Making a Korean AI Agent Feel Like Real People: Training Without Data Using Synthetic Personas

Why Read This from an HCI Perspective

Rather than treating an AI agent as simply a ‘more intelligent model,’ this article shows what kind of people it should target, what tone of voice it should use, and what procedures it should follow. In environments like Korea—where honorifics, regional conventions, and public-service workflows matter—translation alone isn’t enough. For HCI and UX practitioners, it prompts the question of why ‘context-appropriate interaction’ matters more than delivering a ‘correct answer.’ For researchers, it raises key issues around trust, intervention, and designing failure modes.

CIT's Commentary

What’s particularly interesting here is that this example treats generative AI not as something trained on a ‘dataset,’ but as something grounded on an ‘interface foundation for interaction.’ The six million synthetic personas can be training material for the model, but in practice they function more like interface mechanisms that pre-load the tone users expect, the regional flavor, and the public-service context. That said, there are clear trade-offs. The more realistic the system becomes, the more easily individual users may believe it has ‘the right answer for my situation’—which means the system state and its basis must be made more transparent. Especially in public and medical domains, personas should be only the starting point of an answer; the design must also specify when users can intervene and correct the system to ensure safety. This approach can work even more strongly in places like Korea’s tech ecosystem, where many services are deeply integrated into everyday life. At the same time, research is needed on how to validate ‘who the synthetic persona is actually representing.’ Even if LLMs help generate personas or support UX measurement tools, the rigor must be maintained so that the tools do not blur how well users are truly understood.

Questions to Consider While Reading

• Q.How can we verify whether synthetic personas increase real user trust—or instead encourage overconfidence?
• Q.In high-stakes areas like public health and healthcare, what interface patterns are appropriate for intervention paths and error recovery in persona-based agents?
• Q.Where might persona grounding tailored to the Korean context conflict with general principles in global HCI research?