Research Statement

2026-09-13

Hyun Seung Moon · Ph.D. student, KAIST Industrial Design — AI Experience Lab · Minor in Data Science · Updated September 2026

I study how people learn what can't be taught — by talking with AI that plays the other side.

Some of the most consequential human abilities — reading a patient's hesitation, taking a historical figure's perspective, sensing when to concede in a negotiation — cannot be transmitted through lectures or textbooks. They are tacit: grown only through lived interaction, failure, and reflection. Yet opportunities to practice them are scarce, expensive, and high-stakes precisely where they matter most. My research builds AI interlocutors — conversational agents that take the counterpart's role — and studies how conversational simulation with them cultivates social skill, tacit knowledge, and professional judgment.

What I have built and found

Across five peer-reviewed projects, I have designed, built, and evaluated AI systems that inhabit a social role rather than serve as a tool:

HistoChat (CSCW 2025, co-first author) turned historical figures into conversational personas for middle-school history classes. Students stopped treating the AI as an answer machine and began treating it as an epistemic partner — developing historical empathy through dialogue.
A flexible psychiatric AI interviewer (CHI 2026) explored how LLMs can conduct clinical history-taking under expert supervision, evaluated with 1,440 simulated dialogues and 19 clinicians. It surfaced a design principle I keep returning to: some behaviors must be coachable, others must be guardrailed — and systems must know which is which.
My master's thesis built Artificial Social Actors that negotiate on a user's behalf, identifying excessive compliance as the core failure mode of AI in adversarial social roles.
In industry and clinical practice, I deployed these ideas at scale: a synthetic consumer agent system built on 4,000-respondent survey data, deployed inside LG Electronics for four months with 47 practitioners across six departments; and voice AI agents now in prospective pilot studies at Asan Medical Center and Severance Hospital, where medical trainees practice psychiatric interviews with simulated patients on demand.

A consistent pattern runs through all of this work: whatever the domain, the central design tension is the same — how much should the simulated counterpart adapt to the learner, and who controls that dial? Users never want full automation; they want vetting, scaffolding, and calibration. This recurring structure, observed independently in five domains, is the empirical backbone of my research agenda.

Where I am going

My doctoral research takes education as its umbrella: understanding conversational simulation with AI interlocutors as a general mechanism for growing judgment — what Schön called the reflective practicum, and what Polanyi and Goodwin described as tacit knowledge and professional vision. Three questions organize the agenda:

Design — What makes an AI counterpart pedagogically effective: fidelity to a real counterpart, or calibrated deviation from one? When should a simulated patient be difficult, and when supportive?
Supervision — How should learners and educators steer these simulations? I am extending the coachable-vs-guardrailed distinction into interfaces where experts tune counterpart behavior the way they would coach a human role-player.
Growth — How do we know judgment actually improved? I combine large-scale simulated evaluation with field deployment in real classrooms and clinics, measuring transfer beyond the simulation itself.

I care about research that survives contact with the real world. Every system I study is built to be deployed — in classrooms, hospitals, and enterprises — because the phenomena I study, tacit skill and judgment, only show themselves there.

Contact: mzes0401@kaist.ac.kr · hyunseungmoon.net