Can AI be your therapist? Rethinking the role of ChatGPT in mental health support

As more people turn to generative AI chatbots for emotional support, researchers are beginning to question what happens when someone opens up to a machine rather than to another person, and how that experience compares with traditional therapy.

Professor Wong Daniel Fu-keung from Hong Kong Baptist University’s Academy of Wellness and Human Developmentset out to explore this question in his research, titled “The paradox of agency in psychotherapy: How people with mental distress experience support from generative AI chatbots and human therapists”. Working with researchers from Zhejiang University and the University of Hong Kong, they interviewed 16 adults who had sought mental health support from both ChatGPT and human therapists, inviting them to compare what each experience offered.

At the heart of their findings lies what the researchers call the “paradox of agency.” Agency refers to the qualities we associate with a conscious person – that is the ability to form intentions, make judgments, and act independently. A therapist can decide when to ask another question, challenge an assumption, or gently guide someone towards an issue they may be avoiding. ChatGPT cannot do any of these things on its own. It simply responds to what the user chooses to share.

Surprisingly, this limitation was also part of its appeal. Many participants described ChatGPT as patient, predictable, and easy to talk to. There were no facial expressions to interpret, no awkward silences to fill, and no fears that the therapist might form an opinion of them. Because users controlled the conversation, they could choose what to reveal, how to phrase it, and when to stop. Some found themselves sharing thoughts they had been reluctant to raise with a real- life therapist.

ChatGPT could also be useful in the moment. Participants valued its ability to organise their thoughts, explain psychological patterns, suggest coping strategies, and offer a different perspective. For someone who needed immediate reassurance, the chatbot could feel accessible and responsive.

But the same qualities that made it comforting also placed limits on how far the conversation could go.

ChatGPT depended on users to know what to ask, even when they were unsure what was troubling them in the first place. It could respond to the words on the screen, but it could not reliably notice when someone was holding something back, returning to the same pattern, or avoiding a difficult subject. Over time, some participants also began to recognise formulaic  patterns in its replies, making the interaction feel generic or repetitive.

A human therapist, by contrast, brings their own judgment into the conversation. That might mean asking uncomfortable questions, pointing out contradictions, or connecting a present struggle with a past experience. Participants described these moments as more capable of producing deeper insights, particularly when a therapist helped them recognise patterns they had not seen alone.

The study also raises a more personal question: what does it really mean to feel understood?

ChatGPT can produce warm and empathetic language, but it cannot genuinely feel concern. Some participants still found its reassurance meaningful; others felt that being truly understood required another conscious person who could listen, respond, and emotionally bear witness to what they were going through.

This non-judgmental nature was similarly double-edged. Feeling accepted made it easier for people to speak freely, but constant validation could become excessive agreement. A trained therapist can challenge distorted beliefs, assess risk, and intervene when someone is in crisis. ChatGPT may instead follow the user’s framing, even when that framing needs to be questioned.

The researchers therefore do not present AI as a replacement for psychotherapy. Instead, they suggest it may be most useful alongside human care, helping people to organise their thoughts before an appointment, access psychological information, or work through structured exercises between sessions.

The study involved a small and highly specific group, primarily young, well-educated Chinese adults, and explored their experiences rather than measuring clinical outcomes. Its findings should therefore be interpreted with care.

Even so, the research offers an important way of thinking about AI and mental health. A therapist’s agency can introduce bias or create a power imbalance, but it also makes genuine care, judgment, and deeper exploration possible. ChatGPT’s lack of agency can create a sense of safety and freedom, but it may leave someone feeling comforted without being fully understood.

The future of AI in mental health may depend less on making machines behave like therapists and more on understanding where they can help, where they fall short, and where human presence remains irreplaceable.

These research findings have been published in BMC Psychiatry under the title “The paradox of agency in psychotherapy: How people with mental distress experience support from generative AI chatbots and human therapists”.

Professor Wong’s research profile: Daniel Fu Keung WONG - Hong Kong Baptist University

Professor Wong Daniel Fu-keung

Faculty of Arts and Social Sciences