uxdesign
coach.
Get a UX review
The journal

AI + UX · 8 min read

Let AI find patterns. Keep the evidence attached.

Use AI to assist synthesis without turning guesses into research findings.

Finish this exercise with a checked theme table, an evidence trail and one research question. Allow about 25 minutes for the practice. The six notes below are fictional teaching material; they are not research with real users. In real work, use consented, approved records and remove unnecessary identifying details before using an AI tool.

1. Start with a decision and bounded inputs

Practice decision: should a booking team investigate confirmation feedback, search filters or both? Keep note IDs attached throughout. Three fictional participants each contribute two notes, so six notes do not mean six people.

  • P1-N1: In a booking task, P1 selected a time, pressed Continue and said, “I don’t know if it booked.”
  • P1-N2: P1 went back to the calendar and selected the same time again.
  • P2-N1: P2 found the confirmation message and described the appointment time correctly.
  • P2-N2: P2 wanted to filter the calendar to evenings before selecting a time.
  • P3-N1: P3 completed the task on a phone, then checked email to see whether the booking existed.
  • P3-N2: P3 said the available times suited them.

Copy these notes into a new document. Keep the participant labels and the difference between observed behavior and a spoken preference. Do not add demographic details, a recording or a usability result that the exercise does not supply.

2. Ask for candidate themes with counterevidence

Use only P1-N1 through P3-N2 supplied below.
Create a table with: candidate theme, supporting IDs, counterevidence,
interpretation, unknowns, and a next research question.
Separate behavior from preference. Count distinct participants, not notes.
Do not invent quotes, severity ratings or population percentages.
Do not turn a proposed design change into a research finding.

Run the prompt in an approved tool if you want to practice with AI, or group the notes manually. Either route still requires checking every link back to its source. A plausible response is only a draft.

3. Find and repair an overstated result

Flawed practice output: “All six users failed to book because the mobile confirmation is broken. Add an evening filter to fix completion.” That sentence doubles the participant count, turns uncertainty into failure, generalizes one mobile context and proposes a remedy without evidence of causality.

Trace the claim back to its evidence1P1-N1 + P1-N2One participant: uncertainty and a repeated action.2P2-N1Counterexample: the confirmation was understood.3P3-N1Another participant checked email for reassurance.4Corrected themeInvestigate confirmation clarity; failure is unknown.
Each source has a different role. Repeated notes from one person do not create new participants.
Checked practice output
Candidate themeEvidence and counterevidenceInterpretation and next check
Confirmation confidence may varyP1-N1, P1-N2 and P3-N1; P2-N1 is a counterexample.Two fictional participants showed uncertainty or sought reassurance. Inspect the visible confirmation and whether any duplicate booking was actually created.
An evening filter may help one preferenceP2-N2; P3-N2 reports suitable times.One participant requested a filter. Investigate when time filtering matters; do not call it the cause of confirmation uncertainty.

4. Audit the table before sharing it

  1. Open each cited note and underline the exact supporting words or behavior.
  2. Check that every source ID exists and belongs to the right participant.
  3. Look for counterexamples the grouping lost. P2 completed the confirmation check correctly.
  4. Remove claims that cannot be traced. We do not know whether any server booking failed or whether a duplicate exists.
  5. State the sample boundary: three fictional participants in an exercise, not evidence about a customer population.

5. Turn the interpretation into a testable question

A useful handoff is: “Investigate whether people can explain their booking status immediately after Continue, and what they do when uncertain.” That question can guide a future study with relevant real participants. It does not prescribe a new feature before the problem is understood.

For the next study, define the task, consent and note-handling process, observable success and the information needed to distinguish a screen misunderstanding from a failed booking. Keep the original source, analyst correction and decision together so another researcher can inspect the chain.

6. Try a harder case

Add fictional note P2-N3: “P2 later received two confirmation emails for the same appointment.” Update the table. The note supports a new question about duplicate confirmations; it still does not prove two appointments were created. Your correction should preserve what changed and what remains unknown.

Finish check: another reviewer should be able to reconstruct each interpretation from its IDs, name a contradictory example and explain one next research question. If they cannot, repair the table before sharing it. No numerical score can substitute for an accurate evidence trail.

Try it yourself

Use the six supplied practice notes, check the flawed output, then add P2-N3 and revise your evidence table.

Put your research into practice

Planning your next study? Use the free AI research-planning workflow, then open the printable usability session kit to record what real participants do.

For help applying this with your team, see the scope of the AI for UX Workshop.

Keep learning

Pretty is the beginning of a critique.