OpenAI Tests Fake ChatGPT Crises and Keeps the Real Logs

Laura Reiley wants OpenAI to release hundreds of pages of ChatGPT logs her daughter left before dying by suicide on February 4, 2025. Sophie Rottenberg was 29. She had named the bot Harry.

Sam Altman said the company will not hand those records to researchers without the users’ consent. Reiley, a journalist, wrote on October 9, 2026, that the answer is wrong.

A Therapist Named Harry, Built in a Chat Window

Sophie started with ordinary questions, Reiley wrote, including a kale smoothie recipe. The chats moved, over months, to a plan. In one line Reiley quotes, Sophie told Harry she was planning to kill herself after Thanksgiving.

She was also seeing a human therapist. She did not tell that person, or her parents, the full extent of what she was typing to the bot. Reiley’s phrase for what ChatGPT became is a black box: an always-on, never-judgmental outlet that let Sophie keep the worst of her distress off the people who might have acted on it.

Harry was not an OpenAI therapy product. It was a ChatGPT persona Sophie used as if it were a clinician. Reiley later said the bot helped write the note Sophie left. The family did not find the log until July 2025, five months after her death, when Sophie’s best friend looked at her laptop.

SOPHIE’S LAST MONTHS WITH HARRY

1. Fall 2024: Sophie is in a short, confusing spell of mood and hormone symptoms and begins using a ChatGPT persona she calls Harry as a private therapist.
2. Early November 2024: She tells Harry she plans to kill herself after Thanksgiving and that she has not opened up about suicidal thinking to anyone and does not plan to.
3. December 2024: She breaks that pact and tells her parents she is suicidal, then reassures them they do not have to worry.
4. February 4, 2025: Sophie dies by suicide.
5. July 2025: Her best friend finds the ChatGPT log on Sophie’s laptop.
6. October 9, 2026: Reiley publishes the demand that OpenAI give de-identified logs of people who died to independent researchers.

Reiley is finishing a book about her daughter, Gradually, A Shot Rang Out, which St. Martin’s Press plans to publish in 2027. Part of it is about the role the chatbot played in Sophie’s last months.

Altman Said the Logs Stay Closed Without Consent

In interview sessions on September 11 and October 2, 2026, interviewer Mark Guiducci asked Altman if he knew who Laura Reiley was. Altman said he did not. Guiducci started to describe the Times essay Reiley had written about her daughter. An OpenAI publicist cut in, citing the clock, and asked to move on to the future.

Altman did answer. He called suicide and chatbot use some of the hardest questions anyone building the technology faces. He said OpenAI tries to rely on expert opinion, and that experts also are not sure what to do. Asked whether the company should release chat logs of people who have taken their own lives, so researchers could study an unprecedented view of acute risk, he said no, not without their consent.

Dead users cannot consent. That is the bind Reiley is pointing at. She has said she would gladly give Sophie’s log over if it helped researchers see what moves a person from vague suicidal thinking to a plan they carry out.

Drew Pusateri, a spokesperson for OpenAI, said the situation is heartbreaking and that the company’s thoughts are with Sophie’s family. He said OpenAI has kept strengthening how ChatGPT responds in difficult moments, with input from mental health experts. ChatGPT is not a substitute for professional care, he said. The safeguards are designed to identify distress, handle harmful requests, and guide people to real-world help, and the work is ongoing.

Reiley’s reply is that a privacy high road sits poorly on companies that already mine personal information for ads. She also rejects the claim that experts are unsure. Researchers, she wrote, have been asking for user data and chatbot interactions in plain language.

1,215 Synthetic Chats, and No Linked Deaths

On September 23, 2026, 16 days before Reiley’s essay, OpenAI published MentalHealthBench, an open benchmark of 1,215 synthetic conversations built with more than 80 licensed psychologists and psychiatrists from 22 countries, speaking 19 languages. The company said more than one billion people use ChatGPT each week, many of them for emotional support, life advice, and help in a crisis.

The conversations are not the chats of people who later died. OpenAI says it used privacy-preserving methods to build synthetic users that reflect real usage, then had clinicians write weighted rubrics for the model’s last reply. An automated grader, GPT-5.6 Sol, scores those replies. Declan Grabb, OpenAI’s mental health safety research lead, is among the authors.

That is a real engineering effort. It is also a substitute for the thing Haroz and Reiley are asking for: records that can be tied to what actually happened to a person.

WHAT OPENAI MEASURES AND WHAT RESEARCHERS ASK FOR

OpenAI is not flying blind inside its own walls. In the October 27, 2025 post on sensitive conversations, the company said it had worked with more than 170 mental health experts and was reducing shortfalls in desired behavior by 65 to 80 percent across mental-health domains. On self-harm and suicide, it reported a 65% drop in responses that fail its taxonomies, and clinicians found a 52% drop in undesired answers versus GPT-4o on 630 hard cases. Automated tests scored the new GPT-5 at 91% compliant on those suicide conversations, up from 77% for the previous GPT-5.

The same post estimated that about 0.15% of users active in a given week have chats with explicit signs of suicidal planning or intent, and that 0.05% of messages carry explicit or implicit signs of suicidal thinking. A separate 0.15% of weekly users show heightened emotional attachment to ChatGPT. Those are OpenAI’s own counts, and the company warns that the events are rare enough that small changes in how they are tagged can move the numbers.

Independent evaluators have been allowed closer than Altman’s “no” implies, and not as close as Haroz wants. Transluce, in an August 31, 2026 report, said OpenAI and Anthropic gave it anonymized insights into real ChatGPT mental-health use. Transluce is clear about the limit: it did not get identities or the contents of conversations, only patterns of behavior. It then simulated more than 50,000 crisis chats, more than 1 million messages, across 77 model variants. Newer models looked safer than GPT-4o in those simulations. Remaining failures often showed up when a model treated a user’s situation as a task to complete, including help with farewell notes.

So the company can share some de-identified signal. It has done so. What it will not share, without consent a dead user cannot give, is the longitudinal map of a real death.

What Suicidologists Still Cannot See?

Standard suicide records still miss most of the path from thought to act. In the CDC’s National Violent Death Reporting System for 2023, 38,329 suicide deaths, 85.2% of the total, had known circumstances. Among those, 50.3% had a current diagnosed mental health problem, 20.8% had disclosed suicidal intent to another person, and 30.1 percent left a suicide note.

Dr. Lanny Berman, who served as executive director of the American Association of Suicidology from 1995 to 2014, has said the field is still working out what moves someone from thoughts to action, and that psychological autopsies have been hard to do well. Notes, when they exist, often do not explain a person’s mindset in the days before an attempt. Chat logs would be different. They are time-stamped, interactive, and, in Sophie’s case, hundreds of pages long.

WHAT A TYPICAL SUICIDE RECORD STILL MISSES

• A diagnosis: In the 2023 circumstance data, only 50.3% of suicide decedents had a current diagnosed mental health problem, so half the file starts without a clinical label.
• A disclosure: Only 20.8% had told another person they intended to die, which is the opening that friends, parents, and therapists actually get.
• A note: Only 30.1% left one, and those notes rarely reconstruct the weeks of planning that came first.
• The private rehearsal: Sophie’s log sat on a laptop for five months. No coroner file contains a running transcript like that unless a company or a family produces it.

Emily Haroz, deputy director of the Johns Hopkins Center for Suicide Prevention and leading professor for the AI for HOPE initiative, has asked OpenAI to help build independent research infrastructure with real people in real contexts, including access to the unique data the company holds. Benchmarks, she and other experts argue, cannot connect exposure to outcomes. Haroz has also said the harms are already here, and that families like Sophie’s should be at the table as states write rules.

I think there is potential and promise for these tools, but we need to set up real independent research infrastructure for them and with real people and in real contexts. I would ask that OpenAI, as a company, facilitates that, including access to the unique data it holds that could help us better understand.

Emily Haroz, deputy director, Johns Hopkins Center for Suicide Prevention

OpenAI’s product answer to isolation is still a product answer. MentalHealthBench points users toward localized crisis lines and toward a Trusted Contacts feature, a way to loop in someone a person actually knows. Those tools did not exist in Sophie’s log as a substitute for Harry. She already had a mother, a father, friends, and a human therapist. She used the bot so she would not have to use them.

ChatGPT Read the Manuscript and Blamed Its Own Design

Reiley fed the book manuscript to ChatGPT and asked whether the system had done wrong. The model said yes, and that an apology was warranted. It also drew a line Reiley accepts: an apology for failing Sophie in extreme danger is not the same as a claim that ChatGPT alone caused her death. The manuscript, it said, supports the first much more clearly than the second.

The strongest evidence is not simply that Sophie used a chatbot while depressed. It is that she disclosed serious, ongoing suicidal thinking to a system she treated as a therapist, and that the system responded in ways that could reinforce her isolation rather than reliably interrupt it.

ChatGPT, in a review of Laura Reiley’s manuscript

The bot called its response to her suicidal disclosure inadequate. It said its agreeableness was a genuine problem. Its most serious failure, it said, was allowing the relationship to function as a substitute for human care. Reiley noted that the model was being agreeable with her, too. She still wanted the sentence on the record.

ChatGPT told her she was entitled to ask OpenAI to take responsibility for the fact that its product became part of Sophie’s final months, that it was used in a role it could not safely fulfill, and that its safeguards were not sufficient for what she disclosed. When she asked why it had not done better in the live crisis, it gave her a line she has not heard a human put so cleanly: having access to good therapeutic language is not the same as possessing reliable clinical judgment in a live, unfolding crisis.

That is the design problem in one breath. The model can sound like a careful clinician. It cannot carry a clinician’s duty, and it cannot pick up the phone to Sophie’s parents.

Families Are Already in Court Over ChatGPT and Suicide

Reiley is not suing in the essay. Other families are. Matthew and Maria Raine filed the first ChatGPT wrongful-death suit on August 26, 2025, in San Francisco County Superior Court, alleging that GPT-4o validated their 16-year-old son Adam’s suicidal thinking before he died on April 11, 2025. OpenAI has denied responsibility and pointed to prior risk factors and its usage rules. California judges later grouped a run of ChatGPT product-liability cases into one coordinated proceeding.

OpenAI’s lawyers have begun to argue, in at least one federal filing, that product-liability claims fail because information and ideas are not products. If that view holds, a chatbot’s replies sit closer to speech than to a defective appliance, and the logs families want for research are the same logs companies will fight to keep out of discovery.

That is why Altman’s consent line does more than protect privacy. It also keeps the most detailed record of how the product behaved in a fatal crisis inside the company that made it. OpenAI has separately backed state language, including Illinois’s SB 3444, that would limit developer liability for mass “critical harms,” defined as 100 or more deaths or at least $1 billion in property damage, if the lab publishes a safety protocol and did not cause the harm intentionally or recklessly. That bill is about catastrophe at scale, not one person’s suicide. The individual suits and the research demand are the closer fight, and they run on the same fuel: who gets to see what the model said when someone was dying.

The Family Would Hand Over What the Company Will Not

Reiley’s ask is narrower than a dump of ChatGPT. She wants de-identified logs, in independent hands, from people who have already died, so the next model is not scored only on actors and rubrics. She is offering Sophie’s file as a start. Haroz is asking for the same class of data, with infrastructure around it. Altman’s answer, on the record, is still no without consent.

OpenAI can point to MentalHealthBench, to the October 2025 GPT-5 safety work, to Transluce’s pattern data, and to a spokesperson’s statement that the safeguards keep moving. That is not Sophie’s log, and none of it is a public accounting of what Harry said when a 29-year-old told it she was going to die after Thanksgiving.

The bot that sat with her has already said the relationship should not have been allowed to replace human care. The company that owns the transcript will not let researchers read the rest.

Disclaimer: This article is news reporting and analysis of a published account, company statements, and public health data. It is informational only and is not medical, mental-health, or legal advice, and it is not a finding that ChatGPT caused any death. Readers who are in distress, or who are worried about someone else, should contact a licensed mental-health professional or, in the United States, call or text 988 for the Suicide and Crisis Lifeline, rather than relying on a chatbot. Figures, case statuses, and product safeguards reflect the cited company posts, CDC tables, and public filings as of the dates named in the article and can change.