ChatGPT to Your Taste: How to Tailor It to Fit You

Why Read This from an HCI Perspective

This article reframes AI not as a mere contest of performance, but as an interaction design problem—one that includes how people build trust, when they feel the need to intervene, and what moments make them anxious. In real products, what matters often shifts from ‘accuracy’ to things like ‘how well the system’s state is made visible’ and ‘whether users can immediately recover when it fails.’ For HCI/UX practitioners and researchers, it’s a meaningful prompt to revisit design criteria for AI services where safety is critical.

CIT's Commentary

AI agents and automation features are a lot like driving-assist systems: it’s not the smooth operation that matters most, but the ambiguous moments when they go wrong. So the key isn’t whether the model is ‘smart,’ but whether it clearly shows what it’s doing right now, whether and when users can step in, and whether there’s a recovery path when it fails. Even if a paper presents a clean framework, once you put it into a product, trade-offs emerge—like increased screen complexity and trust issues. On the other hand, approaches commonly used in the field, such as ‘using LLMs to assist with UX measurement,’ can lead to research questions about how to improve measurement convenience while also preserving consistency and avoiding bias in the criteria. In Korea’s mobile and social environment, fast interaction flows are often preferred, so there’s a strong possibility that you’ll need shorter and clearer intervention paths than in global HCI patterns.

Questions to Consider While Reading

• Q.What minimal interface elements are necessary to help an AI agent reveal its current state more effectively?
• Q.When building UX measurement tools using LLMs, how can you preserve both the convenience of automation and the rigor of research?
• Q.In Korea’s mobile service context, which intervention and failure-recovery approaches fit better than global HCI patterns?