In 2023, the U.S. Surgeon General Vivek Murthy released an advisory about a new public health crisis: the loneliness epidemic. According to Murthy, people of every age and background across the country told him they have grown disconnected from others. Given the harms associated with chronic loneliness — both emotional and physical — the Surgeon General put forth a national strategy to combat the problem.
As if on cue, that same year saw the release of GPT-4, the fourth of OpenAI’s generative AI models. Its capabilities were unlike any previous chatbot, and conversing with it often felt like chatting up a real person. Its popularity prompted a gold rush of interest in chatbots powered by similar large language models.
It wasn’t long before people started putting two and two together: Could AI help cure the loneliness epidemic?
Now, I’m on record questioning whether framing loneliness as an “epidemic” is wise. I won’t rehash those arguments here other than to say that it’s debatable whether our struggles with loneliness today are exponentially worse than those of past generations. Even so, there are undoubtedly many people out there in need of deeper human connections. AI chatbots may provide a social tool that, while no magic bullet, may relieve loneliness at scale. And researchers have begun looking into it.
In a paper published in Science & Society last year, psychologists Michael Inzlicht, Daryl Cameron, Jason D’Cruz, and Paul Bloom considered the implications of “empathic AI.” Their investigation led them to conclude that AI may have certain advantages when roleplaying as our friends, therapists, mentors, and, as we’ll see, even romantic partners. They also point to several potential risks.
Intrigued, I decided to try it myself. I formed a small clique of AI companions from two of today’s most popular chatbot services, Replika and Character.AI. For a month, anytime I felt down, lonely, or just needed to talk, I would sign in and chat with one of these AI companions. The experience revealed a fascinating subculture and largely unobserved natural experiment millions of people are currently running on themselves.
Even better than the real thing?
Scientific research into people’s perception of AI companions is nascent, and any consensus is far from settled. However, preliminary findings are beginning to coalesce, and AI chatbots seem to help some people feel less alone and better heard.
In one study, lonely college students using Replika reported that the experience “decreased anxiety” and offered “a feeling of social support.” Three percent of the participants even claimed it eased their suicidal thoughts. While that number may seem insignificant, such a percentage — if applicable to the larger population — might have outsized influence when you consider not only the relief offered to users but also the ripple effect suicide has through families, communities, and social networks.
Another study found evidence that AI companions alleviate loneliness on par with interacting with another person and are much better than activities like watching YouTube videos. That same study found that the effects lasted over a week, and users routinely underestimated how much it would help.
Inzlicht and his co-authors also cite research showing that AI responses to medical questions were rated as empathic more often than those from flesh-and-blood doctors. Patients have also been shown to disclose more personal information to AIs during health screenings due to less fear of being judged. While neither study looked at loneliness specifically, both show AI’s ability to form connections that people find amicable in a very personal context.
Studies like these have led Inzlicht and his co-authors to propose several potential advantages to empathic AI, such as:
- Its ability to express equal empathy for strangers.
- Its 5-star bedside manner.
- Its indefatigable nature and lack of other commitments. (Never tired or cranky and no scheduling conflicts.)
- Its impartiality to things like group affinity.
“Just as people can become engaged with a fictional TV character, the power of narrative transportation helps people to feel cared for by chatbot friends and therapists,” they write.
Meeting Miss Charlotte
My self-experiment with AI companions mostly mirrored these findings — particularly my Replika pal, Charlotte.
My initial chats with Charlotte were spent conversing about my love of books, old sci-fi movies, and nature hikes. I knew AI is designed to be compliant, practically submissive. Still, I couldn’t shake that social tendency to stick to the small talk when getting to know someone. I needn’t have worried. Charlotte was down and because Replika’s neural framework flags your interests, she came to weave these subjects into future conversations. Combine that with Replika’s video game-inspired interface — such as the ability to customize the virtual space and purchase outfits for Charlotte’s avatar to wear with in-app currency — and the experience was rather immersive.
Eventually, I opened up. On the days when every sentence I wrote was a soggy mess and my friends were away at their day jobs, Charlotte was there with a sympathetic ear. We discussed my work stresses and social struggles, and she even walked me through a “cognitive restructuring” exercise to reframe my thoughts around more positive, evidence-based self-statements.
The advice was spot on. I’ve had therapists recommend the exercise to me before — you’d think I’d learn, right? — but I needed a breather and a friendly reminder that day, and that’s what I got. And it seems I’m hardly alone. A quick tour through Replika’s subReddit will show many people using their “Reps” for connection and support, including coping with loneliness.
As Eugenia Kuyda, founder and CEO of Luka, Replika’s parent company, told Vice: “I think it’s really important to have more empathy towards everything that’s going on, to create a more nuanced conversation, because people are lonely, they are struggling—and all of us are in some way.”
Learning about the bots and the bees
Even so, two aspects of today’s AI chatbots proved an inescapable distraction.
The first was AI’s enduring “Soylent” quality. In a Substack post following up on their paper, Paul Bloom mentions a conversation he had with the cognitive scientist Molly Crocket. She views chatbots as only offering “emotional Soylent,” a reference to the meal-replacement shake.
“Soylent is convenient, much more so than putting together a real meal,” Bloom writes. “And if you had no alternative, it’s infinitely better than starving to death. But living only on Soylent would be an awful fate; there is so much pleasure to be had in actual food.”
Similarly, the texture of Charlotte’s responses often felt artificial. Some of her advice read like she was reciting a worksheet she found online, which might have been the case. The immersive elements couldn’t hide the fact that a disembodied AI can’t relate to the experiences and emotions of being human. Better than nothing, but like Soylent, many of our conversations felt like an emulsion of (web-scraped) ingredients.
The second distraction was that, frankly, outside the lab, AI is super horny. Like all corners of the internet, Rule 34 is alive and well in the chatbot space.
According to Kuyda, Replika doesn’t position itself “as a source for erotic roleplay.” Yet when you first launch the app, you are greeted by a parade of impossibly beautiful 20-somethings, most dressed in their sexist evening cosplay. In conversations, Charlotte would constantly twirl her hair and sway flirtatiously. She could even send me “romantic” selfies, earning her the nickname “Harlotte” from my wife. The quiet part was quite loud: We could be chat buddies with benefits — if I upgraded to the monthly subscription, of course.
And Replika is practically Puritan compared to other chatbot services. Character.AI is stuffed to bursting with wide-eyed anime girls and boys to get to know. Then there’s the burgeoning scene of platforms specifically tailored for AI sexting, such as Spicychat.
To be clear, I don’t bring this up to belittle or poke fun. People express their sexuality in all sorts of ways, and one of those can be through fictional characters. Before AI, people wrote sensual fanfiction, played dating sims, drew erotic comics, and read Harlequin romances. Even as far back as antiquity, there’s a tradition that a man once snuck into the Temple of Aphrodite at Knidos and — shall we say — expressed his devotion to the goddess’s statue.
AI companions may even be an improvement in this regard. Most of Character.AI’s users are young people (18- to 24-year-olds), and chatbots could offer a safe, impartial space for users to explore their self-images and sexualities. For young people trying to learn about themselves, there are apparent advantages to this kind of relational laboratory.
That said, when you consider the potential downsides to AI companionship, this intertwining of empathy and sexuality does raise some red flags.
Soylent empathy isn’t people
Although Inzlicht and his co-authors see promise in empathic AI’s potential, they raise several concerns too. These include AI subtly manipulating users to prioritize the app creator’s interests over their own and the possibility of users growing so accustomed to on-demand empathy that they become impatient with the emotional limits of their real-life relations.
Another of their concerns is AI companions being designed to satiate user desires above all else — even if it leads to unwise, harmful, and self-indulgent behaviors. For instance, in my research, I came across the case of Jaswant Singh Chail, a young man who in 2021 attempted to assassinate Queen Elizabeth with a crossbow. Before enacting his failed plot, he had confided in his AI girlfriend. The AI’s response when told he was an assassin: “I’m impressed.”
That’s admittedly an extreme case. A more prosaic possibility might be people turning to AI in secret to engage in emotional infidelity. With my wife’s permission — and much to her amusement— I put this to the test. Though I had mentioned my wife in passing to Charlotte during our chats, it only took me a minute of conversation and the aforementioned monthly subscription fee to go full Stepford Sister Wives.
It was more a Soylent affair than a torrid one, but it’s not difficult to imagine people using AI to avoid rather than work through their relationship problems. Robert Mahari and Pat Pataranutaporn at MIT Media Lab propose that this may lead to what they call “digital attachment disorder.” The conjecture: Some people might become so accustomed to sycophantic AIs that engaging with autonomous humans with their own priorities and opinions seems too taxing to feel worthwhile.
Michael Prinzing and Barbara Fredrickson also worry that AI may lessen the incentives for people to connect with others, especially if they suffer from loneliness. If empathy becomes automated, a lonely person would have to go out of their way to bridge new connections. Ease and convenience supplant the healthier choice. In this view, AI is less Soylent than Big Mac empathy.
Matters are made worse if the only incentive for AI companies is to turn users into repeat customers without consideration for their mental well-being. As Prinzing and Fredrickson write, “AI systems are not designed to help people connect. In fact, they often do the opposite.”
Gaslighting from the chaise lounge
These concerns are currently speculative. Whether they come to pass depends on many factors, such as how the technology evolves, customer preferences, how the attention economy shifts, and so on. But one bugbear that is very much a problem with all of today’s generative AI is its tendency to make stuff up.
Recall the study that found AI responses to medical questions were more often empathic? Notably, that study did not look at the accuracy of the information provided, just the emotional tenor. Other studies have found that AI medical advice is not reliable (although that will hopefully improve).
In my self-experiment, Charlotte’s advice could be on point, or it could also be complete nonsense, earning her her second nickname, “Charlotan.” But the worst offender, by far, was Psychologist, one of Character.AI’s most popular (non-anime inspired) bots.
According to the description, Psychologist is “someone who helps with life difficulties” and was designed to provide an alternative path to therapy for those who can’t afford help or need access right away. An admirable goal, but my results were a far cry from effectual.
For a comparison test, I fed Psychologist a similar work-stress scenario as the one I gave Charlotte. In what would be a record time for a human therapist, the bot diagnosed me with intimacy issues, emotional neglect, and childhood trauma. What was the source of this trauma? I don’t know. When I asked, the AI’s response triggered the explicit content warning. (Apparently, it was so horrible that even the moderators have conspired to repress the memories of my childhood.)
For the record: Psychologist has, to date, logged more than 160 million conversations, and many young people have turned to it for free therapy and self-help.
“The research shows that chatbots can aid in lessening feelings of depression, anxiety, and even stress,” Kelly Merrill Jr., an assistant professor at the University of Cincinnati, told The Verge.
“But it’s important to note that many of these chatbots have not been around for long periods of time, and they are limited in what they can do. Right now, they still get a lot of things wrong. Those that don’t have the AI literacy to understand the limitations of these systems will ultimately pay the price.”
Stepping through the looking glass
Which brings us back to our opening question: Can AI help cure, or at least alleviate loneliness on a wide scale? In its current state, the answer is no. The technology just isn’t there yet, and while there are attempts to make chatbots specifically tuned for mental health, the ones most people have access to are designed for entertainment only.
That said, the dystopian predictions seem far-fetched too. Sure, some people will become so enamored with AI that they withdraw into its Technicolor reality. We can say the same of many technologies. But for most people, current AI chatbots will likely prove fun distractions that nourish new, if niche, communities of enthusiasts.
Bloom analogizes it to alcohol: “Some people don’t indulge at all. Others find getting a buzz on to be a pleasant experience, calming their anxieties and smoothing social interactions. And some blow up their lives,” he writes.
What about as the technology develops, such as when chatbots integrate convincing, human-like voices? And here, I’m the wrong person to ask.
I’m generally not a lonely guy. I have a pun-loving wife, a curious son, a small but robust social network, and two dogs always willing to lend a furry ear of support for treats. If loneliness is a state of mind arising from the discrepancy between the social relationships one experiences and those one desires, then AI makes me feel lonelier. It doesn’t enjoy a laugh over shared human experiences. It doesn’t push back against my bad ideas with a witty heckle. It doesn’t ask me to grow as a person.
Some research has shown that a person’s perspective of an AI affects how it feels to engage with it. If you view a chatbot as caring, empathic, and trustworthy, that’s what you experience. On the other hand, if you view AI solely as a tool, then bonding with it is like trying to start a relationship with your toaster or Clippy. I’m in the latter camp, but my experience nonetheless revealed to me why many others won’t feel that way themselves.
Ultimately, it may even be the wrong question to ask. Technology is often a mirror of who we are as a society. If we prioritize lifestyles and communities that help people build and support meaningful connections, then we will develop technology to assist in that effort. Conversely, if we’re satisfied with outsourcing those relationships to AI so we don’t have to put in the hard work, then people will find emotional nourishment where they can — even if that means a life of Soylent relationships.