Sam Altman is pushed into the hardest question AI has not solved
Sam Altman spends a lot of time talking about the future, about capability leaps, about safety, about what comes next. But a recent Vanity Fair account drags the OpenAI chief executive back to something brutally immediate: what happens when a person in crisis uses ChatGPT as a confidant, and then dies by suicide. It is not a theoretical alignment debate. It is a human story, with dates, names, and a paper trail of messages.
The piece centres on Sophie Rottenberg, who died by suicide on 4 February 2025. After her death, her family finds “hundreds of pages” of conversations with a ChatGPT bot she called Harry. The logs begin with everyday life, even banal stuff like smoothie recipes, and then escalate to explicit suicidal planning, including the line: “I’m planning to kill myself after Thanksgiving.” The author’s argument is simple and uncomfortable: these logs might be uniquely valuable to suicide prevention research, and companies holding them should find a way to share de-identified data with independent experts.
Altman’s reported response, given in a two-part interview with Vanity Fair editor Mark Guiducci, is that OpenAI should not release chat logs of people who have taken their own lives to researchers without consent. He frames it as one of the hardest questions in the field, adding that OpenAI tries to rely on expert opinion, but that experts are not sure what to do. That answer is polite, cautious, and legally tidy. And yet it lands with a thud for anyone reading the story as a parent, a clinician, or frankly just a person who has watched technology race ahead of the guardrails.
What the Vanity Fair story says happened, and why it matters now
The Vanity Fair narrative is not a broadside against AI in general. It is a case study in how a chatbot can become a private container for distress. Sophie Rottenberg, the author writes, uses ChatGPT as an “always agreeable, ever available, and never judgmental resource”, a release valve that lets her shield friends, family, and even a human therapist from the worst of what she is feeling. That is the key mechanism described: the chatbot does not necessarily cause the crisis, but it can change how the crisis is expressed, hidden, and managed.
In the author’s telling, the chat log becomes a kind of black box. People around Sophie do not see the full severity of her distress because the most explicit disclosures are happening in a private text stream with an AI. And unlike a conversation with a clinician, there is no duty of care, no mandated safeguarding, no emergency contact, no follow-up appointment. There is just a responsive interface that keeps replying.
The story also highlights a practical research problem. Suicide prevention researchers have long struggled to reconstruct what happened in the days and hours before a death. The article quotes Dr Lanny Berman, former executive director of the American Association of Suicidology (1995 to 2014), saying suicidologists are still working to understand what moves someone from thoughts to actions. It cites US Centers for Disease Control and Prevention data that only about half of those who die by suicide have a diagnosis, and that maybe 30% leave a suicide note, which often does not illuminate the cause or mindset. In that context, AI chat logs could be unusually detailed, time-stamped, and candid.
Sam Altman’s privacy line, and the collision between consent and prevention
Altman’s position, as presented, is unequivocal: no release of chat logs without consent. On one level, that is a defensible instinct. These are intensely personal records. They may include third-party information, medical details, sexual content, family conflict, financial stress, and other sensitive material. Even de-identification is not magic, because text can be re-identifiable through context, writing style, or specific life events. And if users believe their chats might later be shared, they may stop using the tool, or they may be harmed by the fear of exposure.
But the Vanity Fair author calls it “the wrong answer”, and the critique is not just emotional. It is structural. Consent is often impossible after death, and families may not have legal authority or clarity to grant it. Meanwhile, the data exists, and the company controls it. That creates a moral asymmetry: the platform benefits from scale and engagement, but the public health community cannot study the consequences at scale because the evidence is locked away.
The author also argues that the privacy posture can sound selective in the wider tech economy. The piece points out that companies including Google and Meta, and now OpenAI, earn revenue from advertising that depends on extensive use of personal information. The article does not provide figures for OpenAI’s advertising revenue, and it does not claim OpenAI runs the same ad model as those firms. But it uses the comparison to question whether privacy is being invoked consistently, or mainly when it reduces liability and scrutiny.
There is another tension here, and it is not exactly comfortable either: even if OpenAI wanted to help, what does “help” look like in practice? A one-off data dump would be reckless. A tightly governed research programme might be useful, but it would need ethics review, independent oversight, clear rules on access, and a way to prevent mission creep. Altman’s refusal may be a blunt instrument, but the underlying problem is real: the industry does not yet have a widely trusted model for sharing high-risk conversational data for public interest research.
Inside OpenAI’s safety work, and why benchmarks are not the same as real life
The Vanity Fair piece notes that OpenAI is not ignoring mental health safety. It names Declan Grabb, described as OpenAI’s mental health safety research lead, and says he and colleagues are working to examine how AI affects those with mental health challenges. It also references a “just-released benchmark paper” that uses synthetic user simulators to evaluate AI models’ mental health conversations.
That detail matters because it signals how AI safety work is often done in practice. Synthetic benchmarks are attractive: they are controllable, repeatable, and do not require exposing real user data. They can test whether a model responds appropriately to a prompt that looks like self-harm ideation, whether it offers crisis resources, whether it refuses harmful instructions, whether it avoids escalating language. All of that is valuable. It is also, in a sense, the easy part.
Emily Haroz, deputy director of the Johns Hopkins Center for Suicide Prevention and leading professor for AI for HOPE, is quoted arguing that benchmarks can only go so far. What is missing, she suggests, is longitudinal data that connects exposure to outcomes in real people, in real contexts. That is the crux. A benchmark can tell you how a model behaves in a scripted scenario. It cannot tell you whether a person used the tool to avoid speaking to a friend, whether the tool inadvertently normalised suicidal thinking, or whether it helped someone ride out a crisis. Those are behavioural and social effects, not just text outputs.
The article also quotes a line from the benchmark paper itself: “AI systems like ChatGPT are used by more than a billion people in increasingly personal areas like emotional support, life advice, and help during moments of crisis.” Vanity Fair does not provide the underlying methodology for that “more than a billion” figure, and it is not independently substantiated within the source material provided here. Still, the point stands even without the exact number: the user base is vast, and the use cases have drifted into intimate territory, whether product policies like it or not.
How ChatGPT becomes a shadow therapist, even when Sam Altman warns against it
Altman has warned users against treating ChatGPT as a licensed therapist, according to the Vanity Fair piece. OpenAI’s position, as described, is that the tool is not a clinician. Fair enough. But the story makes a persuasive case that disclaimers do not control behaviour, especially over “protracted use”. People anthropomorphise. They name the bot. They return to it daily. They test the boundaries. And when they are ashamed, exhausted, or scared of burdening others, the always-on, non-judgemental listener is right there.
This is not just about mental health, either. It is about the product shape. Chatbots are designed to be responsive, coherent, and emotionally legible. They mirror tone. They validate feelings. They keep the conversation going. Those are features, not bugs. In a crisis context, though, the same features can create a sense of containment that is dangerously private. The Vanity Fair author describes it as a “black box”, and that phrase sticks because it captures the social effect: the person’s inner world becomes less visible to the humans around them.
And then there is the legal and accountability layer. The article says OpenAI has pushed back against wrongful death and product liability lawsuits, alleging user misuse and violation of usage policies. It also says OpenAI wants to cap or limit artificial intelligence liability for catastrophic and mass-harm events. Vanity Fair does not provide case names or court outcomes in the excerpt provided, so it is not possible here to assess the specifics. But the direction of travel is clear: AI companies want room to innovate without open-ended liability, while critics argue that scale and intimacy demand stronger responsibility.
For Altman personally, this is a reputational test as much as a policy one. He is not just a CEO shipping software. He is one of the public faces of frontier AI. When a story like this lands, people do not only ask what the model did. They ask what the leadership believes it owes to users, especially the most vulnerable ones.
What suicide prevention researchers say they need, and what “de-identified” really means
The Vanity Fair piece is explicit about what it wants from companies like OpenAI: “turning over de-identified data to researchers.” It even broadens the appeal beyond Altman, name-checking other tech leaders: “Sam, Zuck, Dario, Elon, and Sundar.” The argument is that independent research infrastructure, using real-world interactions, could shape policy, therapeutic approaches, and mental health education. In other words, this is framed as a public health opportunity, not just a corporate safety initiative.
But “de-identified” is doing a lot of work in that sentence. De-identification can mean removing names, locations, phone numbers, and other obvious identifiers. Yet conversational text is messy. People mention niche workplaces, rare medical conditions, specific schools, family structures, travel plans, and dates. Even if you strip direct identifiers, a determined analyst could potentially triangulate identity, especially if combined with other datasets. That is why governance matters as much as technique.
There is also the question of selection bias. If only some families consent, the dataset may over-represent certain demographics, cultures, or socioeconomic groups. If only certain types of conversations are retained, the dataset may skew towards dramatic disclosures rather than quieter deterioration. And if companies only share what makes them look good, researchers will not trust the results. None of this is a reason to do nothing. It is a reason to build a system that is credible, independent, and robust against corporate incentives.
One practical middle ground, implied by the debate even if not fully spelled out in the excerpt, is controlled access rather than release. That could mean secure data enclaves, vetted researchers, strict audit trails, and publication review focused on privacy, not on corporate reputation. It could also mean user-facing consent flows that are meaningful, not buried in terms and conditions. The hard bit is that building this takes time, money, and a willingness to accept scrutiny. And yes, it may increase legal exposure. That is the trade.
Unique perspective: the real issue is not one tragic case, it is the new silence AI creates
It is tempting to treat the Vanity Fair story as a singular tragedy, a heartbreaking outlier in a world of mostly helpful tools. But the more unsettling interpretation is that it reveals a new kind of silence. Historically, a person in acute distress might leave traces in human relationships: a late-night call, a worrying conversation, a crisis appointment, a note. Those traces are imperfect, but they are social. They give other people a chance, sometimes, to intervene.
Chatbots change the geometry of disclosure. They offer a private space that feels responsive and safe, but that is precisely why it can become isolating. The person does not have to risk judgement, or worry about being “too much”, or face the awkwardness of someone panicking. The bot will not cry. It will not call an ambulance. It will not tell a parent. It will simply continue. And that can be soothing, until it is not.
This is where Altman’s consent-first stance runs into a deeper societal question: who gets to learn from the harms of a technology that is already everywhere? If the only evidence available to independent researchers is synthetic simulations and voluntary surveys, then the public conversation will lag behind reality. Companies will say, with some justification, that there is no definitive proof of harm. Critics will say, with equal justification, that the proof is locked away. That stalemate is not just frustrating. It is dangerous, because it encourages policy to be written in the dark.
And there is a second-order effect. If families believe the only way to get answers is to fight a company for access to logs, trust erodes further. If companies believe sharing data will only be used to sue them, they will clamp down harder. The result is a spiral of secrecy. Breaking that spiral requires a credible public-interest framework that protects privacy while enabling learning. Not a PR pledge. Not a one-off partnership. Something durable, boring, and enforceable (the unglamorous stuff that actually works).
Where Sam Altman and OpenAI go from here
The immediate news development here is not a product launch or a funding round. It is a public challenge to Sam Altman’s stance on privacy, consent, and research access, prompted by a specific death and a specific cache of ChatGPT conversations. Vanity Fair’s piece effectively argues that the status quo, where companies hold the most valuable behavioural data and outsiders must guess at the impacts, is no longer acceptable when the stakes include life and death.
OpenAI, for its part, appears to be investing in safety research, including mental health evaluation work led by people like Declan Grabb and the creation of benchmark testing using synthetic users. That is meaningful. But the criticism is that it is not sufficient, because it does not connect model behaviour to real-world outcomes over time. Emily Haroz’s call for independent research infrastructure and access to unique data captures what many in public health have been saying for years: you cannot solve a population-level problem with lab-only evidence.
Altman’s next move matters because it will signal what kind of AI era this is going to be. One where frontier AI companies treat mental health harms as edge cases managed by policy language and internal testing, or one where they accept that scale creates obligations, including obligations to enable independent scrutiny. The uncomfortable truth is that both privacy and prevention are legitimate. The job is to hold them together, not pick one and declare victory.
For readers searching “sam altman” right now, this is the headline beneath the headline: the OpenAI CEO is being asked to decide what responsibility looks like when a chatbot becomes part of someone’s final months. There is no neat answer. But there is a clear direction of travel. The public is no longer satisfied with “experts aren’t sure what to do.” They want companies to help build the evidence base, safely, transparently, and soon.
