Using AI for UX Research — What I Got Right, What I Got Wrong, and What I’d Do Differently

A few months ago I started using AI for the unglamorous parts of UX research — synthesizing interview transcripts, analyzing surveys, clustering feedback, generating personas. The output was fast. It looked done. I almost trusted it.

Then I started reading the AI’s summary next to the actual interview.

This article is structured that way. On the left, you’ll see the AI’s clean version of a user moment. On the right, what I noticed when I was actually in the room. The gap is the article.

AI: “The user is generally satisfied with the application.”

Me: What they actually said was “yeah, it’s fine.” There was a four-second pause before they said it. They looked at the camera, then away, then said the words like they were choosing a polite version of something heavier. I knew, in that pause, that they were not satisfied. I knew it because I was sitting there. The AI logged a sentiment. I saw a person deciding not to tell me the truth.

AI: “User checks the dashboard regularly in the morning.”

Me: What they actually said was “I always check it in the morning… actually, no, I usually forget. I know I should though.” That’s three different things in one sentence. An aspiration, a behavior, and a feeling about the gap between them. The AI averaged them into one tidy bullet. The bullet is wrong. None of those three things is “checks the dashboard regularly.” The interesting data is the gap between what they think they should do and what they actually do. The AI flattened the gap into a sentence and lost the entire insight.

AI: “User finds the application functional.”

Me: What they actually said was “it’s not that the app is bad, it’s just — “ and they trailed off. They never finished the sentence. The unfinished sentence is the data. Whatever they could not say is exactly what we needed to find out. The AI looked at an unfinished sentence and produced a complete summary anyway. It filled in the gap. The thing it filled the gap with was an assumption.

AI: “User would benefit from a simplified interface.”

Me: They never said simplified. They said “overwhelming.” That word has a different temperature. Simplified is a design instruction. Overwhelming is an emotional state. One leads to reducing options. The other leads to reducing pressure. They are not the same intervention. The AI converted a feeling into a feature request and lost the part of the conversation that would have actually helped.

AI: Persona: organized, motivated, time-conscious user who values efficiency.

Me: This persona is composed of three things this user contradicted within the same interview. They forgot to do the morning check they said they always do. They paused before describing themselves as efficient. They got quiet when I asked about productivity tools. Real users are not their stated preferences. They are the tension between what they say and what they do. The AI generated the stated preferences and called it a person.

The pattern, after enough of these.

AI is excellent at what was said. It is useless at what was meant. It is excellent at the words. It is blind to the texture around the words — the pause, the trail-off, the contradiction, the temperature of the chosen phrase, the body in the chair. The texture is where the research actually lives. The words are just the receipt.

I think this is the part most quickly going wrong in the way teams are adopting AI for research. The summary looks complete. The themes look organized. The personas look professional. Nothing about the output signals that something has been lost. There is no warning label that says the texture is gone. There is only the polish. And the polish is not the truth.

What I do now, in practice, is simple but uncomfortable to actually maintain.

I read the source before I read the summary. Always. If I read the AI version first, my brain anchors to it, and when I read the raw material I find myself confirming the AI’s version rather than discovering my own. The order matters more than the speed.

I read AI personas as drafts, never as deliverables. The structural scaffolding is fine. The content needs me to rewrite it with the contradictions and tensions and texture the AI smoothed away.

And I ask one question of every AI output before I let it into the world. Can I defend this if the AI disappears? If I can stand behind every claim, knowing where it came from in the source and why it matters, I have done the work. If I cannot, I have not researched. I have forwarded.

The cost of treating AI output as already-reported is that you stop being the researcher. The work feels done. The deliverables look right. But the user has been compressed into a summary, and the thing you needed to know about them — the pause, the trail-off, the word they did not use — never made it through.

Use the tools. The acceleration is real. But the part that matters about user research has not changed and is not going to change. It is the part that happens between the words. The part the AI cannot hear.