Is a chatbot therapist better than nothing?

Recent reports suggest we are experiencing a loneliness epidemic and mental health crisis. Suicide rates in the U.S. have risen over the past two decades; a recent surgeon general’s report flagged troubling statistics on the well-being of American youth; and more than a billion people globally have a mental health condition, according to new World Health Organization data.

Technologies such as smartphones, social media, and chatbots inevitably come up in any discussion about mental health — often as forces for harm. But could these same tools be used to help?

In a recent symposium hosted by the Faculty of Arts and Sciences, four Harvard experts on depression, anxiety, and trauma — Karestan Koenen, Elizabeth Lunbeck, Matthew Nock, and Jordan Smoller — discuss the potential risks and rewards of using tools such as chatbots to narrow the gap between need and treatment access. Gazette senior science writer Alvin Powell moderated the following conversation, condensed for clarity and length. Watch the video for the full conversation.

---

Whenever mental health comes up, inevitably technology becomes part of the conversation. AI-powered chatbots are powerful tools, clearly. Several cases have grabbed headlines, though, of potentially negative impacts — they’ve even been accused of encouraging suicide. How closely do we need to monitor these chatbots for harmful interactions?

Nock: As I think about new technologies, I think about old technologies. A lot of the conversations that we’re having now about chatbots we had a few years ago about social media, we’ve had about TV and telephones, and so on. They are tools. For each of them, we’ve got to figure out how do we use them in a way that maximizes the good and that minimizes the harm.

Lunbeck: I appreciate the nod to history. I’ll bring up another data point. The first chatbot for mental health was the ELIZA machine, which was set up by Joseph Weizenbaum, a computer scientist at MIT, and he found immediately that people reacted to the chatbot as if it were a person. And there’s an apocryphal story of him — his secretary was using it in the interface and asked him to leave the room because she was having a private conversation with the chatbot. And even his colleagues who were computer scientists, he complained that they should have known better. There’s something very alluring about this technology that we don’t fully understand yet all these years later.

Smoller: Chatbots are increasingly good at tapping into some of the machinery that we have ourselves in terms of identifying a connection with something and then anthropomorphizing it.

Lunbeck: People talk about it — them hacking our attachment system.

There’s been a lot of focus on the sycophancy, the way that they are primed to agree with us, to mirror us. A recent study was just reported that the “yes” answers from ChatGPT outnumbered the “no” by 10 to 1.

Nock: But what are we comparing them to? How I learned to do therapy in the room as human to human, I say “yes” a lot more than I say “no.” I do a lot more validating. We’re trained, and the evidence suggests, there’s benefit to validating a person’s experience — “yes and…” I rarely, when I do therapy, say, “No, you’re wrong, and don’t do that.”

Lunbeck: Well, Freud said the analyst should be as a mirror to the patient. And now, we complain about chatbots mirroring, but that is part of what therapists do.

Nock: I think we should acknowledge the positive here. There are a lot of people who have mental disorders and don’t have the access to care. And here comes a technology that not everyone but a lot more people have access to. The possibility, the upside here is incredible.

It seems you are in agreement that there’s great promise here, but are these tools ready for prime time when we hear about these troublesome cases?

Lunbeck: Well, you can talk about suicide. Obviously, we need better guardrails. For a long time we’ve known that relationship is the key to effective psychotherapy. People have said, well, it’s not a relationship with a human, but I’m more focused on what’s the nature of that relationship people are forming with the chatbot? In what ways does it matter, in what ways doesn’t it matter that it’s not human? If it’s giving you something useful that you didn’t have before, who are we to say…?

Koenen: Is that better than nothing?

Lunbeck: If it’s preventing you from seeking help, that’s a problem. If it’s responding to your suicidal ideation with, “yay, go girl,” that’s terrible.

Nock: I see this, as in many areas of health and medicine, there’s an iteration that has to happen, and we think about: Are we ready? Is the treatment ready? There are people dying now. Let’s get the best treatment we can in the hands of people, test it out through the scientific process, see what works, and see if we can make it more effective.

The same approach is needed here. A lot of these platforms weren’t designed to provide therapy, but people are using them for a pseudo-therapeutic kind of relationship. I’d love to see more collaboration between industry and independent scientists to iteratively improve these technologies in ways that improve the health of society.

Lunbeck: I don’t want to lose sight of the dangers, though, because you did start with that. There has to be more attention to regulation and more pressure on the companies. It’s kind of a semantic question. What’s therapy, and what’s emotional support? There are ways to build in some guardrails — if you’ve been on it for five hours, for the bot to say, “maybe you should take a break” or something like that — and we just have to keep pressure on the companies to do that.

Would it be preferable if we boosted human connection?

Smoller: One of the things that we’ve studied is looking for the factors that are likely to be causally protective against developing things like depression. And over and over, two things come up: physical activity and social connection. During the pandemic, we had a study in which people who had good social and emotional support early in the pandemic had half the rate of developing significant depression.

So I think, absolutely. It’s a difficult thing to do. Our society is being structured in the opposite direction too much.

Koenen: And you’ve had studies, I know, where you’ve looked at people at high genetic risk of depression in the military, and found that those in units with high support, high conviction in their military units, even though they had high risk of depression genetically, had less new depression later. So even with high genetic risk, the social support and connection can buffer against that.

We’ve talked about large language models, but what about cellphones? Are there innovative things going on in that area?

Nock: We did a study at Harvard a few years back where we interviewed teens who were getting psychiatric treatment, and we asked them about their social media use and what is helpful, what is harmful. Everyone said there are aspects of it that I love: I connect with other people, I see what the new trends are, I learn skills that I can use for my mental health. And on the negative side, I do social comparison, and I feel lonely, and I feel like I don’t have enough, and kids bully me, and so on.

Kids are using social media. This is the world they’re living in. How do we find out how to maximize health and minimize harm?

Smoller: Social media is not going away. The numbers I’ve seen are that 75 percent of young people get mental health information from social media, and there’s room for raising the game there. I worry about that because there’s so much misinformation and investment in psychiatric illness as being an important part of people’s development, and a lot of expression of distress through the labels of psychiatric illness, often inappropriately. It’s another thing we have to keep our eye on because that’s where people are getting their information, and it’s not always good.

Lunbeck: I couldn’t agree more. It’s become a marker of identity. Through TikTok, there are whole communities around entities that aren’t in the Diagnostic and Statistical Manual of Mental Disorders, but that have been named by users. We have to think about other ways for people to express their emotional vulnerabilities and concerns.

Nock: We teamed up with an industry partner who developed an algorithm to find people on social media platforms who were in distress, and then we experimented with how do we get them to crisis services. And we found that through a little tweak, we could boost use of crisis services by about 25 percent.

I worked with another platform to find blog posts that people found helpful and inspiring, and then randomized people to get those versus not. And found that people who got these blog posts had a significant decrease in their suicidal thinking, and they felt more hopeful, more optimistic, more connected with others. There are ways to use platforms and new technology in ways that boost people’s mental health rather than decline it.

Koenen: We have a study where we’re doing an app-based treatment in Kenya for PTSD. I was a little skeptical but we’re finding positive effects. People’s PTSD symptoms are really going down with this mainly app-based treatment, and they meet occasionally with a facilitator.

That’s by people’s smartphones, not something we could have done a decade ago. And it’s reaching many more people than we could have reached if we were waiting for a therapist trained in the specific evidence-based treatment to see individual patients in their office.

How do you see mental health care changing in 10 years?

Smoller: The problem with mental health care that dwarfs most others is access to care. We have to solve that problem. People with serious mental illness, psychotic illness actually have a shortened lifespan of 10 to 15 years.

My hope also is that we’re going to advance into this possibility of more precise, personalized treatment that’s more effective. And, I hope, a much bigger focus on prevention and early detection because most of what we do now is reactive.

Lunbeck: For years, it’s been the case that about half of the people who seek psychotherapy do not meet the criteria for any diagnosable mental disease. They go for support, for advice, for companionship, for changing oneself, a whole range of reasons.

From where I sit in therapy communities, there’s a lot of worry, will therapists survive this, the new technology? Will there be a role for actual human-to-human interaction? And I’m a big proponent of things happening when people actually talk to each other.

Nock: Thinking about suicide, half of people who die by suicide saw a clinician in the four weeks before they died. So they’re getting to care. One of my biggest questions is — and this is a focus of the center that Jordan and I lead at Harvard MGH — how can we use new technologies to provide better round-the-clock care, reach people when they need it, not in the absence of humans, but in addition to their human connections?

Koenen: And I think the other piece that we need to think about is the brain doesn’t exist separate from the rest of the body. As Jordan mentioned, there is good evidence that exercise can prevent depression.

There’s recently been large clinical trials where they’ve compared a cognitive behavioral therapy — that’s the gold standard for people with PTSD — against trauma-informed yoga. Randomized. They had the similar effects on PTSD symptoms, and there was less dropout from the yoga.

I would have said you can’t recover from PTSD without talking about your trauma. There was no talk about the trauma in the trauma-informed yoga. So what’s going on there? I don’t know, but there is this place for other things besides talking.

Maybe it was the community of doing yoga in the group. I don’t want to trivialize people with psychotic disorders or suffering from severe mental illness and say that yoga is going to cure them, but I think there are a lot of exciting developments in mind and body that we don’t really understand that could also help address this.

So it may be that a future Dr. Chatbot tells you to shut off the chatbot and go for a run, preferably with a whole group of people.

Koenen: Exactly.