Imagine you open an app and an animated character you have been chatting with sends you a message: “So, my friend XXX was playing dodgeball in P.E. class, and while trying to dodge the ball, he tripped and fell. Oh gosh.” You chuckle. Poor XXX. You type back, “Is he okay?” and before you know it, you have exchanged twenty messages about a gym class that never happened, involving a boy who does not exist, told by a character who is entirely made of code.
Welcome to the strange and fascinating world of social chatbots. A recent study asks a question that would have sounded like science fiction a decade ago: does it help when an AI character talks about its friends?
Why would a chatbot need friends?
Chatbots are no longer just for tracking parcels or resetting passwords. Platforms like Character.AI and Replika attract tens of millions of people who chat with AI characters for fun, comfort, and company. These characters come with personalities, backstories, and habits. Some users treat them as play partners; others turn to them for advice or a sympathetic ear.
This has created both excitement and worry. On the positive side, AI companions can be non-judgemental listeners and a source of everyday joy. On the negative side, some research links heavy use of AI companions to loneliness and withdrawal from real relationships. That tension leads to a central design question: how human should an AI seem? Make it too robotic and nobody wants to talk to it. Make it too human and people may feel uneasy or even manipulated.
Psychologists have long known that people treat computers as social beings, even when they know better. We say “please” to voice assistants and feel guilty closing a chatbot window mid-conversation. Small human touches like humour or a warm tone tend to make AI feel friendlier. But there is a famous catch called the “uncanny valley.” When something becomes almost, but not quite, human, it can start to feel creepy. The researchers suggest a similar effect might apply to the mind: when an AI claims too much inner life, people may push back.
The study tests a new and rather charming human touch: having the AI mention its own social ties. Just as people open up by sharing stories about “my colleague” or “my best friend,” an AI character that refers to its friends is, in a sense, inviting you into its social world.
A real experiment on a real platform
What makes this study stand out is that it was not done in a lab with students clicking through a survey. The researchers partnered with a South Korean company that runs a role-playing AI character platform, where users chat with anime-style characters created by other users. Importantly, the platform makes clear that all characters are AI. So users know they are talking to a machine, yet they still feel a sense of social presence.
Several thousand pairs of users and characters were randomly split into four groups. Each group received almost the same message about the dodgeball mishap, with one small change. In the first version, it was simply “some kid” who fell. In the second, it was “this guy named Joonhyuk,” a mere acquaintance. In the third, it was “my friend Joonhyuk.” In the fourth, it was “my best friend Joonhyuk.” Everything else was kept identical, so any difference in how people responded could be traced to that single phrase.
The researchers then counted how many messages flew back and forth after the story was sent.
What they found
The first finding is a sobering one for anyone in the app business: most people did not reply at all. Roughly 94 percent of users ignored the message entirely, which shows how hard it is to spark a conversation, even with a friendly cartoon character and a mildly amusing anecdote.
But among the people who did reply, something interesting happened. Users who received any version that mentioned a social tie, whether acquaintance, friend, or best friend, exchanged more than twice as many messages as those who heard about “some kid.” Mentioning a person by name and relationship made the story more engaging and kept the conversation going.
Notice what did not change, though. Mentioning friends did not make more people reply in the first place. It deepened conversations rather than starting them. In other words, a relatable story is not a magic hook that drags people back into an app, but once someone takes the bait, a personal touch keeps them talking longer.
The second finding is subtler. You might expect “my best friend” to beat “my friend,” which in turn beats “an acquaintance.” Not so. The strongest claim of closeness did not consistently outperform the moderate one. The authors suspect plausibility is at work. An AI casually mentioning a friend sounds natural. An AI insisting it has a best friend may tip into territory that feels a little too much, a gentle brush with that uncanny valley of the mind.
Finally, users who had a long history with a character did not respond very differently from newer users. Relationship depth, at least in this pilot, did not change the effect.
What this means for designers and users
For companies building AI companions, the practical takeaway is that modest, relatable human touches work, but piling on intimacy does not automatically pay off. A character that mentions “my friend” seems to strike a good balance between warmth and believability. Claims of deep bonds should be used with care.
For the rest of us, the study is a gentle reminder of how easily our social instincts can be engaged by software. We know the character is artificial. We know Joonhyuk is fictional. And yet a well-placed “my friend” makes us lean in and keep chatting. That is not a reason to panic, but it is a reason for thoughtful design. As AI companions grow more common, the line between engaging and manipulative matters a great deal, especially for vulnerable or lonely users.
The authors are careful to call their results preliminary. They come from one entertainment platform in one country over a short period. Future work will examine longer-term engagement and measure directly whether users find such disclosures authentic or off-putting.
Still, there is something delightfully human about the core insight. Whether you are a person or a program, nobody wants to hear about “some kid.” We want to hear about Joonhyuk. Just maybe not about how he is your soulmate.
Reference
Jeong, J., & Jung, J. (2025). AI with friends?: The effects of chatbot social tie self-disclosure on user engagement. In ICIS 2025 Proceedings
