AI broke UX research.
Can AI fix it?

AI-simulated users can identify the same major usability issues as real users for most routine evaluations.

If true, this would let us run cheap, fast sanity checks before investing in real user research. The interesting question isn't "does it work" but "where does it break down." My guess is it fails on anything involving trust, emotional response, or domain expertise.

AI moderators playing passive observer roles—and human moderator removal in specific contexts—can improve data quality.

Two related ideas here. First: some participants perform more naturally without a human researcher watching. Second: AI moderators can adopt passive observer roles that humans simply cannot—infinitely patient, never reacting, intervening only when things go off the rails. The question is whether this captures what matters while reducing observer effects. I suspect the answer depends heavily on task type and participant comfort with AI.

We can detect when an AI product has changed enough that previous research findings no longer apply.

Right now, research invalidation is vibes-based. Someone notices the product feels different, maybe. What if we could instrument this—track behavioral signatures over time and flag when drift crosses a threshold? Not sure if it's possible, but worth exploring.

User trust calibration follows predictable patterns that can be measured longitudinally.

People start with some mental model of what AI can do. They use it, get surprised (positively or negatively), and adjust. Over time, their expectations stabilize—or don't. If there's a pattern here, we could design for trust calibration instead of just measuring satisfaction at a point in time. The tricky part is that people are often bad at introspecting on their own mental models, so we'd need both self-reported measures and behavioral proxies—what people say they expect vs. how they actually behave.

AI analysis of qualitative data approaches human-level quality for well-structured research questions.

"AI can analyze interviews" is too broad. The real question is: for which types of analysis, with what constraints, and how do you know when it's working? My hunch is it's good at finding patterns in explicit statements and bad at interpreting what people didn't say.

Surfacing AI confidence signals to users improves their trust calibration.

This connects the "measure the AI" angle with user outcomes. If the AI can tell you when it's uncertain—and you actually expose that to users—does their mental model calibrate better? This is a design intervention that could be tested empirically. The interesting questions: what form should confidence signals take? Do users actually attend to them? Does it help, or just create anxiety?