How AI is changing user research (and what it still can't do)

The conversation about AI and user research splits along an unhelpful line. One side claims AI will replace researchers, citing tools that generate personas and synthesize transcripts. The other side claims AI is fundamentally incapable of doing research, citing the things it gets wrong.

Both miss what's actually happening. AI is good at a specific part of the research workflow and bad at a different specific part, and the people who understand the split are using it to do better work in less time.

What AI is genuinely good at

Synthesis. If you've ever run a research session, you know the next twelve hours: transcripts to clean up, notes to organize, themes to extract, quotes to tag. This used to be hours of plodding work that researchers either did themselves and lost time, or skipped and lost rigor.

AI cuts that work substantially. A model can read forty hours of interview transcripts, surface the patterns that repeat, pull supporting quotes, and flag the disagreements. The synthesis isn't perfect. It tends to over-cluster and miss subtleties a human would catch. But it's a starting point that compresses days into hours.

The same applies to survey analysis, support ticket review, and any other text-heavy work where the question is "what are people saying about this." The volume problem that limited research scope is now mostly gone.

AI is also useful for prep work. Generating discussion guides, drafting screener questions, summarizing prior research on a topic. These are all places where a competent first draft saves the researcher real time, and the researcher's editing produces a better final than starting from scratch.

What AI is bad at

Doing the actual research. A real research session, with a real participant, is a thing AI can't replicate. The participant might not say what they mean. They might contradict themselves. They might do something you didn't expect and the question is whether you notice and follow up.

A human researcher reads the silence after a question, picks up on the tone shift when something hits a nerve, and rephrases when the participant clearly didn't understand. AI agents that simulate participants miss all of this because they don't have the actual thing being studied: a person trying to use your product in their own context.

The deeper issue: AI synthesizes what's already said, and research is often about what people don't say or can't articulate. Latent preferences, unconscious workarounds, the gap between stated and revealed behavior. These only surface when you observe, not when you ask. AI can't observe.

The trap of synthetic participants

There's a category of tool that generates "synthetic users" you can interview. These are useful for one thing: pressure-testing your discussion guide before you run real sessions. You can see whether the questions land, whether the framing biases responses, whether the flow makes sense.

What synthetic users can't do is replace real participants. The synthetic ones are generated from training data and produce reasonable-sounding answers that match what the model has read. They will not surprise you. They will not contradict the consensus. They will not tell you something you didn't already roughly know.

The whole point of research is to find out things you don't already know. A synthetic participant cannot serve that function, by construction.

What this means for researchers

The researchers I see thriving have made a specific shift. They spend less time on synthesis and write-up, and more time on the field work. Running sessions, observing behavior, building relationships with users. The AI absorbs the work that scaled poorly. The human time goes to the work that doesn't scale and can't be replicated.

The researchers who are nervous tend to be the ones whose value was tied up in the synthesis side. If most of your job was running transcripts through tags and writing up findings, AI has compressed that work to something a PM can do. That's not a comfortable position, and the response is to push the work upstream. Get closer to the participants. Get closer to the decisions the research is informing.

The risk to watch for

Teams that overuse AI synthesis develop a confidence problem. The output looks comprehensive, sounds confident, and produces clear themes. It also smooths over the actual texture of what participants said. The exceptions, the contradictions, the moments where one participant said something nobody else said but it might be the most important thing all get clustered into "noise" or filtered into a less prominent theme.

Reading the raw transcripts is still the work. The AI summary is a useful index into them, not a replacement for them. Researchers who skip that step are going to make confident recommendations based on synthesized findings that don't match what real users actually said. That's a worse failure mode than the slow synthesis the AI was meant to fix.