Computer rendering of artificial intelligence used in preliminary psychiatric interviews. Courtesy of KAIST
The standard intake interview in a psychiatric clinic is a delicate, time-consuming calculus: A clinician must extract highly structured medical histories while simultaneously evaluating a patient’s emotional state.
Now, researchers in Korea are outsourcing the administrative half of that equation to artificial intelligence (AI), attempting to solve a chronic bottleneck in mental health triage.
A joint team from KAIST and Gangnam Severance Hospital has developed a generative AI assistant designed to conduct preliminary psychiatric interviews. Built on a large language model, the system engages patients before they ever step into a doctor's office, mapping out symptoms and medical histories in real time to build a comprehensive clinical dashboard for the attending physician.
The research — led by Lee Ui-chin and Lee Tak-yeon of KAIST, alongside Kim Eun-joo, a professor of psychiatry at Gangnam Severance Hospital — was presented this month at ACM CHI 2026, a premier international conference on human-computer interaction.
Unlike rigid digital questionnaires, the conversational agent adapts dynamically to patient inputs. To coax out sensitive information, the AI is programmed with basic clinical communication techniques, using empathetic prompts, restatements and clarifying questions to mimic human rapport. In a simulation study mapping 1,440 virtual patient profiles, the system successfully gathered the core diagnostic data required for an initial consultation in under 30 minutes in the vast majority of cases.
The technology arrives amid a broader, polarizing push to integrate AI into clinical medicine. Critics frequently warn that large language models are prone to "hallucinations" and lack the authentic empathy required for psychiatric care.
The researchers emphasized that the system is strictly an assistant, not a diagnostic proxy. Final evaluations and treatment plans remain entirely under human purview. Furthermore, the team acknowledged that the current iteration of the software remains blind to subtle emotional micro-expressions and struggles with highly complex, sensitive disclosures — underscoring that while AI can streamline the paperwork, interpreting it still requires a physician.
This article was published with the assistance of generative AI and edited by The Korea Times.
Source: Korea Times News
