"“What Did It Actually Do?”": Understanding Risk Awareness and Traceability for Computer-Use Agents

Why Read This from an HCI Perspective

This article clearly shows that computer-use AI agents are not just conversational interfaces, but interactive systems that actually change files, permissions, and the state of the environment. The core issue is that users often don’t know what they delegated and what remains afterward. As a result, for HCI practitioners and researchers, it turns into concrete UX challenges around building trust, enabling auditability, and supporting post-incident recovery.

CIT's Commentary

The article’s most important value is redefining ‘safety issues’ as a problem of interaction design rather than model performance. In particular, the findings are compelling in that users want to be able to reconstruct what changed after the fact, not just receive advance warnings. In real products, exposing every action in fine detail increases complexity, while providing only summaries can weaken accountability and recoverability. Here, traceability reads not as mere logging, but as a mechanism for designing points where users can intervene. Even in the domestic context, when applying agent features to platform-style services such as Naver or Kakao, or to startup offerings, the way it shows ‘what was touched’ and ‘what remains’ is likely to have a major impact on trust more than simply demonstrating intelligence. This perspective is also useful when designing LLM-based UX measurement tools: it could lead to meta-tools where the AI first summarizes residual changes or risk signals that are hard for humans to find directly.

Questions to Consider While Reading

• Q.What visual hierarchy would be most effective for increasing post-hoc traceability while keeping the information density manageable for everyday users?
• Q.When separating and presenting permissions, execution, and residual state, where is the boundary that users most commonly misunderstand, and what is the smallest unit needed to reduce that misunderstanding?
• Q.If you use an LLM to automatically summarize an agent’s execution traces, how can you present errors or omissions in the summary in a form that can be verified?