Marcus Andrey Study
Ler em português Library

AI for UX · Step 4 · 5 min read

AI in UX research

Where AI genuinely helps across the research cycle and where people must stay in charge: planning, recruiting, moderating, transcribing, analysing and reporting, the five big risks, and a safe end-to-end workflow.

AI can take the busywork out of research, and it can also take the learning out of it. Scheduling, transcription, data clean-up and first drafts are faster than ever. But if a machine does the listening and the thinking, the team gets a polished report and loses what research is for: understanding people. This note shows where to hand work to AI and where to keep it.

Two trays of stones: synthetic answers almost alike at the back, people's answers of every size at the front, with one odd stone, the answer research exists to find. In NN/g's 2026 study, people gave more than 50 kinds of answer where synthetic respondents gave about 12. Point at a tray to see its spread. Illustration: Marcus Andrey, drawn with Hairline (MIT, Lucas Marques).

What you'll be able to do after reading:

  • Decide, for each stage of a study, what AI can do and what a person must do
  • Name the five big risks of AI in research, and how to avoid each
  • Follow a safe workflow from recording to report
  • Explain to stakeholders why "AI read the interviews" is not research

1. Hand off the busywork, keep the learning

NN/g's Maria Rosala puts it plainly: don't outsource the learning. Hand AI the logistics, keep the parts where understanding happens (NN/g, Jul 2026):

Hand offGive AI the busywork

Recruiting, scheduling, transcription, data clean-up, first-draft guides, formatting.

Hours saved, little lost
KeepKeep the learning

Moderating, observing live, debriefing together, interpreting what it means.

Where understanding happens
What to hand to AI in a study and what to keep, following NN/g's Maria Rosala (Jul 2026). Diagram: Marcus Andrey, CC BY 4.0.

Why keep the hard parts? Learning comes from effort. Reading a polished AI summary gives an illusion of learning: you feel informed, but you haven't met the people.


2. Stage by stage

Stage AI can A person must
Planning Draft a discussion guide, screener or survey items Set the research questions and what the business needs to learn
Recruiting and scheduling Screen, schedule, send reminders Decide who you need to hear from, including people who are hard to reach
Moderating Run short structured interviews in some cases (see below) Moderate open, exploratory sessions; follow the surprise
Transcription and translation Speech to text, translation Check key quotes against the recording
Analysis First-pass tagging, clustering, summaries Interpret, weigh, find contradictions and what's missing. See AI-assisted synthesis, step by step
Reporting Draft summaries for each audience Choose the evidence, tell the story, make recommendations
Desk research Summarise existing studies and documents Check every source: AI may cite things that don't exist

What about AI moderators?

NN/g tested two AI-moderated interview tools with ten research leaders from eight countries. Only 3 of 10 said the conversation felt natural; both tools were sycophantic and could only run structured interviews (NN/g, Jan 2026). Possible uses: quick product feedback, teams without researchers, interviews in other languages, screening. Not a replacement for an exploratory interview.


3. Five big risks

  1. Invented evidence. AI "might miss, misinterpret, or even manufacture insights" (NN/g). Trace every insight back to raw data.
  2. Shallow synthesis. AI finds frequent themes, not important ones. One expert's workaround can matter more than ten mentions of "slow".
  3. Synthetic users passed off as research. AI respondents that answer like users. Useful to pilot a guide or brainstorm, never a substitute. See Synthetic users and digital twins.
  4. Participants' data. Recordings and transcripts with faces, voices and names are personal data. Use approved tools only, and anonymise first. See AI governance and data protection.
  5. Losing empathy. If nobody watches or reads the sessions, nobody feels the user's frustration. Watch at least some sessions in full.

4. A safe workflow

1Record

With consent that mentions AI processing.

2Transcribe

With an approved tool.

3Anonymise

Names, faces, customer details; then check.

4Read a sample

People read or watch sessions in full.

5AI first pass

Tagging, clustering, candidate quotes.

6Human synthesis

Interpret, weigh, find contradictions and gaps.

7Trace to source

Each key insight to at least two raw sources.

8Report

With a note on where AI assisted.

A safe research workflow with AI, from recording to report. Diagram: Marcus Andrey, CC BY 4.0.

The rule in step 7, trace every key insight to at least two raw sources, is this course's recommendation, not a standard; adapt the number to your study.


5. Vocabulary

Term Meaning
Discussion guide The plan of questions and topics for an interview
Screener Questions that check whether someone fits the study
Transcript The written record of what was said
Synthesis Turning raw research data into findings and insights
AI-moderated interview An interview run by an AI system instead of a person
Synthetic users AI-generated respondents that imitate users

6. Try it (no code needed)

  1. Take a public interview transcript (many research courses publish sample transcripts).
  2. Read it in full yourself and write your three main findings.
  3. Ask an approved AI tool for "the three main findings, with quotes".
  4. Compare: what did it find that you didn't? What did you find that it missed? Check every quote against the transcript.

Reflection questions: Which part of your research process is busywork you'd happily hand off? Which part, if you skipped it, would change what your team understands about users?


7. If you lead a team

  • Agree which research tools are approved for which data, with your data protection officer.
  • Add "trace to source" to your research definition of done.
  • Once, compare an AI synthesis with a human synthesis of the same study, and share what each missed.
  • Update consent forms to mention AI processing where you use it, with legal.

Checklist

Learn more

Books, courses and videos to go deeper

Synthetic users and digital twins · AI-assisted synthesis, step by step · AI governance and data protection · Prototyping with AI · Synthetic users

Saved only on this device.

↑ ↓ to move · Enter to open · Esc to close