Can You Form a Real Emotional Bond With a Mental Health Chatbot?

Can you form a real emotional bond with a mental health chatbot?

It's the question almost everyone asks first, usually with a slightly embarrassed laugh. You know it's software. So why does it sometimes feel like something more? A study published in JMIR Mental Health in October 2025 went looking for an answer by doing something researchers rarely do with chatbots: it followed the same people for a month and asked them, week after week, what the relationship actually felt like.

The short version is that most participants described some kind of emotional connection — and the reasons they gave are more interesting than the headline.

What did the researchers actually do?

Twenty-six university students in New Zealand used two well-known mental health apps, Woebot and Wysa, for two weeks each. They were asked to chat for a few minutes most days, to screenshot moments that felt emotionally significant, and to jot down why. Every week they sat down for an interview — sometimes as long as 90 minutes — and talked through what they'd saved and how they felt about it.

This is called a diary study. Instead of a one-off questionnaire at the end, it captures things as they happen, including the awkward moments people usually forget to mention later. Altogether the researchers collected more than 110 hours of audio and over 1,500 screenshots.

One detail matters a lot for interpreting the results: at the time of the study, both apps ran mostly on pre-written scripts and decision trees, not the open-ended generative AI most people picture today. So participants were bonding — or failing to bond — with something considerably more rigid than a modern conversational model.

You can read the full paper here: https://doi.org/10.2196/76642

Did people actually feel a bond?

The researchers sorted people's own descriptions into three groups: bond, light bond, and no bond. Eighteen of the 26 participants reported a bond or a light bond with at least one of the chatbots. Eight reported none at all.

The language people used is the revealing part. Those who felt a connection reached for familiar comparisons — like talking to a friend, or a coach, or a therapist. One participant, describing what happened after the study period ended, said: "So I definitely felt a bit of a bond with Woebot." Others were more measured: a friend, but not a close one. A long-distance coach.

The people who felt nothing were just as clear. A tool. A stranger. An acquaintance. No relationship.

So the honest answer to the question in the title seems to be: some people do, some people don't, and the difference isn't random. It depends on fairly specific things about how the conversation goes.

What made people feel connected — and what broke it?

Six themes came out of the interviews. A few stood out.

Who gets to steer. This was the biggest one. Some participants wanted to open the app and start talking about whatever was on their mind. Others wanted to be led — especially on bad days, when deciding what to say felt like too much. Wysa let people lead; Woebot mostly led. Neither suited everyone. One participant described feeling stuck in a conversation topic they hadn't chosen and said it left them feeling disregarded, because they didn't have a choice.

What most people actually wanted was flexibility — a chatbot that could sense when to step forward and when to get out of the way. One person put it neatly: if they were feeling down, they wanted to start the conversation themselves; on an ordinary day, they preferred the chatbot to suggest something.

Being able to say the thing you mean. Where the app offered only preset buttons, some participants found it efficient and low-effort. Others found it maddening — hunting through a list for the option closest to what they actually felt, and settling for approximately right. That mismatch, between what you want to express and what the interface allows, showed up repeatedly as a reason connection failed to develop.

Warmth that lands, and advice that's useful. People expected care, and they expected the suggested exercises to be worth doing. Sympathy without substance wore thin. So did substance without warmth.

Talking like a person. Casual, everyday language strengthened the sense of connection. Anything that felt scripted or, worse, promotional broke it. One participant noted that the illusion of genuine connection collapsed when the chatbot started nudging them toward paid services.

A space with no judgement in it. This came up consistently as the foundation for everything else. Not being evaluated, not having to manage someone else's reaction, knowing the conversation stays private — that's what let people say more than they otherwise would.

It's worth noticing how ordinary that list is. None of it is exotic AI magic. It's roughly what makes any conversation feel good: being listened to, being allowed to lead sometimes, not being talked at, not being sold to.

Does it matter how you're feeling when you start?

Participants completed a short well-being questionnaire at the start and twice during the study. Twelve of them scored below the threshold that suggests poor well-being.

Both groups — those above and those below the line — included people who formed bonds and people who formed none. The researchers concluded that the capacity to feel connected to a chatbot doesn't appear to depend on how you're doing at the time. That's a modest finding from a small sample, but it pushes back gently against the assumption that only people in distress get attached to these tools, or conversely that distress makes connection impossible.

Is a bond with a chatbot the same as a bond with a therapist?

No, and the study doesn't claim it is. It borrows a concept from therapy research called the therapeutic alliance — the working relationship between a client and a therapist, which decades of research has linked to better outcomes. Researchers are now asking whether something comparable exists with digital tools, and calling that idea the digital therapeutic alliance.

But "resembles" is not "equals." In therapy, the alliance involves a person who remembers you, who is accountable, who can notice what you're not saying. A scripted chatbot has none of that. What this study documents is that people describe relational experiences with these tools — feeling understood, feeling supported, missing it when it's gone. Whether those experiences produce the same benefits as a human alliance is a separate question this research didn't test.

What can't this study tell us?

Quite a lot, and it's worth being straight about it.

Twenty-six people is a small group, all enrolled at one university, aged 17 to 35, mostly from computer science and psychology backgrounds, and paid to take part. That's not the general public. Four weeks is short. And there was no measurement of whether anyone's mental health improved — this was about the experience of the relationship, not about outcomes.

The apps themselves are also a moving target. Because the versions studied were largely rule-based, the findings may under- or overstate what happens with today's more fluid conversational systems. A chatbot that can genuinely follow your lead removes one of the biggest frustrations reported here — but it might introduce new ones the study couldn't anticipate.

Qualitative research like this is good at showing what people experience and why they say it happens. It isn't designed to prove cause and effect, and one study rarely settles anything.

What does this mean if you're thinking about trying one?

Mainly this: if you've felt something like warmth or attachment toward a chatbot, you're not unusual, and you're not being naive. Enough people report it that researchers are now studying it seriously.

It also suggests that fit matters more than features. If a tool insists on steering when you want to talk, or waits silently when you need a starting point, that mismatch is a real reason it isn't working — not a personal failing. Something built to move with you, rather than through a fixed script, is closer to what people in this study said they wanted. That's the thinking behind Serena, who is designed to follow where a conversation goes rather than march it through a checklist.

And the thing people valued most wasn't clever technology. It was having somewhere to say the true version of how they felt without being judged for it. That's a low bar in some ways and a high one in others — and it's worth remembering that a chatbot meeting it is a starting point, not the whole of care.

---

This article is for general information and is not medical advice. It doesn't replace assessment or treatment from a qualified professional. If you're struggling with your mental health, please speak to a doctor or licensed clinician. If you're in crisis or thinking about harming yourself, contact your local emergency number or a crisis line in your country right away.