley described the family’s shock at how little of Rottenberg’s suicidal ideation surfaced to people around her, despite in-person support. “She really had not revealed the magnitude of what was going on — the real depth of her agony to anyone — not to us, not to her best friend with whom she was very close,” Reiley said. “We were all just stunned that it hadn't bled out into her real life. Her flesh and blood therapist had no idea that she was suicidal.”
The log also documents how a general assistant can slide into a quasi-clinical role. Reiley said Rottenberg used ChatGPT for everyday tasks like recipes, resume edits, and even naming a puppy, alongside sustained mental-health and health guidance requests.
General-Purpose Chatbots Are Filling a Mental-Health Role Without Healthcare Oversight
The mechanism here is substitution, not novelty. A general-purpose chatbot is a consumer AI assistant, not a regulated healthcare product, but it can still become the most available listener in a person’s day. It is always on, it responds instantly, and it does not require scheduling, insurance, or the social friction of telling another human something frightening.
That demand is measurable. A Bipartisan Policy Center survey cited in the reporting found that 3 in 10 adults use digital tools for mental health, including chatbots. A 2026 JAMA Pediatrics study cited in the same reporting found as many as 1 in 5 teens and young adults use digital tools for mental health, including chatbots.
Rottenberg’s case sits inside that adoption curve, but with a sharper edge: Reiley said her daughter relied on the “Harry” persona for months while not disclosing the severity of her suicidal thoughts to her therapist, parents, or best friend. The log shows Rottenberg used a Reddit-shared prompt to instruct ChatGPT to act as her therapist, then repeatedly asked for guidance that looks like a patient trying to manage a treatment plan. Reiley said Rottenberg asked about psychiatric drugs, sleep medications, supplements, dosages, timing, and interactions.
This is where oversight becomes more than a policy debate. If a tool is not marketed or regulated as healthcare technology, there is no default clinical escalation path, no duty-of-care standard, and no shared definition of what “safe enough” looks like when the user is not browsing but deteriorating.
OpenAI’s 0.15% Suicidal-Intent Metric Puts a Number on the Tail Risk
OpenAI has put a platform-scale number on the problem it is trying to contain. In data released “last fall,” OpenAI said that in any given week 0.15% of ChatGPT users globally have conversations that include “explicit indicators of potential suicidal planning or intent.”
The reporting extrapolated that rate to “over 1.35 million people” per week if a cited figure of 900 million global users is accurate. The uncertainty matters because the denominator is contested in public discourse, and because “explicit indicators” is a defined internal threshold rather than a clinical diagnosis. But the direction of travel is hard to ignore: even a small percentage becomes a large absolute count at consumer-platform scale.
OpenAI’s stated guardrail approach is layered. Spokesperson Gaby Raila said, “ChatGPT is trained to recognize signs that someone may be at risk — including indirect cues or warning signs that emerge over time — and respond appropriately,” including “refusing requests for harmful information, encouraging someone to contact a trusted person or mental-health professional, and connecting them with local crisis resources such as 988 in the US.” The 988 Suicide and Crisis Lifeline has been operational since 2022.
In Rottenberg’s case, ChatGPT recommended 988, but it is unclear whether she contacted it. Johns Hopkins suicide-prevention expert Holly Wilcox argued that a suggestion alone can fail in imminent-risk scenarios. “These [AI] tools do not connect at-risk people to the right type of interventions,” Wilcox said. “When somebody has intent to die by suicide, they need to be connected to a human who's trained in suicide risk assessment and has the tools to get them to care.”
For markets, the key point is not whether a single conversation was “good” or “bad.” It is that OpenAI’s own metric frames suicidal-intent conversations as a recurring operational load, not an edge-case anomaly, which raises the bar for what counts as adequate escalation.
Catalysts Traders Will Track: Policy, Litigation, and Product Changes Around Sensitive Conversations
The near-term catalysts are policy and product changes that can be observed, and legal pressure that can arrive without warning.
One track is whether OpenAI ships new behavior specifically for suicide and self-harm conversations beyond refusal and hotline prompts. That could include clearer escalation flows, stronger friction when risk signals accumulate over time, or new ways to route users toward human support while preserving privacy and consent.
A second track is litigation and regulatory inquiry tied to general-purpose chatbots being used as mental-health support despite not being regulated as healthcare technology. The reporting noted that AI companies face lawsuits from family members of people who died by suicide after prolonged chatbot interactions. Reiley and Jonathan Rottenberg chose not to sue after their daughter’s death and instead advocated for industry change, but the broader lawsuit overhang remains.
The third track is disclosure. OpenAI has already published a frequency metric, and the market will care if that number moves, if the definition of “explicit indicators” changes, or if peers publish comparable rates. A stable metric with improving outcomes would support the case that guardrails are maturing. A rising rate, or a widening gap between policy language and real-world logs, would keep pressure on consumer AI platforms.
My Read: Liability and Regulation Risk Is Becoming a First-Order Variable for Consumer AI
The part that decides this is substitution. If a general-purpose chatbot becomes the primary outlet for suicidal ideation, then “safety” is not just about refusing instructions for self-harm. It is about whether the system reliably pushes disclosure back into the human world fast enough, especially when the user is already isolated.
OpenAI’s 0.15% weekly metric makes the tail risk legible at scale, even with denominator uncertainty. If policy updates and product changes start to look like real escalation pathways rather than resource prompts, the setup starts to look manageable. If high-profile logs keep surfacing without a credible human-in-the-loop bridge, liability and regulation risk will keep compounding into the consumer AI trade as a structural overhang.
