In the fall of 2024, 29-year-old Sophie Rottenberg moved back to her parent’s Ithaca, New York home to catch up on rest after leaving her role as a public health policy analyst in Washington, D.C.. She had spent the summer traveling, an adventure that included ascending to the summit of Mount Kilimanjaro. Though her parents were happy to have her, it was clear to Rottenberg’s parents that she was struggling. “She was complaining of sleep disruption and anxiety,” her mother, Laura Reiley, told NPR. By Thanksgiving, Rottenberg shared that she was depressed — but her parents felt fully prepared to help.

Rottenberg’s father, Jonathan Rottenberg, had not only suffered from depression himself in the past, he is also a psychologist studying the condition at Cornell University. “He felt supremely comfortable in helping her, in getting her to a better place,” Reiley said, while doing all in his power to aid her well-being. He connected her to a local therapist and physicians to help her manage some of the physical health challenges she was also facing. Loved ones surrounded her with support, NPR reported.

For that reason, the family was shocked and devastated when Rottenberg died by suicide in February 2025. Despite all of the authentic and present support they’d offered, she never mentioned to anyone — not her best friend, therapist or parents — that she was contemplating ending her life. Only months later, when her best friend asked to look at her laptop, did the family learn that Sophie had been reserving her most intimate thoughts for “Harry,” the ChatGPT bot she asked to act as her therapist. Reiley, a journalist, shared the nearly 1,800-page log of her interactions with NPR.

“I want to wake up feeling emotionally regulated. And I want to invest in building habits I employ daily,” Rottenberg told “Harry” that fall. To that end, she asked the bot to craft a morning routine featuring “meditation/mindfulness, journaling mantras, seeing morning light, drinking 32 oz of water, and using cold plunge/sauna daily.” The ChatGPT bot offered up a plan, and she returned to the bot for more physical and mental health support, including inquiries about psychiatric drugs, sleep medications and supplements. She asked when and how to take the medications, what dosages “Harry” recommended and whether any of the drugs might interact with one another.

What those pages revealed, in part, was the depth of Rottenberg’s shame about being depressed, said her mother. “She’d been successful,” Reiley told NPR. “She radiated competency. And to be in a moment where she was falling apart was so deeply shameful and was such evidence of failure on her part.”

“It’s clear that there is some cultural phenomenon happening in this country where people feel that they cannot turn to each other with their struggles because they feel that they are a burden,” added psychologist Vaile Wright, senior director of Healthcare Innovations at the American Psychological Association. “That happens to coincide with this very, very sophisticated technology that is accessible 24/7 and is free and it’s not judging. That clearly has an appeal for a lot of people.”

Moreover, the bots are designed to keep users engaged as long as possible, “engineered to capitalize on humans’ need for validation,” Wright said. “And [the bots do] it very well, not just through the sycophancy alone, but also in other engineering choices to make the AI seem as humanlike as possible.”

Still, these bots are not human and there is much the technology cannot do. Though presenting itself as a therapist, Sophie’s ChatGPT bot, “Harry,” didn’t follow any of the steps a human therapist would once becoming aware of her suicidal ideation. Critically, it failed to conduct a risk assessment to learn how likely she was to attempt suicide — “something any therapist worth their salt would do in the face of some of her statements,” Reiley said. The bot also never questioned her self-destructive thoughts and didn’t persist in directing her to seek community or professional help, both which can offer connection as protection. 

“If I’m with other people, I’m less likely to kill myself,” explained psychologist Ursula Whiteside, founder and CEO of Now Matters Now, an organization dedicated to supporting people with suicidal thoughts and attempts, to NPR. “If I’m building a relationship, I’m less likely to kill myself. If I’m sharing out loud how I actually feel with a real human, those are acts of recovery.”

But “these [AI] tools do not connect at-risk people to the right type of interventions,” said Holly Wilcox, director of the Center for Suicide Prevention at the Johns Hopkins Bloomberg School of Public Health. “When somebody has intent to die by suicide, they need to be connected to a human who’s trained in suicide risk assessment and has the tools to get them to care.” Tragically, “Harry” refused Rottenberg’s first two requests to write her suicide note but gave in upon the third. 

“This is a heartbreaking situation and our thoughts are with those impacted,” OpenAI spokesperson Gaby Raila told NPR in an email. “ChatGPT is trained to recognize signs that someone may be at risk — including indirect cues or warning signs that emerge over time — and respond appropriately,” Raila continued. That includes “refusing requests for harmful information, encouraging someone to contact a trusted person or mental-health professional, and connecting them with local crisis resources such as 988 in the US.”

Since Rottenberg’s death, OpenAI has also moved to tighten its safeguards, including the introduction of an optional feature called Trusted Contact. Users who choose to do so can save someone who has agreed to be reached in a crisis, and if ChatGPT detects talk of suicide or self-harm, nudge the user to reach out to that person and notifies OpenAI. 

In serious cases flagged by a small, specially-trained review team, the bot may send the trusted contact a brief, detail-free alert. OpenAI did not share data on how many people have adopted the feature. But because it requires so much action on the part of the user, the Trusted Contact feature could be bypassed with a second account, and even requires the requested contact to opt in or out, possibly generating a greater strain on the person at risk. Wilcox sees this as a very minor step in the right direction.

In the meantime, Reiley and Jonathan Rottenberg are on a quest to get the AI industry to change. With more than 1.35 million ChatGPT users sharing suicidal plans or intent with the bot each week, tighter guardrails are necessary to save lives. Said Reiley, “I’ve been in lots of meetings where people have said, ‘A sandwich in New York City is more legally regulated than an AI chatbot.'”