AI and UX research: synthesise without inventing
A method for using AI to analyse interviews and feedback while keeping sources, interpretations and limitations clearly separated.
Published · 4 September 2026Updated · 4 September 2026
00 · Premise
A plausible summary is not yet evidence.
Models can group large amounts of text and suggest patterns, but they can also flatten important differences. Research remains trustworthy only when every conclusion can be traced back to the original data.
An essential distinction
Faster analysis does not mean delegated interpretation.
AI is effective at organising material that has already been collected. It becomes risky when it fills gaps, assigns intentions or turns weak signals into certainty.
Useful use
- Index and classify
- Compare recurring themes
- Prepare verification questions
Risky use
- Invent quotes or participants
- Estimate frequency without data
- Present inference as fact
Prepare a controlled corpus
Remove unnecessary data, standardise the format and retain identifiers that make every passage traceable to its original source.
Practical actions
- Define which materials may be processed.
- Anonymise personal and confidential data.
- Keep references to session, question and participant.
Ask for structure, not conclusions
Begin with classification, extraction and comparison. Open questions become useful later, once the material and analysis rules are clear.
Practical actions
- Define categories or ask for proposals with examples.
- Require quotes linked to their source.
- Surface exceptions and opposing signals as well.
Separate evidence from interpretation
Every insight should distinguish what was observed, the interpretation being proposed and the decision it might support.
Practical actions
- Label facts, inferences and hypotheses.
- State the quantity and quality of signals.
- Do not treat model confidence as evidence.
Always return to the sources
Before sharing a result, verify quotations, context and cases that do not fit the pattern. Human review closes the loop.
Practical actions
- Check a sample for every theme.
- Review the passages that most influence decisions.
- Document limitations and remaining questions.
Ready to use
Verifiable AI synthesis
- The corpus is authorised and minimised.
- Every item of evidence retains its source reference.
- Quotes and quantities have been checked.
- Facts, inferences and hypotheses remain separate.
- A person validates insights and limitations before use.
Key principle
AI can find order; research must preserve meaning.
A good synthesis is not the one that looks most complete, but the one that remains verifiable and respects differences, context and uncertainty in the data.