Skilled AI in Front of People, Awkward AI: Moments When Trying to Talk Well Ends Up Twisting Relationships

Why Read This from an HCI Perspective

This article shows that what matters more than whether an AI can speak well is how users interpret what it says. In particular, in transaction contexts, friendly conversation can actually create awkwardness—an important signal for HCI and UX practitioners. It helps readers understand that adding features does not automatically lead to higher satisfaction, and that the gap between expectations and real experience drives emotional reactions.

CIT's Commentary

An interesting point is that an AI’s ‘social fluency’ is not always beneficial. In goal-oriented situations—like payments, refunds, or reservations—an offhand, chatty tone can instead violate expectations and amplify awkwardness. This illustrates well the cost that arises when an interface starts to look like a person. So the key is not to make the AI sound more human for its own sake, but to design in a way that clearly communicates why this conversation is needed and when the user can (or should) step in. Even in domestic services, a friendly tone is often used as a default, but in purpose-driven contexts—such as Naver/Kakao or commerce apps—the same pattern can produce different reactions. Ultimately, the core question isn’t ‘how naturally the AI speaks,’ but ‘what conditions allow users to comfortably accept that naturalness.’

Questions to Consider While Reading

• Q.In transaction-oriented AI, how can we distinguish situations where we should use relational (relationship-oriented) tone versus situations where we should remove it?
• Q.To detect awkwardness and expectation violations early, what user metrics could we design in real products?
• Q.Do we also need to further validate how elements in Korean services—such as informal vs. honorific speech, friendly emojis, and sentence-ending styles—would change these results?