NN/g: AI Cannot Generate the Evidence an Empathy Map Needs

Real empathy-mapping research moves at fieldwork speed — interviews, observation, the slow drift of patterns across days. AI can fill a workshop wall with sticky notes in seconds. The gap between those two tempos is the engine of a new Nielsen Norman Group piece, "AI Can't Replace Real Research in Empathy Mapping."

The article draws a clean line. AI tools can organize research a team has already gathered. They cannot manufacture the evidence an empathy map is meant to capture. NN/g's core claim is short: empathy maps summarize what a team knows about a real user, sorted into four quadrants — Says, Thinks, Does, and Feels.

What is an empathy map?

An empathy map is a shared reference point. Teams build them in workshops, often on a wall, pulling what users say, think, do, and feel into a single page the whole team can argue with. The mechanic is simple: each note lands under its quadrant.

Those notes usually trace back to transcripts, observation logs, or usability clips — the artifacts of talking to people. The NN/g piece treats that origin story as the whole point. A note that cannot be linked to a real moment sits in the wrong quadrant, even when the language sounds plausible.

Where does AI actually belong?

The article does not argue against AI in workshops. It positions AI as a useful organizer: clustering notes, summarizing transcripts, drafting prompts. The failure mode is treating AI as a participant who has talked to users when it has talked to no one.

The piece names the temptation directly. When the workshop is on the calendar and the data isn't there, an AI tool looks like a reasonable fix. The article frames it this way: "Ask it to think like a frustrated user, and within seconds, you have a page full of sticky notes." The wall looks right. The substance underneath is the problem.

What changes for the workshop?

A working rule from the article: if a note cannot be traced to a piece of evidence — a quote, a timestamp, a clip — it doesn't belong on the map. Teams that adopt the rule find their empathy maps harder to build and harder to argue with, in the right way.

Debates shift from "what would our users say" to "what did the participant say, and when." Workshop prompts turn from abstract scenarios to transcripts the team has actually read. AI clusters, summarizes, and scaffolds. It does not produce the evidence the four quadrants depend on.

The full piece sits at nngroup.com, alongside NN/g's existing guides to empathy mapping and to running the workshop itself. The next wave of UX tooling will almost certainly blur that line further, and the practice that survives will be the one that protects the evidence underneath the sticky notes.