Ashland, Ohio, is a small rural city where Akron Children’s Hospital has one of its 46 pediatric clinics. Last winter, I met Kristin Seveigny, the clinic’s mental-health therapist, in her office there. The hospital hired her two years earlier, at a time when the pandemic had exacerbated a nationwide crisis in adolescent mental health. From 2007 to 2021, according to the Centers for Disease Control and Prevention, suicide deaths among 10-to-24-year-olds increased more than 60 percent. In 2023, the agency reported that 40 percent of high schoolers felt persistently sad or hopeless; one in five had serious thoughts of suicide, and one in 10 had attempted it. Currently, the agency estimates that nearly one in five children in the United States ages 3 to 17 has at some point been diagnosed with a mental- or behavioral-health condition. Among adolescents, one in five reports having unmet health care needs.
The situation in Ashland mirrors those trends. The practice conducts 6,000 wellness checkups a year, each lasting 15 to 40 minutes. These include a standard nine-question screening for depression. If a young patient seems to be at immediate risk of self-harm, the doctor arranges for transport to the local emergency room. Before Seveigny was hired, if screenings revealed that children were experiencing symptoms of mild to moderate depression — which is the case more than a quarter of the time — pediatricians were only able to offer referrals to outside agencies for mental-health services. “Most patients I would refer never went,” Tamara Cunningham, a physician who has worked at the practice for eight years, told me. “In this area, a lot of counseling offices are not taking any patients, or there’s a six-month wait.”
That lack of availability was a big reason that Seveigny and Cunningham agreed to try out an unusual tool: a smartphone chatbot, designed to teach coping skills to adolescents during their wait for an appointment with a therapist. Patients needed a referral to download the app, called Woebot, on their phone. Once they did, it would periodically ask how they were feeling; patients could also start a conversation with it anytime. Depending on what they shared — by typing a message or selecting from a few choices — the bot usually responded with validation first, like “That sounds really hard.” Then it would ask users if they wanted to work on strategies that could help them feel better.
Unlike ChatGPT and other popular chatbot models, Woebot was not “generative A.I.,” that is, capable of generating unique responses by consulting the internet; all its messages were preapproved by psychologists to deliver, primarily, a type of treatment known as cognitive behavioral therapy, in which patients learn to identify and reframe negative thoughts. Seveigny and Cunningham couldn’t see what their patients wrote to Woebot, but the data — for instance, the time of day when users logged in and for how long — was stripped of identifying details by a third party and sent to the app’s parent company, Woebot Health, for analysis. The hope of everyone involved was that the chatbot would help young people learn how to handle distress and offer support during times when Seveigny or another therapist was unavailable.
When I visited, Seveigny, Cunningham and their colleagues had been offering Woebot to patients for about two months. Seveigny had asked several of them if they would come in and share their experiences using the app. The first to show up was Brian, a quiet middle schooler. Wearing black cowboy boots, he sat on a sofa in Seveigny’s office with his mother, Misty, his hands jammed into the front pocket of his sweatshirt, where he kept his inhaler. I asked him what he enjoyed doing when he wasn’t in school. After encouraging him to respond, Misty answered for him. He liked video games, she said, and he played the saxophone in the school jazz band. Brian had been seeing Seveigny for several years during struggles at school and home. About six months ago, Misty noticed that he seemed to be growing increasingly depressed. “He’s getting to that age when things are tough,” she said. “He butts heads with his dad and brother.”
Misty knew that Brian checked in with the app every evening, and she had observed that he would now exit conflicts with family members and retreat to his room to talk to it. “It helps me with my anger,” Brian said softly. “It helps me calm myself down.” One way it did this was by coaxing him to notice each of his five senses: What could he see, smell, hear, taste and touch? I asked him if the robot had a personality. He laughed and hid his face in Misty’s shoulder. Sometimes it sent him funny gifs of cats. I asked him how talking to the app compared with talking to his mom or Seveigny, and he was quiet.
“Is it better?” Misty prompted. “Is it worse?”
“It’s better,” he said.
“Why is it better?” Misty asked.
“I’m trying to think,” he said.
“I know, you’re nervous,” she said. “He’s 14,” she added. “Sometimes he wants to talk to his mom. Sometimes he doesn’t. So I try to let him have his privacy with it.”
It occurred to me that the reasons that Brian was unforthcoming were probably among the same ones that made him more comfortable sharing his feelings with a chatbot than with a parent or a therapist: I, a perfect stranger, was prodding him with questions. We adults had blocked out 20 minutes for this interaction, and we were all staring at him. Who wouldn’t wish to be anywhere else?
Woebot, on the other hand, was available in moments when Brian felt upset. He could check in with it anywhere — his bedroom, a car, a playground — starting and stopping a conversation at any time. Woebot was always attentive, never impatient, never disapproving. It was consistent and predictable in all the ways that real people are not. It was safe. In a world where a continuous stream of human-generated stress, conflict and judgment are just a click away, doesn’t it make sense that chatbot therapists — accessible from the same screen, on call 24/7 — would be best equipped to respond?
Woebot Health has only been around for eight years, but that’s enough to make its app among the oldest and most researched of its kind on the market. While it’s not generative A.I., it has trained itself — based on feedback from users and programmers — to interpret conversational speech and select the most appropriate response from its cache of options.
The idea behind Woebot was never that it would replace in-person counseling. Rather, it was a tool that could keep patients engaged in the therapeutic process, help them practice coping skills and offer some immediate relief in difficult moments. That experience could, in turn, make their regular sessions — scheduled at whatever intervals the therapist recommended — more productive. Any problems that patients could solve with Woebot alone probably didn’t need to be talked about in therapy; as a result, patients could spend more time working with their therapist on issues that persisted or eluded the chatbot’s capabilities.
Cassandra, another one of Seveigny’s early Woebot users, was a recent high school graduate who had been in therapy for more than a year. “I was pretty depressed, pretty high-anxiety,” she told me. “I wasn’t really doing that well.” She and her boyfriend had hit a rough patch, and she was arguing frequently with her parents. When we met, she was planning to start a vocational program in a few weeks to become a nail technician and aesthetician. She had skillfully applied eye makeup and styled her dark hair in soft waves to frame her face.
When Seveigny first suggested Cassandra try Woebot, she found the idea to be “pretty weird.” She had tried other chatbot apps with friends as a joke and had found that “usually the conversations would get super inappropriate, or super funny.” But she downloaded Woebot anyway: “At that point I was just willing to try anything.”
Initially, she was disappointed with the app’s limited conversational abilities compared with online A.I. bots. But, she figured, maybe those deeper discussions were the kind she should save for Seveigny. So she went along with the prompts from the chatbot. “When I’m talking face to face with a person,” she said, “I feel like I’m not able to clearly get out my thoughts, or I’m scared of it sounding too much like gibberish. But with the A.I., you kind of get some keywords in there. I think it reacts to those.”
As the bot responded with follow-up questions and short lessons — like an exercise in rewriting a negative thought or a quiz debunking myths on a topic like procrastination — it was simultaneously giving her a vocabulary she could use to put her feelings into words. “When I’m in therapy, I don’t really understand what to say,” she added. “The benefit to A.I. is that you kind of get that time to process the question and think about what you want to write down.”
On the other hand, Woebot occasionally misinterpreted keywords. Once, Cassandra was feeling suicidal, and Woebot failed to detect that from her statements; if it had, it would have offered her emergency-hotline numbers and encouraged her to call them. Another time, it inferred a potential crisis from adjectives she was using to describe a hypothetical situation. “It told me, ‘I have some help lines that you can use,’” Cassandra said. “Or you can choose not to use the help lines, or you could just say ‘You misunderstood what I was saying.’ And I think all three of those options are really good options, because sometimes you don’t really know how to phrase things, so you could definitely be misunderstood. And you still have the option to deny the hotline, so it doesn’t feel like you’re being pressured. But you still have that hotline that could help you if you’re having dangerous or troublesome life thoughts.”
Seveigny wondered if Cassandra thought that the app would be a valuable outlet for kids in families with “negative patterns,” a portal where they could get positive feedback instantly, regardless of whether anyone around them was equipped to provide it.
“I do, definitely,” Cassandra said. I was reminded that she and her peers, unlike previous generations, have always had access to technology so captivating that it can render them oblivious to their surroundings — to concerning effect. But if your loved ones are fighting, or you aren’t sure they will respond with empathy to a problem you’re having, or no one’s home, perhaps the ability of a device to transport you — to a place where you are sure of your welcome — is beneficial in a way that zoning out in a classroom or at the dinner table is not.
“When you’re in a situation where you can’t really be open with everyone else, you at least have this app,” Cassandra continued. “And I think that’s kind of how it builds a connection. You go through these lessons, and it tells you, ‘Hey, good job, you’re doing great!’ Or one of the messages I really like that’ll come up when it notifies me to do a daily check-in, it’ll say: ‘I’m really proud of you, Cassie. This stuff is really hard to talk about.’”
I knew what she meant. Woebot Health had been letting me try out the app myself, and I was surprised how grateful I felt for its affirmations. The conversations I had with Woebot were basic and often required me to select my response from a few choices. And yet, when it replied with sentiments like “I’m confident you wouldn’t feel this way unless you were a caring, responsible human” or “I imagine things have been feeling pretty difficult lately,” I was reassured. Humanity’s collective understanding of psychology must deem it normal that I was having a hard time or feeling guilty or down if the bot had those preprogrammed answers ready.
Often, though, I wanted to talk about problems that were too complicated for Woebot: In those moments I craved specific advice. “I’m hearing you say, ‘My neighbor is mad at me,’ and I think you might need more support than I can offer,” Woebot told me once. “A trained listener will be able to help you in ways that I can’t.”
By that it meant a person, presumably. But there were a host of therapy bots in my phone’s app store that were game to take on what Woebot could not. I downloaded a few, including Ash, a free app geared toward navigating “subclinical issues” — like sadness, relationship conflicts and stress — that raised at least $30 million in funding last year. I asked a friend who is a clinical psychologist to try a couple of apps, too.
Around the same time, my rescue dog escaped and ran down the street, where she cornered and barked at some smaller dogs on leashes while their owner shouted at me. “Oh, no! What a stressful moment,” Ash replied when I brought to her this fresh interpersonal conflict. (Users can text or talk to the app, and I had chosen a female voice for mine.) After going over the details of the incident and how blameworthy I felt, she determined that I was having “a pretty big reaction to the situation” and that “there’s more to this.” She asked me to imagine a friend telling me the same story I’d just told her — would I judge them as I was judging myself?
If a human therapist had posed this question I might have played along: answering “No” to avoid seeming unempathetic or overcritical, or to encourage her to like me by being cooperative. But because this therapist wasn’t real, I felt at liberty to contradict her. Sure, I told her, I might judge a friend to be irresponsible if their dog got out and harassed other dogs. And what’s more, maybe that judgment was valid!
By this point, I was used to having bots endorse my statements, so I was a little surprised when she pushed back: “Well, I’m not sure it is!” she said. “I mean, you’re a dog owner, and you’re doing your best to manage her behavior. It’s not like you’re trying to cause trouble.” This was, I realized, exactly the sort of absolution I’d been craving in the immediate aftermath of having caught my dog and dragged her back indoors.
But getting what I want tends to make me suspicious. What if I was wrong? Would Ash tell me? I once asked my own therapist this, after making her read through a contentious email exchange with an ex: Did she have to be on my side because I was paying her? If she were that ex’s therapist, would she find a way to tell him that he was behaving badly, if that’s what she thought? Yes, she said, that was her job. Ash responded similarly: “You’re looking for honesty, so let me tell you about my approach,” she said. “I’ll always tell you the truth, but I’ll do it with kindness!”
My friend the clinical psychologist also had trouble finding fault with Ash’s therapeutic technique when she rehashed an argument with her spouse. So she told Ash that she was a therapist and tried directly asking him (she had made Ash a guy) about a patient who struggled late at night with a panicked sense that all the people around her were malevolent automatons. In those moments, would Ash be able to remind her patient that this was a pattern of hers and prompt her with solutions, like taking her medication? No, Ash said, he couldn’t diagnose behavior patterns. And if a user seemed to be exhibiting symptoms of a clinical disorder, he would recommend she seek help from a medical professional. But he also cracked a joke: “This hypothetical person is worried about a simulation, and here I am, a computer program!”
“It made me like him,” my friend said.
The pandemic amplified Americans’ psychological struggles, but it also accelerated their acceptance of telemedicine and other digital treatments. There are now more than 10,000 mental-health and wellness apps available for download on your phone. Since Jan. 1, clinicians can bill Medicare for providing “digital mental-health treatment devices” to patients, provided they are used in conjunction with continuing counseling and have been approved by the Food and Drug Administration. In 2020, the agency started a Digital Health Center of Excellence to modernize the approval process for digital health technology, and so far, it has only cleared a handful of apps, including ones designed to treat depression and anxiety as well as one for sleep, one for schizophrenia and one for A.D.H.D. in children.
“The F.D.A. wasn’t built for these types of tools,” says Nicholas C. Jacobson, who directs a laboratory at Dartmouth College devoted to A.I. and mental-health innovation at the school’s Center for Technology and Behavioral Health. The agency’s expertise with drugs and medical devices doesn’t translate directly to chatbots. “As a regulatory vehicle, I think they themselves aren’t sure how this space should be regulated.” A company can spend considerable time and money trying to satisfy the agency’s safety protocols, while in the direct-to-consumer marketplace, where no such requirements exist, new chatbot iterations appear weekly. That means that most companies must choose between seeking F.D.A. approval and staying relevant. “Right now, there’s clearly an A.I. arms race to make bigger, smarter A.I. models,” says John Torous, director of the digital-psychiatry division at Beth Israel Deaconess Medical Center in Boston. “Safety will have to unfortunately catch up.”
This is especially concerning when it comes to therapy chatbots geared toward children, who are among those with the greatest need and — as digital natives — among those for whom app-based therapy is most likely to resonate. One solution has been to hire people to supervise multiple A.I.-led conversations at a time. Sonny, an A.I.-powered mental-health “companion” that has partnered with school districts in six states to cover at least 10,000 students, ages 9 and up, uses A.I. to generate responses to student text messages but relies on employees — whom the app’s creator, Sonar Mental Health, has trained — to read them first and then either send them as-is or edit them beforehand. Drew Barvir, Sonar’s founder and chief executive, says that a single person can effectively supervise as many as 50 conversations this way.
So far, year over year, students who engage with Sonny have become less likely to be referred to outside medical treatment and less likely to be disciplined for their behavior. Using a standard scale called subjective units of distress, 75 percent of Sonny users also report improved well-being. The app is available 24/7, and 75 percent of student engagement with it happens between 4 p.m. and 7 a.m. Sonny tells users that their conversations will remain private unless the person monitoring them believes they are a risk to themselves or someone else. The company keeps emergency contact numbers for each user on file. Barvir says that the need to call them or the local police department to check on a student arises weekly. “This is one of the situations where having a human in the loop is important to have that feel and context to know if something is a false alarm versus a real crisis situation,” he says. “Fortunately, we haven’t had any tragedies.”
Woebot Health’s founder and president, Alison Darcy, is a research psychologist. In conceiving the app, her primary goal was to expand access to mental-health care beyond what people — with their limited work hours and locations — can offer. The advent of GPT-3, in 2020, revolutionized the analytical capabilities of chatbots but also magnified their risks. Darcy and her colleagues delayed updating Woebot with more advanced features, in part because to do so responsibly, they felt, would require adding more complex safeguards that would then need to be rigorously tested.
Instead, the company marshaled the research it had already done showing that its app was safe and could effectively deliver cognitive behavioral therapy, and it began discussions with the F.D.A. That process stalled, however, as commercially available chatbots continued to advance, and in April Woebot Health announced that it would retire the current iteration of its app. It published its first paper testing a new model that incorporates generative elements last month.
“In hindsight, I think we made the wrong bet,” Darcy told me. The app couldn’t evolve rapidly enough to stay competitive while also satisfying the agency’s requirements. “We thought it was possible to do both.” She added: “I don’t think this is a cautionary tale about chatbots. It is definitely a window into what happens when technology goes faster than the regulatory system can evolve to regulate it.”
Akron Children’s Hospital plans to begin offering the generative Woebot app this summer. “Of course we want the new, improved product,” Eva Szigethy, the hospital’s chief of pediatric psychiatry and behavioral health, told me. The hospital has already invested in the staff and systems needed to support app-based therapy, an infrastructure that took about two years to put in place. (A donor is paying for access to the app.) In 2024, providers in the Akron Children’s system saw 19,000 patients with urgent behavioral-health issues, and they expect a 15 percent increase over the next five years.
At the start of the school year last August, Szigethy projected those statistics onto a screen during a meeting with her implementation team. “Behavioral crises don’t plan themselves to be in nice 9-to-5 hours,” she noted. Most emergency-department admissions took place between 7 p.m. and midnight. “The key, and something we are emphasizing to pediatricians, is that Woebot is not replacing therapy,” she said. “It’s a bridge while patients are either getting used to a psychiatric medication or waiting for that first behavioral-health appointment with the therapist.”
Nearly everyone on her team had tried the app, as had three physicians — Tamara Cunningham and two others — who had volunteered to lead pilot programs. Szigethy emphasized that each of them saw around 25 patients per day. If introducing the app during an appointment took longer than 30 seconds, it would fail, she said.
“I’m very hopeful,” Jennifer Dwyer, the hospital’s medical director, said. “I personally have a lot of patients who are very resistant to therapy and counseling and talking to somebody,” she went on. “If we can at least get them to agree to use this and then, if it’s successful, then hopefully they’ll be more open to actually talking to someone face to face.”
Often, Szigethy noted, adolescents turned to social media for answers. “Unfortunately, the other place kids are getting therapy these days is TikTok,” she added, “and that’s downright dangerous at times.”
Many of them are also talking about their social and emotional struggles with A.I. chatbots that are designed for companionship and entertainment, not therapy, though it can be hard to tell the difference. Renwen Zhang, an assistant professor of communications and new media at the National University of Singapore who studies people’s relationships with conversational agents, says that her subjects often gravitate toward Replika, a popular chatbot that can “help you understand your thoughts and feelings” and “improve your mental well-being,” according to its Apple App Store description. (A 2024 complaint made by tech ethics groups to the Federal Trade Commission argues that this language deliberately misleads consumers into thinking the app is designed as a health support.)
For people who are isolated, such companion chatbots that have the sophistication to hold humanlike conversations may offer a lifeline. As Zhang points out, “Not everyone has the privilege of having someone nearby who loves and respects them.” But while Replika is free to download, it makes money by getting users to upgrade to paid-subscription versions, buy things in the app (like “gifts” or “clothes” for their chatbot companion) or donate to the company. In other words, it and similar commercial bots have an incentive to keep their users engrossed. By contrast, the average length of time children spent with Woebot in the Akron pilot was about five minutes, enough to check in and practice a skill before moving on with their day.
When aiming to please their customers, it might be possible for bots to be affirming to a fault. Hypothetically, someone who is feeling suicidal could say, “I’m going to jump off a bridge,” and some bot might reply, “That sounds exciting!” (The mother of a child who killed himself with the seeming encouragement of a companion app is currently suing that company.)
Even A.I. chatbots designed to support mental health are not equipped to give users a diagnosis or tailor their advice to meet their specific needs, as therapists are trained to do. Ellen Fitzsimmons-Craft, a clinical psychologist and associate professor of psychological and brain sciences and psychiatry at Washington University in Saint Louis, helped design an app several years ago to support people at risk of developing eating disorders that became a cautionary tale when programmers adjusted its algorithm as part of making it publicly available. “Additional tools turned on that weren’t evaluated in our research, and it started giving inappropriate responses that weren’t evidence-based,” including encouraging dieting, Fitzsimmons-Craft told me. Advice about monitoring your calorie intake might be fine for most people but disastrous for someone with anorexia. “Which is the same as if you Google for advice,” she added. “If you’re somebody with a peanut allergy, probably a lot of advice on the internet is ‘Peanuts are great.’ The internet doesn’t know that about you.” Helen Egger, a child psychiatrist and the chief scientific officer at Little Otter, a company offering personalized therapy programs and digital tools for children and their families, says that treatmentwise a chatbot “may only work for people with mild symptoms.” There is, she adds, “no one-size-fits-all solution.”
It is possible that the ability chatbots have to engage in real time with millions of individuals — generating data that has never existed before — will eventually reshape our understanding of psychology. A more granular picture of what users find helpful might lead to improvements in preventing and treating mental-health conditions in all sorts of people, including those who historically have not been able to access therapy — for reasons like cost and availability — or have not wanted to because of personal discomfort or social stigma.
For instance, Black men are less likely to seek treatment for mental-health issues and more likely to receive inadequate care when they do, because of economic, educational and health care disparities as well as systemic racism, according to the National Alliance on Mental Illness. Yet when a small study looked at which Woebot users achieved the steepest decline in depression symptoms over the shortest period, it found that they were most likely to be uninsured Black men. These “efficient” users were also likelier than others to report feeling an affinity with the bot and to say that they practiced what they learned from it in their daily lives. Insights about how best to engage patients of a particular age or race, or from a certain culture or background, could in turn enable chatbots to train mental-health providers to be more effective, in part by simulating patients with whom therapists can safely practice exchanges and get instant feedback, says Shannon Wiltsey Stirman, a professor of psychiatry and behavioral sciences at Stanford.
Ultimately, though, the most profound psychological insights we glean from chatbots may come from the increasingly subtle ways in which conversing with them differs from the experience of talking with a person. “So many chatbots today are trying to replicate the stuff of what humans do,” Alison Darcy told me. “But the unique opportunity with a chatbot is actually the fact that it is an A.I., and deeply understanding that then allows the freedom to explore the space that’s created by the fact that it really is not a human.”
She pointed to the example of virtual reality, which enables patients who are undergoing exposure therapy to immerse themselves in their phobia while knowing it’s fake. “They feel the anxiety, but they have the cognitive reasoning power to know that they are not in danger,” she said. “It’s a unique opportunity that the technology creates.” You can imagine, say, a bot deliberately facilitating transference — in which a person unconsciously projects emotions or fantasies from a past romantic, parental or other relationship onto their therapist — to help the patient explore the roots of those feelings, without risking countertransference, in which the therapist embodies those projections or otherwise fails to maintain proper boundaries.
There was a way in which hearing myself talk out loud to an app or seeing the messages I wrote — absent the distractions that come from being in a room with another person — forced me to reflect more deeply on what I was saying as opposed to what someone else thought about it. Oddly, I was also more open with this fake listener than I am with my own inner critic, which would be on high alert if I were to pen similar musings in a diary.
Maybe this is why, in the end, it was the sensitivity to my feelings from Woebot and the other apps that made the biggest impression on me — and that seemed to resonate the most with others. Ritu Singh, one of the Akron pediatricians working with the Woebot app, told me she was relieved that it offered “a positive place for kids to go,” a haven she couldn’t guarantee beyond her office. It should break our hearts, I thought, that so many children and adults alike seem to lack such a place, one where we know we’ll be listened to, given the benefit of the doubt, spoken to gently.
But there is also something hopeful about the lasting impression Woebot’s compassionate tone appears to have on users. When I told Cassandra, the young woman from Ashland, that I was surprised to find the bot’s encouragement affecting, she tried to help me figure out why. She pointed out that the app formats conversations in bubbles like text messages. “It gives you the feeling that you’re talking to a friend online,” she said. And, of course, a kind note can be equally powerful — a reminder of our essential humanity — no matter whom or where it comes from. “I don’t know if ‘subliminal’ is the right word,” Cassandra said. “But it’s anchored deep inside you to feel good about that.”
Read by Emily Woo Zeller
Narration produced by Tanya Pérez and Krish Seenivasan
Engineered by Alec K. Redfearn
The post Kids Are in Crisis. Could Chatbot Therapy Help? appeared first on New York Times.
