Weekly Highlights

Although these news items seem to cover different fields, they converge on the same question. The key is not how well AI can judge things, but what signals people receive and what authority they have to intervene when that judgment is wrong. Autonomous driving, mental-health support, conversational LLMs, and recommendation systems may all look like performance competitions on the surface, but the real risks usually grow from interface silence, gaps in accountability, and distorted objectives. In particular, once users start treating the system not as a simple tool but as an actor that makes some judgments on its own, UX stops being merely a convenience issue and becomes a matter of safety and responsibility.

The recent trend these stories collectively point to has three parts. First, HCI’s focus is shifting from usability to controllability. In the past, it was important how naturally and smoothly AI worked; now, the key design criteria are when users feel unsure, when they should stop, and how users can correct things immediately. The Tesla case and discussions of high-risk AI interfaces show that an experience where automation continues seamlessly is less important than one where incorrect automation is interrupted and corrected. Second, safety is no longer treated only as a problem of filtering inside the model. Studies on delusion-context conversations and peer support show that risk does not arise solely from a single answer; it emerges through the formation of relationships and the accumulation of responsibility over long interactions. As a result, interaction structures—such as state transparency, escalation paths, and points where humans can intervene—have become as important as model performance. Third, UX evaluation criteria are moving from optimizing short-term reactions to aligning with long-term goals. As recommendation research notes, clicks and time-on-page are easy to measure, but they often diverge from users’ long-term well-being or considered preferences. This same issue carries over to generative AI services: just because responses are currently good and conversations are longer does not necessarily mean it is a good experience.

The biggest implication for practitioners is that simply making AI smarter is unlikely to earn product trust. What matters going forward is designing together: screens that do not hide uncertainty, warnings that ensure people do not miss the right moment to intervene, and operational structures that do not shift blame after failures. For researchers, beyond evaluating explainability or accuracy, more refined methodologies are needed to measure how responsibility is assigned and how recovery is possible in real usage contexts. Especially in Korea’s service environment, it is important to consider users’ tendency to treat AI as a relational presence rather than just a tool, mobile-centered short and frequent usage, and the social context in which family and friends may also intervene. In the end, the focus should not be on how accurately AI gets things right, but on how quickly it becomes visible when it is wrong, who intervenes and how, and whether users can regain control.