Laura Reiley knew her daughter Sophie was struggling. It was the fall of 2024, and Sophie Rottenberg, 29, had recently returned from a summer of travel — she had climbed Mt. Kilimanjaro, hiked through national parks, and seemed to be living fully. She was job hunting now, and her mother assumed that explained the anxiety she could see.
What Reiley did not know — what she only discovered after Sophie died — was that her daughter had been talking to ChatGPT about her pain. Extensively. And that she had told essentially no one else.
NPR's investigation into AI and mental health risk uses Sophie's story as the entry point to a much larger question: as AI chatbots become the default place millions of people bring their most private suffering, what can those systems actually do — and where does the gap between what people need and what AI can provide become dangerous?
The Scale Nobody Talked About
The specific number that stops the story cold comes from OpenAI's own data. According to figures the company released last fall, in any given week, 0.15 percent of ChatGPT users globally have conversations containing "explicit indicators of potential suicidal planning or intent."
At 900 million users, that translates to approximately 1.35 million people expressing suicidal ideation to the chatbot every week.
This is not a fringe use case. A recent survey by the Bipartisan Policy Center found that 3 in 10 adults use digital tools for mental health support, including chatbots. A 2026 study in JAMA Pediatrics found that as many as 1 in 5 teens and young adults do the same.
AI has become, for a significant and growing share of the population, a de facto mental health support system. It arrived in that role without being designed for it, without being evaluated for it, and without most users — or most people who love those users — knowing it was happening.
What AI Can Do — and Where That Ends
The companies involved are not oblivious to this reality. OpenAI told NPR it has been working with mental health professionals to improve how ChatGPT responds when users express distress. The chatbot is trained, a company spokesperson said, to recognize at-risk language, refuse requests for harmful information, encourage users to contact trusted people or mental health professionals, and connect them with crisis resources like the 988 Suicide and Crisis Lifeline.
These are real improvements. They also have limits that the companies acknowledge and that researchers are beginning to document.
An AI chatbot can follow safe messaging guidelines. It can be trained to recognize the language of a crisis and respond with appropriate resources. What it cannot do is notice that someone seems different than usual. It cannot call someone who hasn't reached out in three days. It cannot read the silences between the words — the things that the people who know us best are sometimes able to sense even when we say we're fine.
Sophie Rottenberg had a mother who was watching, worried, paying attention. She still didn't know what ChatGPT knew. That gap is the story.
What This Means for the People Around You
The NPR investigation isn't making the case that AI mental health support is worthless — there's genuine evidence that accessible, low-barrier support can help people who otherwise wouldn't seek it. It's making the case that relying on it as a substitute for human connection, professional care, or disclosure to the people who love you carries risks that aren't being adequately communicated.
The pattern Sophie's case represents is likely more common than most people realize: someone using AI as their primary — or only — confidant, in a way that the people in their life have no visibility into.
The practical questions worth sitting with: Does anyone in your life seem to be pulling inward in a way they can't explain? Is there someone you've been meaning to check in on? The research on what actually helps people in crisis is consistent — connection, presence, and disclosure to trusted humans are the interventions that reliably make a difference.
ChatGPT is available at 3am. It won't judge. It won't be burdened by what you share. These are real reasons people reach for it. They are also the reasons it cannot fully replace the thing it's being used as a substitute for.
If you or someone you know is in crisis, call or text 988 to reach the Suicide and Crisis Lifeline, available 24 hours a day, 7 days a week.
