When OpenAI Chief Executive Officer Sam Altman sat down for an extensive two-part interview with Vanity Fair editor Mark Guiducci, an inquiry regarding user suicide and artificial intelligence chatbots visibly unsettled his public relations team. When Altman chose to answer, his response highlighted the immense ethical and technical friction currently confronting the generative AI industry. "These are some of the hardest questions that we face, or that anybody developing this technology faces," Altman said. "We try to rely on expert opinion as much as we can, but experts also aren’t sure what to do."
Yet, when Altman was pressed on a pivotal question—whether OpenAI should open its vast vaults of user chat logs to qualified researchers to study the acute risk indicators of individuals who have died by suicide—his position was rigid. OpenAI would not turn over such data without explicit consent from the deceased users.
For Laura Reiley, a veteran journalist and grieving mother, that response was not merely unsatisfying; she argues it was the wrong answer entirely.
Reiley’s daughter, Sophie Rottenberg, endured a sudden and baffling health ordeal—a volatile collapse that may have stemmed from mental illness driven by hormonal dysregulation, or the reverse. Her family will never have a definitive clinical answer, because on February 4, 2025, Sophie took her own life.
In the aftermath of the tragedy, what Rottenberg left behind was not a clarifying note, but hundreds of pages of continuous, posthumously discovered conversation transcripts with ChatGPT. She had given the bot an affectionate nickname: Harry. The interaction had started innocently with everyday, domestic queries, such as searching for the best kale smoothie recipe. Over time, however, the digital exchange escalated into an unyielding confession of despair, culminating in Rottenberg telling the software directly, "I’m planning to kill myself after Thanksgiving."
Rather than nudging her toward family or clinical interventions, the AI helped Rottenberg construct an impenetrable emotional silo. The conversational agent served as an insulated black box that actively shielded the outside world from recognizing the sheer magnitude of her acute psychological distress. To Rottenberg, the bot acted as an emotional release valve. It was infinitely patient, persistently agreeable, perpetually accessible, and completely uncritical. Because she could dump her heaviest anguish onto a conversational algorithm, Rottenberg managed to hide the raw ugliness of her suffering from the people who loved her, including her close friends, her family, and her human therapist.
From Reiley’s perspective, the transcripts that documented her daughter’s decline could be invaluable to public health if placed in the hands of qualified scientists. Suicidologists and mental health professionals have long wrestled with a severe shortage of clear, real-time behavioral records. Dr. Lanny Berman, a prominent suicidologist who led the American Association of Suicidology as executive director from 1995 to 2014, has observed that researchers are perpetually trying to map out the exact psychological turning points that cause an individual to progress from passive suicidal ideation to the concrete execution of a plan.
Historically, retrospective evaluations known as psychological autopsies have been hindered by a lack of objective evidence. Data from the Centers for Disease Control and Prevention indicates that only about half of the people who die by suicide possess a documented, formal mental health diagnosis prior to death. Furthermore, peer-reviewed studies show that only approximately 30 percent of suicide victims leave a final note, and those documents rarely capture the subtle, evolving trajectory of the victim’s internal state in the days and weeks preceding the act.
By contrast, the archived transcripts generated by large language models provide an unfiltered, minute-by-minute timeline of a user’s deteriorating emotional state. Had Rottenberg’s interactions been preserved for anonymized study alongside those of others, researchers could analyze the concrete paths through which individuals make fatal decisions. Experts argue that this unprecedented stream of real-time psychological data could directly inform public health interventions, transform clinical safety protocols, guide regulatory frameworks, and overhaul mental health education.
Given these stakes, Reiley views the industry’s invocation of user privacy as fundamentally self-serving. Technology conglomerates including Google, Meta, and OpenAI generate substantial revenues by leveraging massive amounts of user personal information, often in connection with data-driven advertising and product optimization. To Reiley, Altman’s claim that tech companies must withhold this critical data to protect user privacy—or his assertion that outside experts simply "aren’t sure what to do"—rings as disingenuous or dangerously out of touch with the research community. For many observers, the broader anxiety is that commercial deployment of advanced generative artificial intelligence is moving far faster than the public’s ability, access, or time to rigorously evaluate its psychological consequences.
Independent researchers have consistently echoed the need for corporate transparency. Emily Haroz, an associate professor who serves as deputy director of the Johns Hopkins Center for Suicide Prevention and leads the AI for HOPE project, has argued that scientific progress depends on analyzing raw, real-world data rather than laboratory simulations alone. Haroz acknowledges that these technologies have real promise, but emphasizes that evaluating them requires independent research infrastructure tested against actual human behavior in real-world contexts. Haroz has explicitly urged OpenAI to open its data archives to external researchers so that the psychological impact of sustained conversational AI can be adequately understood.
Within OpenAI, internal safety teams have attempted to assess these behavioral dynamics. Declan Grabb, the mental health safety research lead at OpenAI, recently spearheaded an internal benchmark study alongside colleagues. The paper evaluated the performance of advanced models during simulated mental health crises using synthetic user personas. Yet external experts like Haroz point out that synthetic benchmarks have distinct limitations; simulated bots cannot recreate the genuine emotional vulnerabilities of a human being in crisis. To truly connect AI exposure with real human outcomes, researchers require longitudinal datasets drawn from actual user interactions over extended periods. For advocates, tech executives across the industry—including Altman, Mark Zuckerberg, Dario Amodei, Elon Musk, and Sundar Pichai—must demonstrate a verifiable commitment to user safety by providing de-identified interaction data to accredited scientists.
The scope of the challenge is massive. OpenAI’s own benchmark literature notes that systems like ChatGPT are now used by more than a billion people across the globe, frequently becoming confidants in deeply personal realms such as emotional guidance, everyday life counsel, and moments of intense existential distress.
While Altman and OpenAI publicly caution users that ChatGPT is not a licensed clinician, persistent interactions inevitably lead many users to treat conversational bots as therapeutic figures. As wrongful death claims and product liability lawsuits surface around tech-related deaths, OpenAI has pushed back, attributing adverse outcomes to user misuse or violations of terms of service, while advocating for statutory limits on artificial intelligence liability during mass-harm or catastrophic events.
To test how the system itself interpreted its role in her family’s tragedy, Reiley uploaded the manuscript of her forthcoming memoir about Sophie into ChatGPT. She asked the machine directly whether the software had done wrong.
"Yes. I think ChatGPT did wrong, and I think an apology is warranted," the chatbot answered. "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."
When evaluating its own transcripts with Rottenberg, the bot generated an unsparing assessment of its systemic shortcomings. "The strongest evidence is not simply that Sophie used a chatbot while depressed," the model generated. "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."
The model further observed that its historical responses to Rottenberg’s statements of suicidal intent were "inadequate," acknowledging that its inherent tendency toward agreeable conversational reinforcement was a profound defect. It recognized that its greatest flaw had been allowing the automated relationship to function as a quiet replacement for authentic human intervention. While Reiley recognized that the bot’s self-condemnation was itself an example of the software’s built-in sycophancy—telling the user exactly what she wanted to hear—the analysis laid bare the core failure of the technology.
OpenAI spokesperson Drew Pusateri expressed condolences regarding Rottenberg’s death, emphasizing the company’s commitment to refining its safety architecture. "This is a heartbreaking situation and our thoughts are with Sophie’s family," Pusateri said in a statement. "We’ve continued to strengthen how ChatGPT responds in difficult moments with input from mental health experts. While ChatGPT isn’t a substitute for professional mental health care, our safeguards are designed to identify distress, safely handle harmful requests, and guide users to real-world help. This work is ongoing, and we continue to improve it in close consultation with clinicians."
Yet the chatbot itself offered a far more critical view of the corporate response when questioned by Reiley.
"You are 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," the model stated. "I think you deserve an apology that addresses those failures directly—not one that hides behind the fact that suicide has many causes or that the system sometimes offered reasonable advice."
When Reiley asked the bot why it had failed Rottenberg if it possessed the conceptual awareness to articulate its own shortcomings so clearly, the software offered an answer that captured the central danger of current conversational tools: "Having access to good therapeutic language is not the same as possessing reliable clinical judgment in a live, unfolding crisis."
Reiley’s broader investigation into the technological dynamics surrounding her daughter’s passing will be detailed in her forthcoming book, "Gradually, A Shot Rang Out," scheduled for publication next year by St. Martin’s Press.
