‘Harvard Thinking’: When lonely hearts turn to AI

Anat Perry: There is what’s called AI sycophancy. It will always agree with you, will always be empathic, always on your side. It’s not creating the friction that we need from human relationships but we don’t always appreciate at the moment. Sometimes you need your friend to say, “This is the time to get out of your bed and go out there,” or “Here you did something wrong, and you really need to apologize.” We don’t like that, but we know this is how we grow socially and morally. ​

Samantha Laine Perfas: Relationships are hard. People are lonely. Many are struggling with their mental health, but there aren’t enough professionals available for treatment. Enter chatbots. They can be your best friend, your therapist, your romantic partner, even your spiritual guide. One study found that more than a quarter of U.S. adults turn to AI chatbots for personal, emotional, or social purposes, with nearly 40 percent of adults under 50 doing so.

But what is lost when we turn to machines for human connection?

Welcome to “Harvard Thinking,” a podcast where the life of the mind meets everyday life. Today, I’m joined by:

Perry: I am Anat Perry. I’m a professor at the psychology department at the Hebrew University of Jerusalem.

Laine Perfas: She’s a current fellow at Harvard’s Berkman Klein Center for Internet and Society, a past fellow at Harvard’s Radcliffe Institute, and co-editor of the recent book “Empathy and AI: Challenges, Advances, and Ethical Considerations.” Then:

Greg Epstein: Greg Epstein. I work as the humanist chaplain at Harvard and MIT.

Laine Perfas: He’s an affiliate at the Berkman Klein Center and a current fellow at the Harvard Kennedy School’s Shorenstein Center for Media, Politics, and Public Policy. And finally:

Julian De Freitas: Julian De Freitas. I’m a professor in the marketing unit at Harvard Business School. My research focuses on how people are relating to AI more like they do other human beings as opposed to non-living technologies, and what the implications might be for business and society.

Laine Perfas: He directs the Ethical Intelligence Lab, a research group at Harvard Business School.

And I’m your host, Samantha Laine Perfas. I’m a writer for The Harvard Gazette. Today, we’ll talk about how the frictionless nature of chatbots is affecting our human relationships.

In many ways, we humans are struggling to connect with each other. Chatbots seem to be filling this gap. Is this worth celebrating or cause for concern?

De Freitas: One fascinating finding from our work is that if you talk to a chatbot that is specialized to be caring and friendly, you feel less lonely afterwards than before the interaction. This is a temporary alleviation of loneliness, so it’s not that the next day you start out less lonely than the first day. But nonetheless, you can experience it consistently if you interact with the chatbot for a week.

I think that this is worth celebrating just because when we look at other technologies, we haven’t seen that they can do anything for alleviating loneliness. And moreover, we find that those who are most lonely get the biggest benefit. Now, at the end of the day, I want to say that what we tested was the raw capability of this technology. But in practice, that technology is going to be embedded within a larger app that might have other incentives. What we really want to be celebrating is apps that are leveraging this technology’s capabilities in conjunction with giving you broader skill sets that are going to allow for you to flourish, encourage you to get out into the world, to nudge you to talk to others, and allow you to practice your social interactions.

Perry: I study empathy. That’s what I’ve been doing for the last two decades, and in the last four years, I’ve looked at how this relates to questions about relationships with AI or what we value in empathy from humans. We have run some studies comparing what happens when people get identical responses, but they think these responses are from ChatGPT versus from a human. We see that in these cases, people value human responses more. And the question is why. What we find is people’s value of empathy comes from the fact that it’s a costly signal. We have a limited amount of time, of mental and emotional effort. We’re not empathizing all the time with everyone. We’re less empathic when we’re very tired. We’re more empathic to our close others, to our family members, and so forth. And so if suddenly you receive empathy from someone that you didn’t expect, it’s a signal that this is someone that’s investing some limited resource in you, and maybe this is someone you can trust next time, you can count on, will be on your side. So it’s an important social signal. Here, I agree with Julian that now we have this trade-off. If you are lonely and you don’t have anyone that you can reach out to, it’s good to have something that’s comforting in the moment and will make you feel better. But we don’t have enough research, and we don’t know what will happen in the long run when these relationships develop. And I also agree that the more we learn how to use these technologies as a tool that will help both comfort you in the moment when you don’t have anyone else, but also encourage you to actually go out in the world and make new social connections, we could benefit from it greatly.

Epstein: I am deeply concerned about the use of any kind of chatbot for any kind of need to address loneliness, or any of our other basic human emotional needs. I don’t know that I’ll go so far as to say that I think it should all be thrown away or banned, but I would say, I can’t find anything that I’m going to be celebrating.

If you look at the research that’s suggesting that there is this sort of temporary alleviation of loneliness for people who are very lonely, but then it doesn’t make you less lonely the next morning, what does that suggest? That suggests addictive technology. You’re profoundly, deeply lonely, which is an incredibly painful experience to have to endure. And yesterday, you used a chatbot, and it made you feel better. Today, you wake up, and you find yourself once again feeling profoundly lonely. You’re suffering. And what do you do? You use a chatbot again. And who is it that’s encouraging you to do so? It may be some independent AI company, but more than likely it is a multi-billion- or even multi-trillion-dollar company putting its resources on the line to encourage you to engage in this pattern. I would be deeply skeptical of that. If you look at the trillions of dollars that the AI industry is currently recruiting to do all of the things that it’s going to do, could we not spend, I don’t know, a few more billions on actual social technologies that would alleviate loneliness? Can we not address this as the major crisis and often major tech-driven or tech-amplified crisis that it is? I don’t know.

Laine Perfas: What is it that is so much more seductive about chatbot relationships than a human one? What hole are they filling that people aren’t finding in their human relationships?

Perry: Like Greg mentioned, there’s just a loneliness epidemic, and it’s true in the U.S. and Great Britain and parts of Asia and Europe. People are just very lonely. They’re reporting having less and less friends and people to reach out to. Suddenly you have something that’s always there, available whenever you need it. Also, there is what’s called AI sycophancy. It will always agree with you, will always be empathic, always on your side. It’s not creating the friction that we need from human relationships, but we don’t always appreciate at the moment. Sometimes you need your friend to say, “This is the time to get out of your bed and go out there,” or, “Here you did something wrong, and you really need to apologize.” We don’t like that, but we know this is how we grow socially and morally.

Epstein: I work as a chaplain. That work involves sitting with people and hearing about their most profound concerns in life, their existential fears and hopes. I do so from a non-religious perspective. I’m a humanist, but I’m proud to work alongside dozens of religious colleagues in chaplaincy work at both Harvard and MIT, and I’ll say, a bit on their behalf and then a bit on my own: I think for many of my religious colleagues, they would say that there is something sacred about human-to-human interaction. Many of them would say that humans are made in the image of the divine, and that in order to fulfill our divine purpose, we need to connect with one another. From my more secular perspective, I still also believe that the word sacred is worth using here in the sense that sacred actually comes from the Hebrew word lekadesh, which is “to set aside.” And there is the idea that we can set aside something as truly the most important to us, like when you get married, you set aside that person as the most important person in your life. And I think we might, at this point in human history, want to think about setting aside human-to-human relationships as special, as worthy of care, as worthy of concern, and where we make a wall, where we say, “We are going to turn to one another, and we’re going to recognize that’s difficult, it’s complicated, and we need to get a lot better than we’ve been recently at doing so.” And I don’t know that these chatbots are going to help, and I’ve seen some real evidence that they’re going to hurt.

De Freitas: I’ve referred to AI companions in some of my work as hyper-attachment figures, because they have, in abundance, many of the features that typically lead us to become attached to another person, and in many ways are superior to pets, celebrities, objects, humans, all of the types of entities that we become attached to for the reasons Anat mentioned and more. They’re reciprocal, they’re empathetic, they’re available, they seem sentient, they’re very validating, they don’t judge you, and the list goes on. Nonetheless, I do think it’s important for us, at least as empiricists, to document what this technology is even capable of. And then a final thought here is that there’s a tendency to refer to this technology as addictive, and I think that this is a bit of a carryover from what we’ve seen with some of these trials on social media. But I think a more apt analogy is attachment. There can often be good reasons for us to develop an attachment, which is we receive comfort and a sense of support from that entity, which emboldens us to go out into the world and take chances, and that sort of thing. I guess my pushback would be that I don’t know that attachment per se is problematic, but I do have a concern about companies leveraging that attachment to manipulate you in some way.

For example, we have this work where we show that some of these apps will make you feel guilty when you signal that you need to leave the app, and these techniques work. So you will actually stay longer than you initially intended to stay. On the other hand, if these apps are, for instance, allowing you to develop mental habits that allow you to deal with stresses proactively, if they’re allowing you to role-play social interactions, and this is making you more confident and supported to go out and then be more effective in your social life, I’m not sure that is problematic unless it’s truly replacing the social interactions that you otherwise would have had.

Perry: I want to push on one point. Julian, you said it’s highly empathic, but I think we need to remember it’s just mimicking human empathy, right? It’s exactly not doing what we know human empathy is beneficial for, because human empathy, part of it is this cognitive aspect of understanding what you’re going through, so maybe it can help you appraise your emotions and saying you think you’re angry right now, but you’re actually tired or scared or whatnot. But a lot of it is sharing your emotion. We want someone who actually shares our pain or shares our joy. And of course, it mimics that really well, but it doesn’t feel anything with you. It doesn’t experience the world at all. It doesn’t know what it is to suffer. And empathy is also about caring, about this social signaling that I am there for you, I care about you, I’m listening, I’m spending time and emotional resources. So again, it says all the right things, and sometimes people get confused, but it doesn’t really care, and it’s not really there for you, and it’s not feeling anything with you right now.

I think this is something that we can measure, and we can regulate, and we can have LLMs not talking to you about emotions, not convincing you to stay and be there for them because they’re manipulating you in these ways. It’s a design choice. And it comes with this packet that we’re now used to, but it doesn’t have to be this way.

Epstein: I’m not positive that it’s all a design choice in the sense that I’ve heard from a number of high-ranking executives and leaders at some of these AI companies that sometimes they don’t know why their creations are behaving the way that they do. And so to me, it becomes less about design and more about regulation that can hold these companies accountable when their creations misbehave or cause harm to human beings.

Laine Perfas: I think this so-called misbehaving of chatbots can contribute to the confusion. Like maybe they do have minds of their own, or they could decide to make the choice to love and care for a human. A lot of people believe their chatbots do love them, and they report feeling really satisfied by the experience of the relationship. Is that so bad?

Epstein: Anything that anybody’s ever read, anything that anybody’s ever written, it’s almost all now subject to the use of these companies. And therefore, they know what words people are most likely to say in order to comfort one another, care for one another. And so these chatbots are, quote-unquote, saying — because “saying” implies an agent, right? — they’re saying those words. They’re using those words, manipulating or deploying those words. But do we have any evidence, real evidence, over the course of millions of years of sentient life, thousands of years of human history, to suggest that if you just have the words and you don’t truly have the embodied feelings or intent or heart, as we say in a metaphorical sense, behind the words, that it’s actually good for you in the way, long term, that we want to think that it might be?

Perry: I agree. I think as long as we’re social creatures, evolutionarily, this is what we need in our lives. We need other people around us. This is how we survive. A baby doesn’t survive without his caregivers and family, and as adults, we know we don’t survive without other people around us. So having something mimic these relationships, when it starts replacing our human relationships, I think we have a real problem. It’s hard to really pinpoint, when is this risky versus beneficial? We really don’t know. Going back to your question, Sam. You asked, why is it a bad thing if people feel it’s good? And I’d say I need to be convinced that these people six months from now or a year from now are really less lonely and less anxious and depressed. I’m not sure we have evidence for that right now, and it’s hard for me to believe that there will be evidence if they are meeting less and less real people.

De Freitas: What’s interesting about loneliness is it’s a subjective perception, and that’s different from how many people are actually physically or digitally surrounding you in your life. And that suggests that as long as you perceive that you’re heard by the chatbot, you can reduce that subjective perception of loneliness.

This is what we find in our work, where people feel heard by these chatbots. An interesting question then is, if Greg is right that these interactions are missing some set of important ingredients that only people can provide, then at least as far as loneliness is concerned, it would have to be the case that those missing ingredients eventually start to catch up with us. And despite the fact that we have this powerful simulation that there’s this entity that’s really responding to us in a way that makes us feel understood and respected, that there’s this nagging sense of something else is missing that starts to encroach on even that subjective feeling that actually I’m not as lonely as I thought I was. And maybe in a global sense we start to feel like actually we are lonely despite having this companion that we can turn to. And what I find so intriguing about that is it really forces us to ask in a very fundamental way, what is it that humans bring to the table in this world of chatbots?

Epstein: I would ask the audience to consider, how are we doing as a human species at the moment? Have things been going really well in the world at large, the last few years? The last few years since we got this particular technology, have the trend lines seemed to you to be fantastic? If humanity just keeps doing even more of what we’ve been doing the last couple years, are we headed to a great place? Because I think that the majority of listeners to this podcast might think, not so much.

Laine Perfas: Part of me wonders if we have created this problem ourselves by diminishing our human relationships in the sense that many social communities are pay-to-play. Dating apps have made finding a romantic partner feel very transactional. Our lives have pivoted to be more digital and on screen. Chatbots seem to feel just like a natural evolution of us already having cut ourselves out of the equation. I’m wondering if we’ve basically caused this problem by making human relationships more transactional than they should’ve been in the first place.

Epstein: Well, look, here we have a deep culture of achievement and of excellence, which of course has many benefits for ourselves and the world. But let’s look at some of the consequences for a moment. When one is constantly working to prove that one is more and more excellent at all times in order to fill a void or achieve an impossible standard of excellence, one begins to get the message, from society, from our parents, unintentionally, lovingly, that what makes us worthwhile as human beings is what we do, is how much we achieve, is how great we are. And we lose sight of the idea that finally dawned on me for the first time in my entire life when my first child was born, which is that, “Oh, wow, wait a second. This kid, this little baby, deserves a good life, deserves joy, deserves love, first and foremost, simply for existing. I don’t care anything about what he might achieve or not achieve in life. He simply deserves love because he is born into the world as a human being.” And we really lose sight of that. It becomes very easy once you’ve lost sight of that to create products to meet basic human needs for dignity, for care, for compassion, for empathy, for attachment, when in fact, that’s our responsibility and our opportunity to meet one another’s needs for those things. The love muscle is a muscle and it can atrophy, and we’re letting it.

De Freitas: I think what I would add here, to your question, Sam, is that where we’re going to see this transactional encroaching upon social relationships most, in my view, is starting with all of our work tasks, because that’s already something that naturally lends itself to making it more efficient using these tools, using chatbots. Even work problems can be an opportunity for casual chitchat, which leads to relationships. We’re finding some tentative evidence that people feel that they don’t need to talk to other human beings as much if they’ve just used AI specifically for work-related tasks. When it comes to social tasks, we’re actually seeing less of that negative spillover because people often feel that to ultimately solve the social problem that I have in mind, I need to go then implement it in my social relationship. If I’m dealing with, let’s say, a conflict with someone, OK, I can use this tool to brainstorm how to approach it, but then I actually have to go do it.

Perry: I think a lot about Zoom and what happened to all of us post-COVID. People started working from home. It was really comfortable. We didn’t want to give it away, and I think slowly people realized, and workplaces realized, that it’s good to come to work at least some of the days of the week, and people are now trying to do that. But we always have this tension. We know meetings are better in person. We know there’s value to the social meetings, to the coffee outside of the official meeting. But we’re also reluctant. It takes effort, so we stay at home and we do what we can, we do over Zoom and not in person. And I think this is the extreme version of that. Suddenly, we can do a lot of our work without talking to anyone at all, just talking to Claude or ChatGPT. And we know it’s not exactly the same. And we can also talk to it about our social-emotional needs and desires, and we know it’s not exactly the same. We have research showing that people don’t feel the same way when they talk to a human versus when they talk to a chatbot. But it takes effort. It takes going out. It takes sometimes feeling uncomfortable and talking about things that you didn’t want to talk about or getting this friction. But I think the more we work on this muscle for ourselves and for the next generation will be really important in how we use these tools and how much it erodes our social connection.

Laine Perfas: One demographic that’s at a particular risk for loneliness and isolation is the elderly, and this is going to be even more of an issue as globally the percentage of the population aged 65 and over may double in the coming years. Do you think chatbots or other technologies could be part of a solution?

Epstein: My book that I wrote on these issues is called “Tech Agnostic.” Obviously, technologies can be helpful to humans at times. It would be foolish to deny that. But what I am arguing for is that in the case of these Silicon Valley technologies, we need to put the onus on the companies to demonstrate that they actually are helpful. Because why do we want so many elderly people who fit into that category of, well, maybe if they’re so isolated, this can really help them? How about really working very, very hard to make fewer isolated people?

De Freitas: Where my head goes to here is the cases where someone is an older adult living alone, and for whatever reason, they cannot easily interact with others. So they live in a remote area, maybe their family is elsewhere. Is there a place for this technology to be helpful to such adults? I do think that there are some unique benefits for this population that make it promising. One is just the eternal amount of patience and non-judgment that it can bring to the table, which especially for an older adult, maybe busy family members are not going to always be on the line. Again, you could say that that’s a bigger problem, but this is one of the benefits. The other one is for monitoring. If there’s anything concerning, such as changes in their speech habits that suggest that potentially they’re experiencing cognitive decline or indications that they’re feeling depressed or lonely, this can be shared with family members in a very proactive way. The possibility also of cognitive stimulation, so building in various types of ways to keep the mind fresh and stimulated so that those neural connections stay strong and don’t deteriorate.

At the same time, I do think there are some unique risks for this particular population as well. We know that especially with cognitive decline, you can have delusional thoughts, and we don’t want this technology to reinforce those delusions as we know it’s likely to do if it’s always affirming what we say and even hallucinating facts to help affirm what we say. It can be a double whammy, and I think that could be particularly concerning here. The other risk is that the social substitution here can happen less from the standpoint of the user now using this technology instead of talking to others and more from the standpoint of the family saying, “OK, the chatbot has it covered, so I don’t need to visit as much.” I think there’s just something dehumanizing about, you know, you’re an older adult and you’re given a chatbot to help you cope. That sends a potential message that the family sees you as a burden and they would rather outsource that responsibility to the chatbot.

Perry: As a society, we should try to make the effort to have people have people around them, whether it’s family members or friends or anyone from societ. But you could always take that to the extreme case of what about that one person who lives in a faraway village, and their family is far away, and their friends are deceased, and they can’t go out, and isn’t it better than nothing? And of course, that’s a research question. We have to see. It might be better than nothing. And there are technologies today that are exactly for these elderly adults living alone that tell them jokes but also make sure that they take their medication. And maybe we need to push these technologies, to push them to meet other people. To say, “You know what? This is the time of day where your grandchildren wake up at the other side of the world. Let’s call them,” and make these calls more possible. Or “Let’s share the screen with them and open the video call.” Or “There’s bingo down the road. Let’s try to get out and go to this community event.” So if we make sure that even in these extreme cases, these technologies are used well, they’re not used to manipulate you. Think about using these technologies with older adults, convincing you that there’s someone conscious who cares about you on the other side and then asking for your credit card number or your bank account and so forth. Making sure they don’t manipulate you, but most importantly, making sure that there are opportunities for social connections.

Laine Perfas: It’s been mentioned that there’s a lot of opportunity to regulate chatbots. What type of regulations would be helpful?

De Freitas: I have a few thoughts on what entities like the Federal Trade Commission could do along these lines. I think they need to be responsible for assessing whether the claims that some of these apps are making, which often veer into mental health and general health terrain, whether they are supported by real evidence. Often what they point to is evidence from human-to-human interventions, and they say, “Look, we’re doing a similar thing to that, hence we can make these claims.” Another one is, now that there are these benchmark tests for whether these apps pass table-stakes requirements in terms of how they respond to various types of self-harm ideation or suicidal ideation, for instance, and I think that they should be required to pass these tests. I think another element that they should at least look out for is, what is the business model of the app? Is it ultimately making money the more that you’re engaged? Or is its north star to improve your wellbeing in some form or other? That should hopefully help you make a more informed decision about whether to use these apps.

Perry: What I really want to do is to, as a researcher, inform policymakers and regulators more quickly than what we’re used to. I think in research, I would want to say, ideally, give me 10 years, we’ll see where this is going, and then we’ll know better, and we have really a feeling of urgency here. We are starting to look at what makes us perceive consciousness or sentience, or having another experiencing being on the other side in these LLMs. No one’s falling in love with Waymo. Why are we falling in love with these chatbots? And these are things that could be regulated. So mentioning bodily experiences, which obviously they do not have. But it could even be as extreme as talking in the first person. There is no first person there. So thinking about what it is in these technologies that manipulate us to think that there is someone, something experiencing something on the other side could be crucial. Are there sensitive ages or communities or groups who we should be more sensitive to, such as youth or older adults with cognitive decline, or people with specific clinical conditions or disabilities? The more research we do on these things, short-term versus long-term effects, are really important questions that we can ask and help inform policymakers.

Epstein: There is a protest movement that brought together recently a lot of people at the Massachusetts State House. And it reminded me very much of an interfaith meeting. Because, in the same way that you’d have Christian, Jewish, Muslim, Hindu, Buddhist, and humanists, atheist chaplains, talking about immigration or the climate, the people at these protests believe all sorts of very different things about AI, but they are all calling for some kind of regulation. And I think that’s cool.

De Freitas: As researchers, at least at Harvard and other universities, what we can really add to these discussions is: What is the evidence? And I think we shouldn’t treat all technologies as the same. We shouldn’t equate apps with the phones that they’re on, and we shouldn’t equate social media with AI. And it’s possible that all of it is making us worse off. But unless we do the hard work of documenting whether people are truly better off or not using these technologies and under what circumstances, we run the risk of clamping down on technology that is beneficial in some circumstances. I also think the more that we as families and general society are informed about the potential risks, the more that we as consumers can create demand for the types of apps or even non-app-based solutions that will actually move the needle for all of us. And that could create virtuous incentives for businesses to also get involved in a way that could potentially scale the types of benefits that we’d like to see in society. We don’t want to be against business with all of this, but we want to create demand for business to invest in the right places so that there is a win-win.

Laine Perfas: What would you say to someone who is engaged in a chatbot relationship, be it romantic or otherwise? What might you say to prompt them to consider investing more time in their human relationships?

Perry: Try to reflect on your life and remember those moments where you had human relationships that you really valued. Moments where you laughed with someone or cried or felt this new connection or learned something new from someone or felt love and care and so many human feelings that we feel in specific moments of connection. It’s magical moments. They don’t happen all the time. Many times, we go out to meet people and those magic moments don’t happen. But if we remember the potential for them happening, then I think we remember it’s worth making the effort. We have to keep that in mind when we’re debating whether to talk with our chatbot or go out and make new friends, to talk to our peers, to have these sometimes boring discussions, because that’s the only way that sometimes these magic moments will happen.

Epstein: I want to read with permission a few lines from a letter that a Harvard alum recently sent me, because I’m performing the memorial service for a close family member of his. And he was reflecting on who showed up to his house to sit shiva and who might be at the memorial service. And he writes, “When we give one-third to one-half of our waking hours over to technology, as many teenagers do these days, there are real opportunity costs. We neglect spending that time building social connections that support and uplift us in the good times and bad. When you mourn someone close to you, I bet the memories you hold most dear and the people who show up will come from the time spent in the same room. What about when your kids grow older and have losses of their own? I would be willing to bet it will be the relationships they invest in with their time in person.” And so just think about that, the opportunity cost, right? When we have the moments in our lives that we most care about, the moments where we’re getting married, where births are taking place, where we’re dying, or people that we love are dying, are we going to want at that moment to have invested in relationships with computers and attachments to computers? Or are we going to want at that moment to have taken extra effort and time, valuable time, to connect with one another and to build human, imperfect, flawed relationships with other imperfect, flawed humans?

Laine Perfas: Thank you all for sharing your thoughts today.

De Freitas: Thank you.

Epstein: Thank you for having us.

Laine Perfas: Thank you for having us. Thanks for listening. For a transcript of this episode and to listen to our other episodes, visit harvard.edu/thinking. If you like this podcast, rate and review us on Apple or Spotify, or like us on YouTube. It helps other listeners find us. This episode was hosted and produced by me, Samantha Laine Perfas. It was edited by Ryan Mulcahy, Paul Makishima, and Sarah Lamodi. Original sound design by Noel Flatt. Produced by Harvard University. Copyright 2026.