Can synthetic interviews establish qualitative saturation?

Name the decision behind the saturation claim

A buyer may want to discover possible objections, understand why customers leave, or decide whether another round of human interviews is needed. Those are different jobs. Ask the supplier to finish this sentence: we have enough evidence to stop doing ___, for this audience and this decision, because ___.

First decide whether saturation is appropriate for the chosen qualitative approach and what the team means by it. In their information-power methods paper, Kirsti Malterud and colleagues argue that sample adequacy depends on the research aim, how specifically the sample fits it, the role of theory, the dialogue and the analysis strategy. Their proposal concerns qualitative interviews with people; it is not a formula for choosing a synthetic record count.

Our recommendation is to agree on the intended claim before generating more material. A screen that supplies questions for a human study can be useful without claiming that customer experiences have been exhaustively understood. This checklist is a proposed buying aid, not a certification standard or a validated stopping rule.

Separate finding a theme from understanding it

In an analysis of 25 human interviews, Monique Hennink, Bonnie Kaiser and Vincent Marconi distinguish identifying the range of codes from developing a richer understanding of what those issues mean. Their study found that identifying codes could precede understanding their meanings; some codes required further interviews, and one still gained new dimensions in the final interview. It did not test synthetic interviews, and its interview counts are not universal thresholds for a buyer to reuse.

A stable list containing 'price' and 'convenience' leaves important questions open. Does price mean the total bill, an unpredictable fee, or poor value for something unused? Does convenience mean less preparation, easier cancellation, or delivery at the right time? Different explanations can imply different product decisions even when the theme label stays the same.

For an analysis that uses a codebook, ask to see how codes were defined and revised, which passages support an interpretation, and where cases disagree. If the team claims depth of understanding, inspect explanations, conditions and exceptions as well as a count of new labels. Do not impose a single code-count rule on every qualitative approach.

Check whose answers the repetition describes

Record whether the material comes from people, AI-generated respondents, or AI helping to interview or analyze people. The AAPOR Code of Professional Ethics and Practices distinguishes AI-generated cases from human research participants and requires disclosure of how generated responses were produced. Applying that distinction here means keeping a generated transcript separate from evidence that a person described an experience.

Generating another interview does not recruit another person. Generated material may suggest new questions, but repetition within it does not demonstrate coverage of the real customers you intended to understand. Check the supplied profiles, source material, interview guide, stimuli and relevant model or service settings. Ask which audiences or experiences those inputs leave unresolved.

Changing prompts or profiles can help explore whether the generated themes depend on the setup. Report that as a check of the simulation. It does not, on its own, establish that the simulation contains every important human perspective. Keep generated explanations labeled as hypotheses until the evidence supports a stronger interpretation.

Fictional example: four themes in 24 generated interviews

Imagine a meal-kit team exploring why people pause a subscription. A supplier generates 24 interviews using profiles of active subscribers. Analysts assign four recurring themes: price, menu choice, preparation effort and packaging waste. The final 12 generated interviews add no new theme labels. All details and counts in this example are invented; they are not a Lewsearch study or a suggested sample size.

The team can describe that observed coding pattern in the generated material. It cannot infer from that pattern alone that it has understood why people pause. The setup used active subscribers, while the business question concerns pausing. A broad label such as 'preparation effort' could also hide different explanations.

In a fictional human follow-up, a paused customer explains that unpredictable work shifts make delivery timing difficult. That could change the next design question without showing how common the problem is. The example illustrates a possible gap in audience and meaning; it does not establish that a particular human interview count will uncover every missing issue.

A defensible next step is to review the audience definition and investigate relevant pausing experiences with people. Keep the generated themes as candidate questions. Repeating the active-subscriber simulation at a larger record count would not, by itself, resolve the audience mismatch.

Request a short evidence record before stopping

Ask the research owner to document these six items. The record should explain the decision rather than decorate it with a large interview count:

• Purpose: what decision will the findings support, and what research activity is being stopped?
• Audience: whose experiences matter, which groups contributed human evidence, and which remain unexamined?
• Method: what analytical approach is being used, and what does 'enough' mean within that approach?
• Material: which records are human or generated, what informed the simulation, and what changed between runs?
• Interpretation: what supports each important explanation, what contradicts it, and what is still only a hypothesis?
• Next evidence: what missing perspective or unresolved explanation could change the decision, who will investigate it, and when will the team reconsider stopping?

If a supplier offers a validation comparison, agree on the qualitative task being evaluated and keep the relevant human findings out of tuning and selection. Matching broad theme names is a narrower result than reproducing explanations, disagreement or decision-relevant exceptions. The blind-validation guide covers outcome separation; it does not certify a qualitative saturation procedure.

Use the human-and-synthetic reporting guide to keep sources and counts distinguishable. A generated quotation should not become a customer testimonial or an additional human participant when it moves into a slide.

Scope the next study to the remaining uncertainty

If the task is to improve an interview guide, synthetic responses can be scoped as exploratory material for a researcher to review. If the decision depends on experiences the current human evidence does not cover, recruit or otherwise investigate the relevant people. If the team knows the broad topics but not why they matter, investigate explanations and exceptions rather than treating another repeated label as closure.

Bring the intended audience, interview guide and stopping decision to a Lewsearch demo. Discuss what an exploratory synthetic exercise could contribute and which findings would still need checking with people. The methods sources cited here do not establish that Lewsearch has validated a rule for ending human qualitative research.