Back to docs
Interview Techniques

Recipient Design: Who Your Participant Thinks They Are Talking To (2026)

Every answer is designed for an audience. The audience your participant pictures is set by cues you control, and changing it changes what they disclose, even when the interviewer and the channel stay identical.

People do not produce answers in a vacuum. They design each answer for a specific listener, and the listener they have in mind is not necessarily the one you intended. Conversation analysts call this recipient design; sociolinguists studying the same effect call it audience design. Either way it is one of the few variables in interview research that is invisible in the transcript, large in the data, and almost entirely under your control.

The headline for a research team: the participant's belief about who is listening changes what they disclose, independently of the interview channel and independently of who is actually listening. That independence is what makes it a separate lever from mode, and it is the reason two studies can run on identical software with identical questions and produce different answers.

The cleanest evidence available: belief with everything else held constant

Most comparisons of research settings change several things at once. A phone survey and a face-to-face survey differ in channel, in interviewer presence, in pace and in privacy, so when the answers differ you cannot say which difference did it.

One study isolates the variable almost perfectly. Gale Lucas, Jonathan Gratch, Aisha King and Louis-Philippe Morency published It is only a computer: Virtual humans increase willingness to disclose in Computers in Human Behavior in 2014 (volume 37, pages 94 to 100). Participants were interviewed by an on-screen virtual human during a health screening. Every participant saw the same virtual interviewer. The only thing that changed was what they were told about who was behind it: some were led to believe a human operator was controlling it, others that it was fully automated.

The result, in the authors' own words: "compared to those who believed they were interacting with a human operator, participants who believed they were interacting with a computer reported lower fear of self-disclosure, lower impression management, displayed their sadness more intensely, and were rated by observers as more willing to disclose."

Nothing about the interface changed. The channel was constant, the questions were constant, the face on the screen was constant. What moved was the participant's model of the recipient, and that alone moved fear of disclosure, impression management, emotional expression and observer-rated openness. The paper has since been cited several hundred times, and it remains the sharpest demonstration that the audience is a design variable rather than a confound.

The audience is usually plural

The mistake teams make is assuming there is one recipient. There are normally at least three, and a participant is managing all of them at once.

AudienceWho it isWhat the participant does about it
The immediate recipientWhoever or whatever is asking right nowAdjusts tone, length, vocabulary, how much context to supply
The overhearing partyWhoever they believe will read the transcript laterCensors specifics, avoids naming colleagues, softens criticism
The self as audienceTheir own self-imageAvoids answers that would require admitting something unflattering

Roger Tourangeau and Ting Yan, reviewing the survey literature in Psychological Bulletin in 2007, describe respondents editing what they report "to avoid embarrassing themselves in the presence of an interviewer or to avoid repercussions from third parties." Those are two different audiences inside one sentence, and they need different interventions. Removing the interviewer does nothing about third parties. Promising confidentiality does nothing about the person in the room.

The third row is easy to overlook and is why anonymity alone underperforms. A fully anonymous participant is still talking to themselves.

What actually sets the picture

The participant builds their model of the audience from a handful of cues, most of which are set before the first question is ever asked. In rough order of influence:

  1. Who the invitation says the study is from. An invite from a named product team reads very differently from one sent by a research platform. This single line does more work than anything in the guide.
  2. What they are told happens to their answers. Whether the transcript is attributed, anonymised, quoted, or shared internally. Vagueness here is not neutral - participants fill the gap with the most cautious assumption.
  3. Whether a person is present. Live moderation implies an immediate social audience with reactions to manage.
  4. The framing of the opening question. A first question that asks for an evaluation signals you are here to grade our work. One that asks for a story signals you are here to describe your week.
  5. Incentive framing. If the reward appears contingent on the sponsor being pleased, the sponsor becomes the recipient regardless of what the consent screen says.

Note how few of these are properties of the interview software. Recipient design is mostly set in the recruitment and framing layer, which is why it survives a change of tool and why it is so often left unexamined.

How this differs from three things it is often confused with

This distinction matters, because the three are usually managed separately and the fixes do not transfer.

It is not mode. Mode is the channel - voice against text, phone against web. Recipient design is the participant's belief about the listener. The Lucas study holds the channel fixed and moves only the belief, which proves the two are separable. If you want the channel question, see mode effects and voice vs text interviews.

It is not interviewer bias. Interviewer bias is variance introduced by what the moderator actually does - their wording, their reactions, their follow-ups. Recipient design operates even when the interviewer does nothing at all, because it lives in the participant's head. A perfectly consistent interviewer still constitutes an audience.

It is not epistemic territory. Epistemic territory concerns whether a question is asking about something the participant is entitled to know. Recipient design concerns who they are telling. A question can sit squarely inside the participant's territory and still get a managed answer because of who is listening.

Making it a controlled variable instead of an accident

The practical goal is not to eliminate recipient design - that is impossible, since every answer is designed for someone. The goal is to fix it deliberately so it does not vary across your sample, and to set it where you want it.

Hold it constant across the sample. If half your participants were recruited through a support ticket and half through a public panel, they are imagining different audiences and their answers are not comparable. This is a real threat to any study that mixes recruitment channels, and it is invisible unless you look for it.

State the audience explicitly. Tell participants exactly who reads this and in what form. A specific, slightly less flattering truth outperforms a vague reassurance, because a vague reassurance is replaced by the participant's own worst-case guess.

Separate the sponsor from the asker where it matters. For high-stakes evaluative questions, the answers improve when the participant is not addressing the party being evaluated.

Record the framing as a study variable. Put the invitation text and the confidentiality language in the study record alongside the guide. When two waves of a tracker disagree, this is one of the first things worth checking, and most teams cannot check it because nobody wrote it down.

How Koji handles this

Koji gives you the pieces to set the audience deliberately rather than by accident.

  • No moderator by default. Koji runs interviews without a human present, which removes the immediate social audience - the exact manipulation that moved disclosure in the Lucas study.
  • One fixed framing for every participant. Because Koji asks every participant the same questions in the same words from the same study setup, the audience cues do not drift between the first interview and the fiftieth the way they do across a week of moderated sessions.
  • Explicit data handling you can state up front. Koji supports anonymising transcripts, so you can tell participants something concrete about the overhearing audience instead of leaving them to guess.
  • Six structured question types for the items where framing matters most. open_ended, scale, single_choice, multiple_choice, ranking and yes_no are all available, and a sensitive magnitude collected as a scale or yes_no item carries less audience pressure than the same thing said aloud. The structured questions guide covers how each is asked and analysed.
  • Voice and text in one study. Participants who would manage their image more carefully aloud can answer in text, and because the wording is identical the two remain comparable.
  • Every transcript retained verbatim. When you suspect an audience effect, Koji keeps the raw answers so you can compare how participants recruited through different channels actually phrased things.

An honest limit: Koji cannot control what your invitation email says, and that is the single most influential cue in the whole list. The platform sets the interview conditions; the recruitment framing is still yours to get right.

Frequently asked questions

What is recipient design in research interviews?

Recipient design is the way people shape an answer for the particular listener they have in mind. In research it means the participant is not simply reporting facts, they are producing an account calibrated for an audience - the interviewer, whoever they think will read the transcript, and their own self-image. Because the calibration happens before the words come out, it is invisible in the transcript and has to be managed at design time.

Is this just social desirability bias under another name?

No. Social desirability bias describes the direction of the distortion, which is toward whatever looks good. Recipient design describes the mechanism and, crucially, the input variable: which audience is being addressed. Two participants can both be managing their image and distort in different directions because they picture different listeners. Naming the audience gives you something to change, where naming the bias only gives you something to regret.

Does an AI interviewer remove the audience effect?

It removes one audience and reduces the pressure from another. The evidence on believed-automated interviewers shows lower fear of self-disclosure and higher observer-rated disclosure when participants think no person is behind the questions, and Koji runs without a moderator by default. Participants still know an organisation will read the results, so the overhearing audience remains and still needs to be addressed through what you tell them about data handling.

Can recipient design affect a study where nothing is sensitive?

Yes, and this surprises teams. Audience shapes how much context participants supply and how long their answers run, not only whether they tell the truth. Someone who believes they are addressing an expert insider will skip explanation that someone addressing a general audience would spell out. That changes the richness of your qualitative data even when nobody has any reason to conceal anything.

How do I tell whether recipient design is affecting my results?

Compare answers across recruitment sources for the same question. If participants recruited from a support queue systematically differ from panel participants in phrasing, length or willingness to criticise, and the questions were identical, the audience cue is the most likely explanation. This is why holding recruitment framing constant, and writing it down, is worth the small effort it costs.

What is the single highest-leverage thing to change?

The invitation. It sets who the study appears to be from, what participants believe happens to their answers, and whether the sponsor is also the asker - three of the five cues that build the audience picture, all decided before anyone reaches your carefully written guide.

Related Resources

Related Articles

Anonymizing Customer Interview Data: A Practical Guide for Privacy-Safe Research

Five operational techniques for handling PII in AI customer interviews — from intake-time anonymization to stakeholder-safe quote sharing — without sacrificing research signal.

Epistemic Territory in Interviews: Ask People About What Is Theirs to Know (2026)

Participants answer questions outside their knowledge as fluently as inside it. How to map the epistemic territory of each question and read agreement tokens as evidence.

Interviewer Bias: How Moderators Distort Research (and How AI Removes the Variance)

Interviewer bias is the distortion caused by a moderator's wording, reactions, expectations, and characteristics. Learn the types, the evidence, mitigation techniques, and why an AI interviewer eliminates interviewer variance.

Mode Effects: When Letting People Choose Voice or Text Changes the Answer

Pew randomly assigned 3,003 people to phone or web and got answers that differed by up to 18 points on identical questions. Here is what that means when your respondents pick their own mode.

Social Desirability Bias: What It Is and How to Eliminate It in Research

Social desirability bias makes people tell you what sounds good instead of what is true. Learn what causes it, why it quietly wrecks product decisions, and the seven evidence-based ways to reduce it — including why AI-moderated interviews get more honest answers.

Structured Questions in AI Interviews

Mix quantitative data collection — scales, ratings, multiple choice, ranking — with AI-powered conversational follow-up in a single interview.

Voice vs Text Interview: When to Use Each Mode

Choosing between voice and text mode for your AI interview? This guide breaks down response depth, completion rate, audience fit, and cost — plus a decision matrix that tells you which mode wins for each research scenario.