AI’s Two Faces: Ethical Challenges Exposed in the Age of AI

Why Read This from an HCI Perspective

This article reframes AI safety not as a question of whether a model is ‘smart,’ but as an HCI question of where and how users place their trust—and where things fail in interaction. Issues such as hallucinations, bias, jailbreaks, and alignment/over-agreement are not merely technical defects; they are failure modes of interaction. For practitioners, it raises new challenges around trust design and intervention pathways. For researchers, it calls for new tasks in measuring safety and defining evaluation criteria.

CIT's Commentary

The interesting point is that AI safety ultimately boils down to interface and usage context. Even with guardrails or human-in-the-loop training, risk remains if users cannot understand the system’s state. In particular, for Korean-language services, it’s difficult to simply transplant English-centered benchmarks; evaluation must reflect Korean speech styles, relationship contexts, and platform usage habits. The dataset examples Team Naver has released prompt a question beyond ‘a safe model’: what does it mean to build a product that looks safe and can intervene safely? Furthermore, even if LLMs are used to assist UX measurement tools, the tools themselves—along with their bias and reproducibility—must be rigorously validated.

Questions to Consider While Reading

• Q.How far should users trust an AI’s answers, when should they doubt them, and how can we make them ask the right questions?
• Q.When evaluating safety in Korean-language services, what failure modes differ from English-speaking standards?
• Q.As guardrails get stronger, convenience may drop—how should that trade-off be designed in real products?