“When I see Jodie, I feel relaxed.” — Examining the Impact of a Virtual Supporter in Remote Psychotherapy

Why Read This from an HCI Perspective

This article shows that AI is not just a simple automation tool, but an interaction design that helps users feel safe and opens the door to speaking up. In sensitive, safety-critical contexts like remote psychotherapy, it’s possible to read how screen layout, pacing of responses, and intervention pathways can either build or break trust. For HCI/UX practitioners, it offers a sense of designing the boundaries of an experience beyond mere functionality; for researchers, it raises questions about how to validate human–in-the-loop intervention structures.

CIT's Commentary

The most interesting point is that the system aims for a ‘less disruptive companion’ rather than a ‘smarter AI.’ During therapy, it speaks almost not at all, and only briefly intervenes before and after sessions—an approach that clearly demonstrates how, in safety-critical systems, the interface itself becomes a safety mechanism. However, rule-based dialogue comes with a trade-off: instead of protecting boundaries, it can reduce the flexibility of empathy that users expect. In a real product, striking this balance is the core challenge. It can be applied right away to remote counseling in Korea, healthcare apps, and conversational agents, but in platform environments like those of Naver or Kakao, transparency design—such as how privacy notices are presented and where user-intervention controls are placed—will likely be even more strongly required. Interestingly, the next research questions raised here are less about how much LLM capability to include, and more about how therapeutic boundaries are maintained even with LLMs, and how the system recovers when things go wrong.

Questions to Consider While Reading

• Q.A structure that is nearly silent during therapy can make real users feel stable, but it may also prevent deeper empathetic experiences. What interaction principles could be used to design this balance?
• Q.Rule-based dialogue is safe, but it can easily feel rigid. Even if you introduce some LLM capabilities, what kinds of failure-recovery mechanisms are needed to reduce boundary violations and misunderstandings while still achieving naturalness?
• Q.How do the moments when users accept an AI supporter as a ‘friend’ versus when they see it as a ‘tool’ differ? How should we measure the impact of that difference on trust, reliance, and therapeutic outcomes?