In research, we spend a lot of time thinking about who is in the sample and what questions we ask.

We obsess over quotas, tweak discussion guides, and debate whether a question is leading, loaded or just a bit clunky – and rightly so. If we aren’t speaking to the right people, or asking them the right things, we’re unlikely to get to the right answers.

But we spend far less time thinking about who (or what) is doing the asking.

Rapport, trust and moderator skill can make or break a conversation. We already think carefully about a moderator’s skill: whether they can build rapport, probe well and create a comfortable space. But how often do we stop to think about how the moderator shapes that space, and what their presence might change about what participants feel able to say?

We wanted to explore whether matching participants with their moderator by key identity markers like ethnicity or gender could help create more comfort, more psychological safety and, ultimately, deeper insight.

Using Quasai, Simpson Carpenter’s live AI humanoid avatar moderator, allowed us to isolate and compare different moderator conditions more consistently than would have been possible with multiple human moderators.

The core of our experiment was simple: does who asks the questions change the answers?

The experiment

The study involved 289 UK adults from Black/Afro-Caribbean, South Asian, East/Southeast Asian, White/Caucasian and Middle Eastern/Mixed-Heritage backgrounds, split across three arms. In one group, participants spoke with a voice-only AI moderator. In another, participants chose from a range of different racial appearances for the same AI avatar moderator. In the third, participants were assigned an AI avatar moderator matched to them by ethnicity.

The topic was inclusion in the UK. Participants were asked about their own experiences, their perceptions of inclusion, and how the moderator made them feel. At the end, they were invited to reflect on whether the moderator’s identity or presentation had influenced their responses.

What we found

The findings were not neat and tidy, which is probably why they were useful.

First, AI itself appeared to create a different baseline for comfort. Across groups, participants described feeling less judged, more able to be themselves and less concerned with giving the ‘right’ answer. For some, the absence of a human interviewer seemed to reduce the pressure to perform, impress or edit themselves.

That doesn’t mean AI automatically creates better research, but it does suggest that, for some sensitive topics, AI moderation can lower the threshold for honesty and help participants move more quickly into reflective, personal or even confessional territory.

Second, visible matching mattered, but not in a simplistic way. For some Black and Asian participants, being matched with an avatar who “looked like me” created an immediate sense of ease and recognition. The matched moderator seemed to reduce some of the emotional labour of the interview including the need to anticipate judgement or how much of themselves to reveal.

Among some White participants, matching created a sense of recognition with a different effect: reduced self-censorship. Several participants acknowledged that they may have chosen their words more carefully, or held back more, if the moderator had looked different to them – particularly when discussing diversity and inclusion. Perceived similarity revealed how quickly people become aware of what they should (or shouldn’t) say when the topic feels sensitive.

The matching effect was also reflected in the data itself. Avatar-led conditions generated longer, more detailed responses than the voice-only baseline, suggesting that visible moderator presence helped participants expand, explain and disclose more. This effect was even stronger in interviews where the moderator was matched to participant. Engagement also moved in the same direction, with participants in matched conditions more likely to say they felt understood and found it easy to explain themselves.

That doesn’t mean participants are only willing to speak to someone who looks like them. It highlighted the tiny calculations people make in any conversation: how much to explain, how much to edit, how vulnerable to be, and whether the person asking the questions is likely to understand the answer.

Third, the study showed the risk of superficial representation. For experimental control, voices and accents were held constant across avatars. This meant some participants noticed a disconnect between how the avatar looked and how it sounded. For some minoritised participants, that felt jarring rather than inclusive.

This is a vital lesson as AI avatars become more common. Avatar design isn’t just cosmetic, voice isn’t neutral, and ‘default’ definitely isn’t neutral either. Representation only supports inclusion when it feels coherent, credible and culturally sensitive.

So what?

For the research industry, the implication is not that every participant should always be matched with a moderator who looks like them. Our experiment demonstrated what superficial matching could already unveil – as humans we have the agency, empathy and responsibility to move beyond that, to help us get closer to the truth.

When it comes to moderator matching, we should ask ourselves: what characteristic is most relevant to the research objective, and where might shared experience change what participants feel able to say?

More broadly, we need to stop assuming that the research environment is neutral by default.

Matching can deepen insight, but only when used deliberately. A diverse sample is not the same as an inclusive research experience. If a participant feels guarded, misunderstood or emotionally exposed, we may capture the answer they feel able to give, rather than the answer that reflects the full truth of their experience.

Used thoughtfully, tools like Quasai can help us test what creates comfort, where people hold back, and how different modes of participation affect the depth of response.

The bigger lesson is not about replacing human researchers, it is about reminding us how much human context matters.

Participants are making judgements about who is listening, what they might understand and how their answers might be received, which means diverse research teams are not a nice-to-have. They are part of the method: shaping the questions we ask, the moderators we assign, the assumptions we challenge and the meaning we take from what participants tell us. With the aid of technology, we can continue to expand what’s possible in conducting inclusive and ultimately more insightful research.

Participants are always reading the room, even when the room is digital or AI-led. As researchers, we should be reading it too.

Annabelle Jones is senior research manager at Simpson Carpenter