Why People and AI Often “Misalign” on Digital Platforms

Why Read This from an HCI Perspective

This article reframes AI not as a technique for “getting it right,” but as an interaction that changes how people behave. In particular, it highlights how signals that look clear—such as clicks and time spent—may not actually reflect users’ true intentions. This connects to a common UX reality: experiences that work well while you’re using them, but leave you feeling uneasy afterward. When evaluating everyday systems like recommendations, feeds, and agents, it makes you ask again what, exactly, should be optimized.

CIT's Commentary

The core point of this piece is that even high-performing models can cause bigger problems when they chase the wrong objective. Especially in systems like recommendation feeds, where user behavior becomes training data again, even a small nudge strategy can eventually spread into collective outcomes. So rather than only trying to improve explainability, we should also design pathways that let users see the current state of the system and intervene. And because it’s hard to measure a “good experience” using only clicks or time spent, even when using LLMs, it’s important to rigorously validate the validity and bias of the measurement tools themselves. In industry, engagement may look better immediately while long-term trust erodes; conversely, turning that tension into research questions creates practical challenges such as how to reliably collect and reflect users’ reflective preferences.

Questions to Consider While Reading

• Q.When different metrics—like clicks, time spent, and satisfaction—pull in different directions, which combination comes closest to real user value?
• Q.What kind of interface is needed in recommendation feeds or AI agents so users can understand what the system is currently optimizing and intervene?
• Q.When building UX measurement tools with LLMs, where should we draw the line between convenience gained through automation and research rigor?