What an interview gets that a model cannot

Nielsen Norman Group put the case plainly in August 2026: an empathy map built from a model's output is a map of the model.

The argument is not that AI has no place in research. It is that the line between organising evidence and inventing it is easy to cross without noticing, because the output looks identical on either side of it.

Separating synthesis from generation

Synthesis operates on evidence a team actually collected. Transcripts from sessions that happened, support tickets that real customers filed, survey responses from real respondents. A tool that clusters forty transcripts into themes, surfaces the quotes behind each theme and flags the ones that contradict each other is doing work a researcher would otherwise do by hand, and doing it faster.

Generation operates on nothing. A tool asked to describe what a mid-market finance manager finds frustrating about invoice approval will produce a fluent, plausible and entirely unsourced answer. The output is a summary of how that role is described on the internet, which is a different thing from what the people using the product actually do.

The reason this is hard to police is that both outputs arrive in the same format, with the same confident tone, and both fit neatly into the same slide.

The test that works on a live project

One question separates them: for any claim in a readout, can someone name the session it came from?

If a finding traces to participant four at minute eleven, it is evidence. If it traces to nothing, it is a hypothesis, and it may well be a good one, but it cannot carry a design decision on its own. The discipline is to keep the traceability visible in the artifact rather than in the researcher's memory, which in practice means every theme on a map or a matrix carries its source identifiers.

Synthesis operates on evidence a team collected. Generation operates on nothing. Both arrive in the same format.

BrilliantUX editorial principle

Keeping the interview in the process

What an interview produces that no model can is the thing the participant did not know to say. A person describing their workflow will narrate the version they believe they follow. Watching them work surfaces the workaround they stopped noticing, the spreadsheet that sits beside the product, and the step they do twice because they do not trust the first one.

Those are the findings that change a product, and they exist only in observed behaviour. No amount of modelling recovers them, because they were never written down anywhere to be modelled from.

The workable arrangement is to use AI on the volume problem and keep people on the discovery problem. Let a tool cluster and index what was collected. Keep interviews, contextual inquiry and observation as the way the evidence gets created in the first place.