Short answer: Choose between focus groups and interviews by asking who the disclosure risk comes from, not how sensitive the topic is. When participants share the stigma, groups surface more sensitive material (Guest et al., 2017). When the risk comes from the people in the room, groups suppress it entirely (Namey et al., 2023). Per participant, interviews still generate broader content.
\nLast updated: July 17, 2026.
\n\nAsk most researchers when to use focus groups instead of interviews and you get a version of the same rule: sensitive topics go one-on-one.
It is in the textbooks, it is in the training decks, and it is the default that method-choice conversations fall back on when the schedule is tight.
The rule is intuitive. It is also contradicted by one of the only randomized comparisons ever run on the question — and confirmed by another rigorous study that reached the opposite conclusion. Both are good studies. The contradiction between them is the most useful thing in this literature, because resolving it produces a rule that actually predicts which method works.
\nThis guide walks through what the evidence says on focus groups vs interviews, why the two headline studies disagree, and what changes now that AI moderation has removed the cost argument that used to decide the question by default.
\n\nThe strongest comparison available is a randomized study by Greg Guest, Emily Namey, Jamilah Taylor, Natalie Eley and Kevin McKenna, published in the International Journal of Social Research Methodology in 2017. The design is unusually clean: 350 African-American men in Durham, North Carolina were randomized into either a focus group arm or an individual interview arm and asked the same open-ended questions about health-care-seeking behavior (Guest et al., 2017).
\nTwo findings matter.
\nFirst, on breadth: at the event level, focus groups and interviews generated similar numbers of unique items in a free-listing task. But per person, individual interviews were more productive. A focus group only kept pace by pooling eight-or-so people into one session — each individual contributed less than they would have alone.
\nSecond, and against the textbook rule: several types of sensitive and personal disclosure were more likely in the focus group setting, with some sensitive themes appearing only in focus groups. Notably, no sensitive theme emerged more frequently in the interviews. The authors also note these disclosures were unsolicited — participants volunteered personal information rather than answering a direct question about it.
\n\nNow hold that against a systematic comparison published in Evaluation and Program Planning in 2023, which tested method performance for eliciting sensitive information in a USAID justice study in Haiti, comparing 16 focus group discussions with 46 individual interviews against a shared elicitation guide (Namey et al., 2023).
\nThe result inverts Guest et al. Participants in individual interviews were statistically more likely to raise themes relevant to gender-based violence, and sensitive themes around sexual violence, economic dimensions and restorative justice did not emerge in the focus groups at all. The authors conclude that prioritizing interviews may be appropriate for socially sensitive topics.
\nSo: one rigorous study says groups pull more sensitive material out; another rigorous study says groups make it vanish. Both cannot be a general law about sensitivity.
\n\nHere is our reading, and it is an interpretation rather than a finding either paper states: the variable that flipped was not how sensitive the topic was. It was where the social risk of disclosing was located.
\nIn the Durham study, the room was full of peers who plausibly shared the experience being discussed. Admitting you avoid the doctor in front of other men who avoid the doctor is not a confession — it is joining in. The group functions as permission. One person going first lowers the cost for everyone, which is exactly the mechanism that produces unsolicited disclosure.
\nIn the Haiti study, the topic was gender-based violence in a community context. The group is not an audience of fellow-sufferers who normalize the disclosure; the group is the community that could judge, gossip, or retaliate. The risk of speaking is carried by the people in the room. Under those conditions, silence is not shyness. It is a rational assessment, and no amount of moderator skill talks someone out of it.
\nThat gives a test that is more useful than \"is this sensitive?\", because almost everything worth researching is sensitive to someone.
\n| Ask | If yes → | Why |
|---|---|---|
| Does disclosing carry social risk from the other participants (judgment, gossip, retaliation, professional exposure)? | Interviews | The group is the threat. Themes will not merely be muted — they will be absent, and you will not know they are missing. |
| Do participants share the experience, so that disclosure reads as solidarity rather than confession? | Focus groups | Peers give permission; one disclosure unlocks others. This is where groups beat interviews on sensitive content. |
| Is the object of study the interaction itself — norms, shared vocabulary, how people argue each other into positions? | Focus groups | Interviews cannot produce data that only exists between people. |
| Do you need maximum content per participant (rare or expensive recruits, small N)? | Interviews | Per-person breadth is where interviews reliably win. |
| Are participants professional peers, competitors, or in a reporting line with each other? | Interviews | A special case of row 1 that B2B research walks into constantly. |
The test is diagnostic, not a scorecard: row 1 is a veto. If disclosure risk sits inside the room, no number of yeses further down the table rescues the focus group.
\n\nHistorically, method choice was often settled before any of this came up. Focus groups were how you talked to 40 people on a budget, because a moderator's hours were the binding constraint. Eight people per session was arithmetic, not methodology.
\nAI moderation weakens that constraint: sessions run asynchronously and in parallel, so 40 interviews no longer cost 40 moderator-hours. The honest consequence is not that focus groups lose — it is that you no longer get to pick one because the other was unaffordable. If the Guest per-person finding holds, and interviews are no longer priced out, the default for exploratory breadth arguably shifts toward interviews, with focus groups reserved for cases where interaction is the point.
\nWhat AI does not change: the disclosure calculus in row 1. Participants suppress information because of who else is present. Automating the moderator does not remove the other participants. Claims that AI moderation eliminates social-desirability bias should be read carefully — an AI moderator can be more consistent and never gets bored, but in a group setting the audience is still there.
\n\nQualitati is an AI user research platform for product, UX and customer insights teams, built to run either side of this decision rather than to sell you one of them.
\nPricing is published per credit, with 30 credits free on signup and no credit card required.
\n\nRead this section before acting on the framework.
\nThis is for researchers, product teams and insights leads choosing a data-collection method for a study that is already scoped. Do not use the Disclosure Risk Test to settle a question you can test cheaply: if you can pilot both arms on a handful of participants, do that instead — the framework is for when piloting both is not on the table. And do not use it for questions where neither method is right. If you need incidence or magnitude, you need a survey.
\n\nNeither, as a general rule — the evidence goes both ways. Guest et al. (2017) found sensitive disclosures more likely in focus groups; Namey et al. (2023) found sexual-violence themes appeared only in interviews. Ask whether the social risk of disclosing comes from the other participants. If it does, interview.
\n\nPer participant, generally yes. In Guest et al. (2017), interviews generated a broader range of items per person; focus groups matched them only at the event level, by pooling several participants into one session.
\n\nIt depends on saturation rather than a fixed number, and the unit is the group rather than the participant. See our guide to data saturation in qualitative research.
\n\nThey can counter some of it — an AI moderator can consistently check consensus and bring in quiet voices without fatigue. But group presence still shapes what people are willing to say. We covered the evidence in do AI focus groups reduce groupthink?
\n\nYes, and it is often the strongest design: focus groups first to map vocabulary and the range of positions, interviews second to go deep on what surfaced. Just do not let the group stage define which themes are “real” — that is precisely what the Haiti study warns against.
\n\nNo. They are useful for exploration — surfacing tensions and sharpening a guide before fieldwork — but they are not evidence about real participants and should not be reported as such.
\n\nThe focus groups vs interviews decision has been carrying a rule the evidence does not support. Sensitivity alone does not tell you which method to use; the location of the disclosure risk does. When participants share the stigma, the group gives permission and surfaces more. When participants are the risk, the group deletes the theme without telling you. And per participant, interviews remain the broader instrument — which matters more now that AI moderation has stopped making them the expensive choice.
\nRun the Disclosure Risk Test before your next study, and let cost be the last input rather than the first.
\nStart free with 30 credits — no credit card required — or view transparent pricing. Run an AI-moderated interview, focus group, conversational survey, or thematic analysis project on Qualitati.
\n","related":[{"slug":"what-is-triangulation-qualitative-research-2026","title":"What Is Triangulation in Qualitative Research?","description":"Triangulation checks findings against more than one source, method, analyst, or theory. A 2026 guide to the four types, when it helps, and where AI fits.","category":"Research Methods","date":"2026-07-18","readTime":"10 min read","author":"Qualitati Research Team"},{"slug":"negative-case-analysis-qualitative-research-2026","title":"Negative Case Analysis: Definition, Steps & Example (2026)","description":"Negative case analysis means searching for data that contradicts your themes, then revising them. Definition, step-by-step loop, worked example and AI tips.","category":"Research Methods","date":"2026-07-06","readTime":"9 min read","author":"Qualitati Research Team"},{"slug":"large-n-qualitative-research-ai-scale-2026","title":"Large-N Qualitative Research at Scale With AI (2026)","description":"Kvale's 1,000-page bottleneck is breaking. Here's how AI runs hundreds of qualitative interviews without abandoning saturation, rigor, or a defensible audit trail.","category":"Research Methods","date":"2026-07-06","readTime":"12 min read","author":"Qualitati Research Team"}]}