AI for UX · Step 4 · 6 min read
AI-assisted synthesis, step by step
A practical method for using AI in research analysis without losing rigour: prepare and anonymise the data, read first, let AI do a first pass, challenge it, trace every insight to source, and report honestly. With what studies say about AI coding of interviews.
AI can tag a hundred pages of transcripts in a minute. Whether those tags mean anything is your job. Used well, AI is a tireless assistant for the mechanical parts of analysis: first-pass tagging, grouping, finding quotes. Used badly, it produces a confident summary that misses the point and invents the rest. This note gives you a method that keeps the speed and the rigour.
What you'll be able to do after reading:
- Prepare research data so it is safe to use with AI
- Run a seven-step synthesis where AI assists and people decide
- Spot the typical failures of AI analysis: shallow clusters, a giant "Other", skipped passages, invented insights
- Report findings that trace back to real evidence
1. What studies say about AI coding of interviews
- Human-coded excerpts captured61.7%
- Descriptive codes matched63.7%
- Interpretive codes matched57.7%
- A 2026 study compared GPT-4o coding of interview excerpts with human coders: the model captured 61.7% of what humans coded, did better on descriptive codes (63.7%) than on interpretive ones (57.7%), and also found 206 correct excerpts humans had missed (Turner et al., 2026).
- Earlier work found an LLM "can infer most of the main themes" in open interviews, while its author framed the paper as a provocation on the limits (De Paoli, 2023), and that with an expert-written codebook, GPT-3 reached "fair to substantial" agreement with experts (Xiao et al., 2023).
The pattern: AI both misses things and adds things, and it is weakest where interpretation matters most. That is exactly why the method below keeps interpretation human.
2. The method, in seven steps
Anonymise and check.
Before any AI output.
Give AI your codes, not a blank page.
Tags, clusters, quotes with locations.
Contradictions, gaps, the "Other" pile.
Check every key quote at its source.
Say where AI assisted.
1. Prepare and anonymise
Remove names, faces, company and customer details before anything goes into an AI tool, and check the result: NN/g reports a tool that stripped product names by mistake (NN/g). Under European data protection guidance, only truly anonymised data stops being personal data; pseudonymised data is still personal (EDPB anonymisation guidelines, 2026 draft).
2. Read first
Read or watch a sample of sessions in full before you see any AI output. Otherwise its framing becomes yours.
3. Give AI a codebook, not a blank page
Ask it to tag against your research questions and a short codebook. Open-ended "find the themes" invites generic themes.
4. Let AI do the first pass
Tagging, clustering, pulling candidate quotes, summarising each session. Ask for quotes with locations (session, timestamp) for every claim.
5. Challenge it
Look for the typical failures NN/g lists: clusters based on keywords rather than meaning, a large "Other" group, and passages it skipped (NN/g). Ask: what contradicts this theme? Who is missing? What's surprising?
6. Trace every insight to source
Open the recording or transcript for each key finding and check the quote is real and in context. A finding that can't be traced is deleted.
7. Report honestly
Say where AI assisted, how many sessions people read in full, and what was checked. NN/g's PACED framework helps decide how to disclose AI use (NN/g).
3. Tools you'll meet
Research repositories now build AI into synthesis. As of November 2026, for example:
- Condens: AI-suggested tags and highlights, auto-clustering, session summaries, "ask your repository", hosting in the US or EU (Condens)
- Dovetail: chat, agents, theme and sentiment classification, synthesis, redaction (Dovetail)
- Marvin: AI-moderated interviews, answers with citations, analysis tools, automatic blurring of personal details (Marvin)
Whatever the tool, check its data terms (where data is stored, whether it trains on your content) before uploading participants' data. See AI governance and data protection.
4. Vocabulary
| Term | Meaning |
|---|---|
| Coding (tagging) | Labelling parts of the data with themes or categories |
| Codebook | The list of codes and what each means |
| Descriptive code | A label for what was said ("mentions price") |
| Interpretive code | A label for what it means ("feels judged by the app") |
| Affinity mapping | Grouping notes by similarity to find themes |
| Pseudonymisation | Replacing names with codes; still personal data |
5. Try it (no code needed)
- Take three public sample interview transcripts.
- Write a five-code codebook from a research question.
- Code one transcript yourself. Then ask an approved AI tool to code all three with your codebook, quoting with locations.
- Compare your transcript with its version: what did it miss, what did it add, and which of its additions were right?
Reflection questions: Which of your codes are descriptive and which interpretive? Where would you trust the AI's first pass, and where would you redo it?
6. If you lead a team
- Make "read first" and "trace to source" part of every study's definition of done.
- Keep a shared codebook per product area that AI tools can use.
- Review the consent form and tool terms with legal before any AI processing of recordings.
Checklist
Learn more
- NN/g: Accelerating Research with AI · When Should You Disclose AI Use? The PACED Framework
- Research: Turner et al., LLMs for deductive qualitative coding (2026) · De Paoli, thematic analysis with GPT-3.5 (2023) · Xiao et al., LLM-assisted deductive coding (2023)
- Data and consent: EDPB anonymisation guidelines (2026) · MRS and ESOMAR on client data, AI and consent
Books, courses and videos to go deeper
Related
AI in UX research · Synthetic users and digital twins · AI governance and data protection · Affinity mapping · Hallucination