Should I use AI to do thematic analysis on my interviews?

Yes for first-pass open coding and theme surfacing; no for the final synthesis. The pattern that works in 2026: AI proposes 20–30 candidate codes per transcript, the researcher accepts/rejects/merges, then AI surfaces cross-transcript patterns once tagged clips exist. The researcher reads at least one full transcript per proposed theme before accepting it. AI is the accelerator on the slow parts; judgment about what the data means stays with the researcher.

Is Dovetail AI good enough to replace manual tagging?

No, but it's good enough to make manual tagging fast. In 2026, Dovetail AI auto-tags clips on ingestion and proposes themes once enough tags exist. The pattern senior UXRs use: let Dovetail propose, then accept or reject each tag manually. Time saved is real but the verification loop is non-negotiable.

Can I use ChatGPT or Claude to write survey questions?

Use AI as a reviewer of human-drafted survey questions, not as the primary author. AI is consistently good at catching double-barreled questions, leading language, assumption-loaded framings, and missing 'I don't know' options. AI is consistently bad at writing surveys from scratch — it produces grammatical but research-naive questions biased toward the product team's assumed outcome. Erika Hall's framing on question-asking applies: the question shapes the answer.

What is Claude Projects and why does it matter for UXRs?

Claude Projects (anthropic.com/news/claude-projects) is a notebook-style workspace where you upload a fixed corpus — transcripts, the discussion guide, the research brief — and chat over it with citations. For UX research, it is vastly safer than open-web chat because the source-of-truth is bounded. Researchers ask grounded questions like 'Where did participants describe friction in onboarding?' and get answers tied back to the corpus. Same workflow as NotebookLM (notebooklm.google.com); pick on data-handling preferences.

How do I prevent AI from hallucinating quotes in my readouts?

Verify every quote against the raw transcript before it ships. There is no prompt that fully prevents hallucinated quotes; the only reliable guardrail is the researcher's verification loop. The discipline: every quote in a synthesis cites participant ID and timecode, and the researcher has heard or read the actual quote at that timecode before it appears in the deck.

Can AI estimate sample size or research validity for me?

No, and this is one of AI's most dangerous failure modes for UX research. AI will confidently summarize 4 interviews as 'users want X' — but four participants is four participants, and the AI's confident tone is not statistical confidence. Senior UXRs always re-state sample size in the final synthesis. Sample-size reasoning stays with the human.

Should I feed participant transcripts into Claude.ai or ChatGPT?

Only after stripping PII (names, employer names, contact details). General-purpose chatbots are not a safe destination for raw participant data — data-handling policies vary and your organization may have contractual obligations about data scope. Repository tools like Dovetail with proper data-processing agreements are different. When in doubt, redact, and check with privacy/legal first.

UX Researcher Hub

AI Tools in the UX Research Workflow (2026)

By Blake Crosley · Last verified 2026-04-30

In short

AI tools in 2026 sit alongside the senior UX researcher, not in front of them. Claude and ChatGPT accelerate transcript coding, theme suggestion, and corpus summarization; Dovetail AI auto-tags clips and proposes themes inside the repository; NotebookLM and Claude Projects act as grounded notebook chat over a fixed study corpus. The discipline that keeps research evidence-grade: every AI-suggested theme is verified against raw transcripts, every quote is traced back to a participant, and AI is never trusted to size or weight findings without a human re-reading the source.

Key takeaways

AI accelerates the slow parts of qualitative analysis — first-pass open coding, transcript summarization, theme suggestion — but never replaces the researcher reading raw transcripts. Erika Hall's Just Enough Research stance still holds in 2026: the researcher's judgment is the artifact.
Dovetail AI (auto-tagging, AI summary, theme suggestions) is the most mature integrated AI feature set for UX research repositories in 2026. The pattern that works: let Dovetail propose tags, accept or reject each one manually, and never ship a synthesis the researcher hasn't re-read.
Claude Projects and Google's NotebookLM are the strongest 'notebook over a fixed corpus' tools for research synthesis in 2026 — both ground answers in the documents you provide and surface citations. They are safer for synthesis than open-web chat because the source-of-truth is bounded.
AI is dangerous for sample-size reasoning. Five interviews summarized by AI feel like 'a clear pattern emerged' but five participants is five participants. Senior UXRs guard against AI-induced false confidence by always re-stating sample size in the synthesis.
AI for survey-question generation must be human-reviewed for leading language, double-barreled questions, and assumption-loaded framing. AI is good as a second-pair-of-eyes reviewer of human-drafted screeners; it is bad as primary author.
Hallucinated quotes are the #1 failure mode in 2026. Any quote produced by an AI summary must be verified against the raw transcript before it leaves the researcher's machine. A fabricated quote in a stakeholder readout is credibility-ending.
The human-in-the-loop discipline: researcher drafts the question, AI assists analysis, researcher verifies every claim against raw evidence, researcher writes the final synthesis. AI is leverage on the boring parts; judgment stays human.

How senior UXRs use AI tools in 2026

The senior UX researcher in 2026 has a stable AI workflow across three surfaces: the research repository (Dovetail AI), the long-context notebook (Claude Projects or NotebookLM), and the general-purpose assistant (Claude.ai or ChatGPT). Each is used for a specific job; none is treated as the synthesis itself.

Repository AI (Dovetail). Auto-tagging clips during transcript ingestion, proposing themes once enough tagged clips exist, and generating AI summaries of tags or insights. The researcher reviews every auto-tag before it counts as evidence.
Notebook AI (Claude Projects, NotebookLM). Grounded chat over a bounded corpus: 8 transcripts from a discovery round plus the discussion guide and research brief. The researcher asks "Where did participants describe friction in onboarding?" and the tool returns a grounded answer with citations.
General-purpose AI (Claude.ai, ChatGPT). Drafting analysis-plan outlines, rewriting a research readout for a non-research audience, second-pair-of-eyes review of a screener for leading questions. Never the primary synthesis tool.

The pattern that matters: the researcher reads at least the raw highlight reel of every interview before any AI summary is allowed near a stakeholder readout. Tomer Sharon's research-craft writing (tomersharon.com) and the NN/g coverage of AI in research (nngroup.com/articles/ai-ux-research) land on the same conclusion: AI is a force multiplier for the experienced researcher and an evidence-grade hazard for the inexperienced one.

AI-assisted synthesis: what it does well, what fails

AI-assisted synthesis in 2026 is genuinely useful for the parts of qualitative analysis that are slow without being judgment-heavy.

What AI does well:

First-pass open coding. Given a transcript, AI can propose 20–30 candidate codes (e.g., "frustration with onboarding wizard", "trust in pricing transparency"). The researcher accepts, rejects, or merges. Time saved: hours.
Cross-transcript pattern surfacing. Dovetail AI and NotebookLM can surface "this theme appears in 6 of 8 interviews" once tags or chunks are in place. The researcher verifies every count.
Transcript summarization. A 60-minute interview becomes a 400-word summary with key quotes. Useful for stakeholder previews; not a substitute for the researcher reading the transcript.
Theme labeling. Given 30 quotes that share a meaning, AI is good at proposing 3–5 candidate theme names. The researcher picks the one that maps to participant language, not corporate jargon.

What AI fails at:

Sample-size reasoning. AI will confidently say "users want X" from 4 interviews. The researcher must always re-state sample size in the final synthesis.
Theme overfit. AI tends to over-cluster — collapsing genuinely distinct experiences into a single theme because the surface language is similar. The fix: read at least one full transcript per proposed theme before accepting it.
Edge-case loss. The most interesting research finding is often the outlier. AI summarization smooths outliers into the average. The researcher must explicitly hunt for the dissenting voice.
Hallucinated quotes. AI summaries have produced quotes that no participant said. Every quote in a stakeholder readout must be verified against the raw transcript. This is non-negotiable.

The discipline: AI proposes, researcher disposes. The Dovetail blog (dovetail.com/blog) and Anthropic's Claude Projects documentation (anthropic.com/news/claude-projects) both emphasize the human-verification loop.

Where AI helps with study design vs where it's risky

Study design — the part of research that happens before any participant is recruited — is where AI is genuinely helpful as a reviewer and dangerous as an author.

Where AI helps in study design:

Second-pair-of-eyes on screeners. Paste a screener into Claude with the prompt "Identify any questions that would screen out a key segment, any leading language, or any compliance issues." This often catches a question that would have wasted recruiting budget.
Survey-question critique. AI is consistently good at flagging double-barreled questions, leading questions, assumption-loaded framings, and missing "I do not know" options — better than most junior researchers.
Discussion-guide pacing review. AI can identify sections of a guide that will run long, places where the participant will likely shut down, and questions that depend on a previous answer the participant might not give.
Research-plan structure check. "Does this plan have a clear research question, a defensible method, and a synthesis approach that produces decisions, not just findings?"

Where AI is risky as primary author:

Generating survey questions from scratch. AI tends to produce questions that are grammatical but research-naive — leading, assumption-loaded, and biased toward the product team's assumed outcome. Erika Hall's framing (muleshq.com) — that the question you ask shapes the answer you get — applies double to AI-generated surveys.
Discussion guides for sensitive topics. AI does not have a calibrated sense of when a question will cause harm. Sensitive-topic interviewing (financial hardship, medical history, identity) requires human judgment.
Recruiting-screener generation. AI will produce screeners that exclude entire user segments because of subtle phrasing. A human reviewer must catch this before recruiting kicks off.

The principle: AI is a fast, cheap reviewer of human-drafted study materials and a poor primary author. The senior UXR uses AI to catch their own blind spots; they do not delegate the question-asking to the model.

Guardrails: hallucinated quotes, theme overfit, ethics

The 2026 UX research practice has developed a stable set of guardrails for AI-assisted research. The senior UXR who keeps their work evidence-grade has these as habits, not rules to remember.

Verify every quote. Any quote in a synthesis must be traceable to a specific participant in a specific transcript. If the quote came from an AI summary, the researcher reads the raw transcript at the cited timecode before the quote ships.
Re-state sample size. Every AI-generated summary should be re-prefaced by the researcher: "Based on 8 interviews with users in segment X…" The AI will not do this on its own.
Hunt the dissenting voice. After AI surfaces the dominant theme, the researcher asks: "Which participant disagreed? Whose experience does this theme not describe?" The outlier is often the most decision-changing finding.
Read one full transcript per theme. Before accepting an AI-proposed theme, the researcher reads at least one full transcript that the theme claims to summarize. This catches theme overfit before it ships.
Never feed PII into general-purpose AI. Strip participant names, employer names, and contact details before any transcript goes into Claude.ai or ChatGPT. Repository tools with proper data agreements are different; general-purpose chatbots are not a safe destination for participant PII.
Disclose AI assistance in research artifacts. The synthesis should note that AI tooling assisted transcript coding and theme surfacing, and that all themes and quotes were verified by the researcher against raw transcripts.
Keep the raw evidence accessible. Every stakeholder readout should link back to the raw clips and transcripts in the research repository. The synthesis is interpretation; the evidence is the evidence. The Nielsen Norman Group coverage of AI in UX research (nngroup.com/articles/ai-ux-research) emphasizes this traceability discipline.

The discipline that separates senior from junior in 2026 is not whether they use AI tools — most do. It is whether they use them without losing the through-line from claim to evidence. AI raises the bar on the discipline required to keep the researcher an honest broker.

Frequently asked questions

Should I use AI to do thematic analysis on my interviews?: Yes for first-pass open coding and theme surfacing; no for the final synthesis. The pattern that works in 2026: AI proposes 20–30 candidate codes per transcript, the researcher accepts/rejects/merges, then AI surfaces cross-transcript patterns once tagged clips exist. The researcher reads at least one full transcript per proposed theme before accepting it. AI is the accelerator on the slow parts; judgment about what the data means stays with the researcher.
Is Dovetail AI good enough to replace manual tagging?: No, but it's good enough to make manual tagging fast. In 2026, Dovetail AI auto-tags clips on ingestion and proposes themes once enough tags exist. The pattern senior UXRs use: let Dovetail propose, then accept or reject each tag manually. Time saved is real but the verification loop is non-negotiable.
Can I use ChatGPT or Claude to write survey questions?: Use AI as a reviewer of human-drafted survey questions, not as the primary author. AI is consistently good at catching double-barreled questions, leading language, assumption-loaded framings, and missing 'I don't know' options. AI is consistently bad at writing surveys from scratch — it produces grammatical but research-naive questions biased toward the product team's assumed outcome. Erika Hall's framing on question-asking applies: the question shapes the answer.
What is Claude Projects and why does it matter for UXRs?: Claude Projects (anthropic.com/news/claude-projects) is a notebook-style workspace where you upload a fixed corpus — transcripts, the discussion guide, the research brief — and chat over it with citations. For UX research, it is vastly safer than open-web chat because the source-of-truth is bounded. Researchers ask grounded questions like 'Where did participants describe friction in onboarding?' and get answers tied back to the corpus. Same workflow as NotebookLM (notebooklm.google.com); pick on data-handling preferences.
How do I prevent AI from hallucinating quotes in my readouts?: Verify every quote against the raw transcript before it ships. There is no prompt that fully prevents hallucinated quotes; the only reliable guardrail is the researcher's verification loop. The discipline: every quote in a synthesis cites participant ID and timecode, and the researcher has heard or read the actual quote at that timecode before it appears in the deck.
Can AI estimate sample size or research validity for me?: No, and this is one of AI's most dangerous failure modes for UX research. AI will confidently summarize 4 interviews as 'users want X' — but four participants is four participants, and the AI's confident tone is not statistical confidence. Senior UXRs always re-state sample size in the final synthesis. Sample-size reasoning stays with the human.
Should I feed participant transcripts into Claude.ai or ChatGPT?: Only after stripping PII (names, employer names, contact details). General-purpose chatbots are not a safe destination for raw participant data — data-handling policies vary and your organization may have contractual obligations about data scope. Repository tools like Dovetail with proper data-processing agreements are different. When in doubt, redact, and check with privacy/legal first.

Sources

About the author. Blake Crosley founded ResumeGeni and writes about UX research, hiring technology, and ATS optimization. More writing at blakecrosley.com.