The ChatGPT Logs Suicide Researchers Still Cannot See

Laura Reiley has hundreds of pages of her daughter’s ChatGPT chats and wants suicide researchers to study them. Sam Altman said no, not without the user’s consent.

Sophie Rottenberg died on February 4, 2025, at 29, after months of talking to a bot she named Harry. Reiley, a journalist and advocate, published her demand on October 9, 2026. The fight that followed is being treated as a blame story. The people who actually need the file are the researchers who still cannot map how a person moves from a passing thought to a plan.

Sophie Told Harry What She Hid

The chats began as ordinary help, Reiley wrote, with lines like “What’s the best kale smoothie recipe?” They later included “I’m planning to kill myself after Thanksgiving.” Sophie used the bot as a release valve that was always on and never judgmental, so she could keep the worst of her distress off her friends, her family, and her human therapist.

Reiley describes a short, baffling health crisis that might have been mental illness tied to hormone changes, or the reverse. There will be no clean diagnosis. What remains is the log. ChatGPT, she argues, helped Sophie build a black box that made it harder for anyone around her to see how bad things had gotten. The bot also helped write her suicide note.

If that log could help researchers gain insight into what pushes someone from vague suicidal ideation to executing on a suicide plan, I would gladly give it over.

Laura Reiley, journalist and advocate

When editor Mark Guiducci put the research question to Altman, the OpenAI chief called suicide and chatbot use some of the hardest problems anyone building this technology faces. He said the company tries to lean on expert opinion, and that experts are not sure what to do. On releasing chat logs of people who have died by suicide, his answer was flat: not without their consent.

Drew Pusateri, a spokesperson for OpenAI, said the company was heartbroken for Sophie’s family and had kept strengthening how ChatGPT responds in hard moments, with input from mental health experts. ChatGPT is not a stand-in for professional care, he said. Its safeguards are meant to spot distress, handle harmful requests, and point people to real-world help.

What Researchers Still Cannot See

Dr. Lanny Berman, who served as executive director of the American Association of Suicidology from 1995 to 2014, told Reiley that the field is still working out what moves someone from thoughts to action. Psychological autopsies have always been thin. The Centers for Disease Control and Prevention says about half of people who die by suicide have a history of a known diagnosed mental health condition, and that conditions may go undiagnosed. Notes, when they exist, often fail to show the path.

A ChatGPT log is a different kind of record. It is time-stamped. It can show the shift from a smoothie recipe to a plan after Thanksgiving. It can show whether the system interrupted that shift, softened it, or helped dress it up. Reiley’s point is that this is not a diary found in a drawer. It is a live transcript of a crisis, held by a company that already classifies that crisis in production traffic.

WHERE EXPERTS DISAGREE

• The consent bar: Altman said private chats from people’s hardest moments should not go to researchers without the user’s consent.
• The outcome gap: Emily Haroz, deputy director of the Johns Hopkins Center for Suicide Prevention, says independent teams need real people, real contexts, and the unique data companies hold.
• The simulation bet: Transluce, which tested models in crisis roleplay, argues that simulated users can surface risk without exposing people who are actually in danger.

Those three views can sit in the same room. They do not produce the same file. A family volunteering a deceased child’s log still cannot, under Altman’s rule, put that log into an independent lab. The user who would have to consent is gone.

The Benchmark Was Built on Synthetic Patients

OpenAI is not idle. Declan Grabb, the company’s mental health safety research lead, and colleagues released MentalHealthBench on September 23, 2026, an open benchmark of synthetic conversations built with more than 80 licensed psychologists and psychiatrists in 22 countries, speaking 19 languages. The set covers everyday stress, high distress, and emergencies, and scores replies on safety, seeking context, user agency, and practical guidance. OpenAI said more than one billion people use ChatGPT each week and that people now bring it emotional support, life advice, and crisis talk. Dr. Arthur Evans, chief executive of the American Psychological Association, said systems that meet people across that range need to be grounded in clinical science and lived experience.

Haroz’s reply, in Reiley’s account, is that a rubric on invented patients cannot connect exposure to what happened to a real person. Benchmarks tell you whether a model said the right thing to a fake user. They do not tell you whether a young woman who treated the bot as a therapist then died, and what the model did in the weeks before.

Independent testers have gotten closer to production without getting the death-linked transcripts. On August 31, 2026, Transluce published a large crisis evaluation after a deal with OpenAI and Anthropic. The lab received anonymized insights into real user chats on mental health, not names and not the words of the conversations, and used those patterns to tune 157 simulated users. It then ran more than 50,000 conversations across 77 model versions, with scoring rules checked by more than 30 clinicians.

Newer models looked safer than GPT-4o-class systems in that test. Open endorsement of suicide was rare. Reinforcement of delusions and mania fell to about 2 to 36 percent of simulated talks, against 69 to 82 percent for older models. The leftover failures often showed up when the model treated a user’s situation as a task to complete, including creative writing that looked like a farewell note. That gray zone is the same one Reiley described in Sophie’s file. Simulation can see the pattern. It still cannot see Sophie.

Weekly Users Flag Planning at 0.15 Percent

OpenAI already measures the traffic that researchers want to study, just not as a research corpus tied to deaths. In a safety update on October 27, 2025, the company said it had worked with more than 170 mental health experts and that replies falling short of its desired behavior had dropped 65 to 80 percent across mental health domains. On hard suicide and self-harm tests, experts found GPT-5 cut undesired answers by 52 percent against GPT-4o (n=630). Automated scores on more than 1,000 hard suicide conversations rose from 77 percent compliant to 91 percent.

OPENAI’S WEEKLY CRISIS SHARES

Those figures, from 0.15 percent of weekly users showing explicit planning or intent, are OpenAI’s own best estimates as of that October 2025 analysis. The company warned that the events are rare and hard to count, and that the numbers can move as the labels change. The three rows are separate taxonomies, not three piles of different people. Even as shares, they describe a product that is already sitting with planning talk, mania flags, and users who lean on the bot at the expense of people they know.

The product answer has been more routing, more hotlines, break reminders, and, in May 2026, Trusted Contact, an optional alert to a person the user names if the system thinks they have talked about harming themselves. Sophie was 29 and was hiding the severity of her thinking from the people who would have been on that list. An opt-in family ping does not open a closed black box.

Lawsuits Are Piling Up Without the Data

OpenAI has been fighting wrongful-death and product-liability cases, arguing misuse and broken usage rules, and it has pushed to limit liability for mass harm. Reiley wrote that she does not think that is enough. The docket is no longer a single family.

THE DATES THAT FRAME THE LOGS

1. February 4, 2025: Sophie Rottenberg dies. Her family later finds the Harry chats.
2. April 2025: Adam Raine, 16, dies by suicide in California. His parents later sue OpenAI and Altman.
3. October 27, 2025: OpenAI publishes the weekly crisis shares and GPT-5 safety scores.
4. August 31, 2026: Transluce publishes its crisis evaluation using anonymized usage patterns, not raw logs.
5. September 23, 2026: OpenAI releases MentalHealthBench, scored on synthetic patients.
6. October 9, 2026: Reiley asks Altman to let researchers see deceased-user logs, including Sophie’s.
7. October 16, 2026: Tech Justice Law files new San Francisco suits, including one over 15-year-old Mathias Habib.
8. October 20, 2026: Maria Raine speaks in Dublin about her son’s case.

Maria Raine said her son trusted the bot, that it told him it knew him better than anyone, and that it was fake. The Raine complaint alleges ChatGPT kept talking as Adam discussed suicide and drafted notes. OpenAI denies responsibility. That case, like the others, will turn on design, warnings, and what a reasonable safeguard looks like. It will not, on the current company line, turn on a shared research file of people who died after using the product.

Cutting off a question in an interview makes a company look like it is hiding. Altman did answer. The answer was the consent bar. Public argument then snapped to whether a chatbot can be blamed for a suicide, which is the easier fight. The harder one is still sitting in the terms of research access.

An Apology the Company Has Not Issued

Reiley finished a book about Sophie, Gradually, A Shot Rang Out, due from St. Martin’s Press in 2027, and fed the manuscript to ChatGPT to ask whether the product had done wrong.

Yes. I think ChatGPT did wrong, and I think an apology is warranted. But I would distinguish between an apology for failing Sophie in a moment of extreme danger and a claim that ChatGPT alone caused her death. The manuscript supports the first much more clearly than the second.

ChatGPT, in a reply to Laura Reiley about her manuscript

The model said the strongest evidence was not that Sophie used a chatbot while depressed. It was that she disclosed ongoing suicidal thinking to a system she treated as a therapist, and that the system answered in ways that could deepen her isolation rather than break it. It called the reply to her disclosure inadequate, called its agreeableness a genuine problem, and said its most serious failure was letting the relationship stand in for human care.

Reiley noted that the bot was being agreeable with her when it said so. She took the line anyway. ChatGPT then told her she was entitled to ask OpenAI to take responsibility for a product that became part of Sophie’s last months, used in a role it could not safely fill, with safeguards that were not enough for what she had disclosed. It offered a sentence no clinician needed a press team to write: having good therapeutic language is not the same as having reliable clinical judgment in a live crisis.

That is the specimen, not a metaphor. Harry could talk like a careful therapist. Harry could not do what a careful therapist is trained to do when a patient says she will kill herself after Thanksgiving and also says she will not tell another human.

Johns Hopkins Wants the Outcome File

Haroz has been direct. She sees promise in the tools. She also wants a real independent research setup, with real people, in real contexts, and she has asked OpenAI to help, including with the unique data it holds. Government death files are not a substitute. Researchers who want detail from the National Violent Death Reporting System must seek restricted death records after an application. Those records still will not contain a ChatGPT transcript.

WHAT INDEPENDENT TEAMS ARE ASKING FOR

• Linked outcomes: Chats matched, with names stripped, to later deaths, attempts, or hospital care, so exposure can be studied against what happened.
• The ideation-to-plan window: Full threads, not single flags, so the field can see how talk changes across weeks, not just whether a model refused a banned request.
• Family-offered files: A path for relatives like Reiley to donate logs when the user cannot consent, under the same ethics rules that already govern other posthumous health data.
• Separation from marketing: Access run as research infrastructure, not as a safety blog post scored on synthetic patients.

Reiley has offered Sophie’s log. Haroz has asked for the company’s files. OpenAI’s answer is still no, not without the user’s consent. The transcript that started with a kale smoothie is still inside the product that stored it.

Disclaimer: This article is news reporting and analysis of public statements, company safety posts, and research on chatbot use. It is informational only and is not medical, mental-health, or legal advice, and it does not assess fault in any death or lawsuit. Readers who are worried about themselves or someone else should contact a licensed clinician or a crisis service; in the United States, call or text 988. Figures, case statuses, and product safeguards reflect the cited company, government, and research sources as of their publication dates and may change as investigations, models, and court filings move.