AI UX Research Without Invented Evidence

AI UX research should organize and interrogate evidence, not manufacture it. Keep source material, generated interpretations, and new hypotheses in separate layers. Reported findings must trace to existing material; unsupported possibilities remain hypotheses.

This applies to interview clustering, support summaries, and synthetic perspectives. Interviews and support conversations can supply human-origin evidence. Synthetic perspectives supply model output. Combining them in a document does not make their evidentiary status equivalent.

Define an evidence contract

State the question, available sources, relevant population, collection context, and exclusions. Classify attitudes, observed behavior, product records, and synthetic exploration.

Interview excerpts support accounts of what those participants described. They do not establish population prevalence without another basis. Support archives represent people who contacted support, not everyone who encountered an issue.

Name out-of-scope decisions. Onboarding expectations should not quietly become retention claims. Limits help reviewers challenge generated synthesis rather than accepting its fluent framing.

Nielsen Norman Group recommends supplementary hypothesis generation for synthetic users, not replacement of real research. Preserve that distinction when choosing whether AI is processing evidence or generating a perspective.

Prepare a traceable analysis set

Assign stable source identifiers and preserve question and task context. Isolated sentences can lose qualifications that change their meaning. Keep these relationships available.

Use approved data-handling practices. Remove unnecessary personal details and review consent, access, and organizational policies. Record meaningful redactions rather than silently losing interpretive context. Do not assume public visibility gives permission to reuse private research material.

Request proposed themes, source identifiers, excerpts, counterexamples, and uncertainty. Allow “not supported by supplied material.” Instructions help but do not guarantee accuracy; reviewers must verify output.

Synthesize without erasing disagreement

Treat generated themes as candidate groups. Two mentions of “time” can mean loading delay, lengthy setup, or scheduling trouble. Similar vocabulary is not necessarily one design problem.

Actively inspect contradictions. Detailed setup guidance might help someone unfamiliar with the category and frustrate an expert. Preserve differences before recommending universal simplification.

Do not let executive summaries transform partial evidence into consensus. Say “in the supplied interviews” when that is the scope. Calculate useful counts from auditable units, not generated percentages. Never accept invented quotes or participant totals.

A reviewer should be able to disagree productively with the analysis. That requires reasoning and context rather than polished certainty. Distinguish a suggested design response from a research finding: “explain reversibility” is a proposal, not an observation that the proposal works.

Hypothetical example: onboarding uncertainty

Hypothetical example, not a reported result: A team supplies approved, de-identified onboarding excerpts with identifiers. It asks AI for themes without adding participant statements.

The model suggests “users dislike configuration.” Reviewing originals reveals a narrower issue: some accounts concern which settings are reversible. Other excerpts value controls. The broad theme is rejected because it suppresses a distinction.

The synthesis records uncertainty about reversibility within particular supplied accounts. Explaining which settings can change becomes a design proposal, not a tested solution.

Next, Cairo explores a proposed public explanation across synthetic profiles. Generated reactions enter a separate hypothesis section. They are not added to interview evidence, quoted as testimony, or counted as participants.

The follow-up plan asks actual users to explain what they expect after setup and whether they can find how to change a setting. The team could discover the problem lies in navigation rather than explanatory copy. The synthetic lane should remain open to that contradiction instead of determining the outcome in advance.

Add simulation as a separate lane

Cairo explores URLs or ideas with demographic, psychometric, and socioeconomic profiles, bounded journeys, audiences, and reports. Here, it broadens considered questions rather than filling missing interviews.

A synthetic objection about costs can inspire a study question. It cannot establish that an interview participant raised the objection. Keep inputs identifiable, comparisons consistent, and outputs labeled generated interpretation. Rosetta explains the behavioral modeling context.

Cairo does not recruit humans, measure real conversions, certify accessibility, or make live changes. It belongs beside source-based analysis, not inside human-evidence counts.

A synthesis review checklist

• Define the question and evidence boundary.
• Confirm permitted source use.
• Preserve identifiers and meaningful omissions.
• Verify quotations and substantive attributions.
• Review negative cases and competing interpretations.
• Separate findings, recommendations, and hypotheses.
• Keep synthetic output outside human counts.
• Assign unresolved questions to a next study.

AI can create convincing unsupported connections and hide minority perspectives in broad themes. Check framing as well as facts. A real quotation is not sufficient if its context contradicts the claim.

Use the audience evaluation framework for panel assumptions and the AI usability review for task-focused follow-up.

The goal is not a more certain-sounding report. It is an analysis whose claims remain inspectable, proportionate, and useful for decisions. Faster synthesis is valuable only when it preserves what the evidence can actually support.

Next step: Explore your public page or product idea in Cairo, define one question and an audience, and use the simulated report to plan what to investigate next.