What an empathy map can and cannot be built from

An empathy map assembled from a model output is a map of the model. What it takes to tell the difference in a readout.

An empathy map laid out in four quadrants during a research session

The artifact looks the same either way. That is the problem.

An empathy map is a summary of evidence. Filled in from a model rather than from sessions, it is a summary of nothing, and it looks identical.

Nielsen Norman Group made this point directly in August 2026. The reason it needs making is that the artifact is formatted the same way in both cases, so the failure is invisible at the moment it does the most damage, which is when the map is on a wall and a team is making decisions in front of it.

Understanding why the map is vulnerable

Most research artifacts carry their provenance. A transcript is a record. A survey has an n. An empathy map has neither by default. It is four quadrants of paraphrase, and paraphrase is exactly what a model produces well.

It is also an artifact that teams often build without the researcher present, in a workshop, from memory. That habit predates AI and it is what makes the tool substitution feel natural rather than suspicious.

Making the difference visible in the artifact

The fix is structural rather than cultural. Every item on the map carries a source identifier: the participant and the moment it came from.

Items that cannot carry one go in a separate area labelled as assumptions. Not deleted, because a hypothesis is useful, but separated, because the separation is what stops a hypothesis being cited later as a finding.

A map built this way is visibly thinner than a generated one, and the thinness is accurate. Six sourced observations are worth more than twenty fluent ones.

Six sourced observations are worth more than twenty fluent ones. The thinness is accurate.

BrilliantUX editorial principle

Using the tools where they are safe

None of this rules out using AI on the way to a map. Clustering real transcripts, pulling every quote where a particular frustration appears, and drafting a first grouping that a researcher then corrects are all reasonable uses, and they save hours.

The line is that the tool organises evidence the team collected. The moment it supplies the evidence itself, the artifact has stopped describing users.

Source

Nielsen Norman Group, AI Can't Replace Real Research in Empathy Mapping, 28 August 2026.

nngroup.com/articles/ai-empathy-mapping