---
name: ai-moderated-interview-design
description: >
  Designs the probe logic for AI-led qualitative interviews: the probe tree, depth
  and stop rules, non-leading follow-ups, handling of thin, evasive and off-topic
  answers, distress and safeguarding escalation to a human, and AI disclosure and
  consent. Use when someone says "design an AI-moderated interview", "we're running
  automated depth interviews", "write the probing logic", "how many follow-ups
  should the AI ask", "how do we stop the AI leading the respondent", or asks
  whether an AI interview can replace a moderated one.
category: 03 Fieldwork and Data Collection
ref: 03.02
tier: 1
inherits: [K2, K3, K4, K5]
---

# AI-Moderated Interview Design

## 1. One-line description
Specifies in advance the probing logic an AI moderator will follow, because it cannot improvise the judgement a skilled human moderator exercises in the room, and sets the depth, stop, safeguarding and disclosure rules that make the resulting transcripts usable as evidence.

## 2. What this skill is used for

**The research problem it solves.** A human moderator does something in the room that looks simple and is not: hears an answer, judges in under a second whether it is thin, defended, rehearsed or genuinely complete, and chooses a follow-up that opens it without suggesting what the opening should contain. An AI moderator has to be given that judgement in advance, as explicit rules, or it defaults to the two failure modes it is structurally prone to. The first is politeness: it accepts the first answer, thanks the respondent, and moves on, producing a transcript that is wide and empty. The second is fluent leading: it generates a follow-up that summarises what it thinks the respondent means and invites agreement, and the respondent agrees, and the transcript now contains the model's hypothesis in the respondent's voice. Both produce material that reads well and cannot be analysed. This skill exists to make the probing logic a designed artefact rather than an emergent behaviour.

**A standing statement on the evidence base.** AI-moderated qualitative interviewing is new. There is a reasonable body of prior work on adjacent questions (interviewer effects, mode effects on disclosure, computer-administered self-interviewing and sensitive disclosure) that bears on it, and a much thinner, faster-moving and largely proprietary body of work on AI moderation specifically. Almost nothing about comparative depth, rapport, or downstream analytic quality should be treated as settled. Throughout this skill, claims about what AI moderation does to responses are stated at the confidence the evidence supports per **K3**, and several are explicitly hypotheses. A researcher using this skill should expect to be validating it, not applying it.

**Where it sits in the research lifecycle.** After the qualitative design exists and the section purposes are agreed, which is the output of **02.02 Discussion Guide Design**. Before fieldwork. It is the conversion of a flexible instrument written for a person into an explicit one written for a system, and it is not a translation exercise: a guide that works for a human moderator is not an AI interview specification.

**Typical use cases.**
- Converting an agreed discussion guide into an AI-moderated interview specification.
- Designing a large-n qualitative study where the volume rules out human moderation and the alternative is open-end survey questions.
- Adding a depth layer to a quantitative study, where the choice is between an AI interview and an unprobed verbatim box.
- Running qualitative in many markets and languages simultaneously where consistency of probing matters more than moderator flair.
- Pre-fieldwork exploration to generate vocabulary and hypotheses ahead of a smaller human-moderated study.
- Auditing an existing AI interview specification that is producing thin transcripts.

**Who uses it.** Qualitative researchers designing or reviewing automated fieldwork; research directors deciding whether a study should be AI-moderated at all; insight managers commissioning it and needing to interrogate what comes back; mixed-methods researchers adding depth at scale.

## 3. When to use it

- A qualitative study will be run without a human moderator and the probing logic has to exist before launch.
- Sample sizes are large enough that human moderation is not affordable, and the realistic alternative is a survey open end rather than a depth interview.
- Consistency of probing across many interviews, moderators or markets is a stated requirement.
- A quantitative instrument needs a genuine follow-up layer on one or two questions rather than an unprobed text box.
- The topic is one where respondents may disclose more to a non-human interlocutor, and the design wants to test that rather than assume it.
- An AI-moderated study has already run and its transcripts are too thin to analyse, and the probe logic needs rebuilding.
- The study will be reported to an audience who will ask how the interviews were conducted and what the AI was and was not allowed to do.

## 4. When NOT to use it

- **The guide itself has not been designed.** This skill takes objectives, section purposes, primary questions and a time budget as inputs. Producing them is **02.02 Discussion Guide Design**, and the boundary is precise: **02.02 decides what the interview is about, in what order, and why; 03.02 decides what happens after each answer.** A probe tree built without an agreed section brief is a machine with no destination.
- **The topic carries a real likelihood of distress, disclosure of harm, or clinical material.** Bereavement, active mental health difficulty, abuse, self-harm, safeguarding-relevant disclosure about a third party, or anything where a participant may need a response rather than a follow-up question. An escalation path (Step 8) is a control for the unexpected case, not a licence to design a study around the expected one. Where distress is foreseeable, the interview needs a qualified human, and the decision belongs to **13.05 Research Ethics and Consent Design**.
- **The value of the session depends on rapport built over time.** Life-history work, longitudinal relationship-based research, interviews with participants who are guarded, marginalised, or have reason to distrust the commissioning organisation. The working hypothesis, held at low to moderate confidence, is that AI moderation loses most in exactly these settings, because what is being lost is the accumulated warrant to ask a harder question.
- **The objective requires reading what is not said.** Hesitation, contradiction between words and manner, the topic the participant steers around, the answer given quickly because it is prepared. Text-based AI moderation has limited access to these signals and no reliable way to act on them. If the analytic value of the study lies there, moderate it with a person.
- **Group interaction is the point.** Focus groups, co-creation, deliberative work and anything where the data is what participants do to each other's views. An AI interview is a set of parallel individual interviews and should never be described as a group.
- **The objective is prevalence or magnitude.** Large-n AI interviewing tempts people to count, because the sample size looks quantitative. It is not a probability sample, the probing differs between respondents by design, and counting coded themes across it produces numbers with all the authority of a percentage and none of the basis. Use **02.01 Survey Questionnaire Design** for the counting and this for the depth.
- **The population cannot be assumed comfortable with, or capable of, an extended typed or spoken exchange with a system.** Low literacy, low digital confidence, cognitive impairment, children, or contexts where the interaction itself will be misread. Mode suitability is a coverage question and it belongs with **03.01 Recruitment and Sample Sourcing**.
- **Disclosure that the moderator is an AI is not permitted or not planned.** Concealment is not an option, for ethical reasons and because the design consequences of concealment cannot be reasoned about. If the study cannot disclose, it cannot run.
- **The purpose is to produce quotes at volume for a report.** Volume of verbatim is not depth, and a large transcript corpus produced by shallow probing is worse than a small one, because its size invites confidence its content does not support.

## 5. Required inputs

**Required. Without these the skill cannot run.**
- **Research objectives and the section brief for each section of the interview**: what each section is for, what a successful section produces, and how it connects back to an objective. This is the handover from **02.02**. Without it, stop and ask.
- **The primary question for each section**, in respondent-facing wording. Probes are designed against a specific opening, not against a topic.
- **Target population, and the sensitivity profile of the subject matter for that population.** Determines safeguarding design, ordering and whether the study should be AI-moderated at all.
- **Mode and interaction format**: typed text, voice, or mixed; synchronous or resumable; device profile. Probe length, wording and fatigue tolerance all differ by mode.
- **The escalation route to a named human**, including availability. Without a real, staffed route, the safeguarding design is decorative and the study should not field.

**Optional, and what each one adds.**
- **Transcripts from human-moderated interviews on the same subject.** The single most valuable optional input. They show what a good answer looks like on each question, what the common thin answers are, and what a skilled moderator actually asked next, which is the raw material for the probe tree.
- **A pilot corpus from an earlier AI-moderated study.** Lets the sufficiency thresholds and stop rules be calibrated against real answers rather than set by assumption.
- **The analysis plan or intended code frame.** Determines what a transcript must contain to be analysable, which is what sufficiency should be defined against.
- **Known vocabulary from prior qualitative work.** Allows probes to use the respondent's terms rather than the client's, which matters more when there is no moderator to notice a term landing badly.
- **Distress and safeguarding policy from the commissioning organisation.** Sets the escalation thresholds and the wording of any signposting.
- **Prior wave specification**, where the study repeats, so probing is comparable rather than incidentally different.

## 6. Questions to ask before starting

1. **What is the analysable unit this interview must produce, per section?** A reconstructed episode, a decision account with alternatives considered, an articulated reason with its own justification, a description of a context. This is the definition of sufficiency, and without it the depth rules have nothing to test against. Default if unanswered: require one specific recent episode per section, described with enough detail to place it in time and context.
2. **Is human moderation genuinely unavailable, or merely more expensive?** The honest comparison is often not AI interview versus human interview but AI interview versus survey open end, and that comparison usually favours the AI interview. Where the real comparison is against a human interview, the losses in Step 2 apply. Default: state which comparison is being made, since it changes what the study should claim.
3. **What is the worst thing a respondent might say in this interview?** Design the escalation for that, not for the average case. Default: assume at least one respondent will disclose something distressing whatever the topic, and require a route.
4. **How long will respondents tolerate, in this population, at this incentive, in this mode?** Unmoderated conversation has no social pressure holding a respondent in it, so the drop-off risk profile is different from a moderated session. Default: design for a shorter interview than the equivalent moderated one and treat that as a constraint on section count, not on depth per section.
5. **Who reads the transcripts, and will they know an AI conducted them?** Bears on disclosure in the report as well as to the respondent (**K4 §7**). Default: disclose in both directions.
6. **What is the fallback if the AI cannot get a usable answer?** Move on, re-ask once in different words, offer an opt-out, or route to a human. Deciding in advance prevents the model improvising a rescue that leads the respondent. Default: one reframe, then move on and record the section as not obtained.
7. **Is this study comparable to a previous wave or to a human-moderated arm?** Probing depth is the variable most likely to differ and least likely to be documented. Default: assume comparison will be attempted and specify probing tightly enough that it could be replicated.

## 7. Step-by-step methodology

**Step 1. Decide honestly what is gained and lost, and record it.** Before designing anything, write the trade for this study. What AI moderation does well: identical probing logic across every interview, no interviewer fatigue and no drift over a long fieldwork period, no interviewer demographic to interact with the respondent, scale that makes qualitative viable at sample sizes where it usually is not, immediate availability so respondents choose the moment, and consistency across markets and languages. What it does not do: earn the right to ask a harder question, notice that an answer is defended rather than complete, follow a genuinely unanticipated line that turns out to be the study's most important finding, read hesitation and manner, or exercise clinical judgement. There is a further effect that runs in both directions and should be treated as a live hypothesis rather than a finding: some respondents disclose more to a non-judging interlocutor, particularly on socially regulated or embarrassing material, which is consistent with the older literature on computer-administered self-interviewing; and some respondents give less, because the effort of articulating something difficult is only worth making to a person who is visibly receiving it. Both directions plausibly operate in the same study on different topics. A correct result at this step is a short written statement of the expected trade, held at stated confidence, which later becomes part of the methodology disclosure.

**Step 2. Convert each section brief into a sufficiency definition.** For every section, write the condition under which the AI may stop probing and move on. This is the most important artefact in the design and the one most often missing. A sufficiency definition names the components an answer must contain, not a length or a probe count. For an episode section: a specific occasion identifiable in time, what the respondent was trying to do, what actually happened, and what they did next. For a reasons section: a stated reason, the alternative that was rejected, and the basis for the comparison. For a context section: where, when, who else was involved, what else was going on. A correct result is a checklist per section that a reader could apply to a transcript and agree on. If two researchers cannot agree whether a given answer meets it, the definition is not written tightly enough for a machine to apply.

**Step 3. Build the probe tree from the sufficiency definition backwards.** For each section, the tree has a fixed shape: an opening question, a small set of **component probes** (one per missing component in the sufficiency definition, fired only when that component is absent), a small set of **deepening probes** (fired when the components are present but shallow), and an exit. Component probes are the workhorses and they are highly specific: "what happened next", "when was that", "what were you trying to do at that point", "who else was involved". Deepening probes are the harder craft: "what made that the option you went with", "you said it was frustrating, what was the frustrating part", "how did you know that". Write the tree so every node has a defined exit, and so no node can be revisited more than once. A correct tree is one where you can trace, for any plausible answer, exactly which probes fire and in what order, and where the total possible probes in a section is bounded and known.

**Step 4. Write non-leading probes, and test them by inversion.** The dominant risk in AI probing is not rudeness but helpfulness: the model summarises, offers a candidate explanation and asks for confirmation. Four rules hold the line. **Reflect only the respondent's own words**, quoted back, never paraphrased into cleaner language, because the paraphrase introduces the model's interpretation and the respondent will adopt it. **Ask for expansion, not for confirmation**: "tell me more about that part" rather than "so it was mainly about the cost". **Never offer a candidate reason, even as an example**, since a single example collapses the response space. **Never evaluate an answer** ("that's really helpful", "that makes sense"), because approval is a training signal about which answers to give more of. The inversion test: for each probe, ask what the respondent would say if the opposite of the implied thing were true, and whether the probe makes that as easy to say. A probe that makes only one direction easy to say is leading, however neutral its adjectives.

**Step 5. Set the depth rules.** Three numbers and one condition, decided per section rather than globally. **Probe budget**: the maximum number of follow-ups in a section, typically small (two to four is a common working range, and this is a design convention rather than a validated optimum). **Minimum probing**: whether at least one probe always fires, even on an apparently complete answer, which is usually worth it on the study's most important section and wasteful everywhere else. **Escalating specificity**: probes go from open to specific, never the reverse, so the respondent has the chance to volunteer before being asked directly. **Stop conditions**, of which there are four and all four must be implemented: sufficiency met; probe budget exhausted; two consecutive non-responsive answers (which means the probe is not working, not that the respondent is unwilling); and explicit or implicit refusal, which is honoured immediately without a further attempt. A correct result is a rule set where the AI cannot loop, cannot exceed the budget, and cannot continue after a refusal.

**Step 6. Classify answer states and specify the handling for each.** The AI must decide what kind of answer it just received before it can choose a probe. Write the classification explicitly, with the response for each state.

- **Sufficient**: components present, specific, self-consistent. Action: exit the section, no thanks, no evaluation.
- **Thin**: on topic, too general to analyse ("it was fine", "convenience mainly"). Action: fire the component probe for the first missing component. This is the commonest state and the one the probe tree mainly exists for.
- **Partial**: some components present, one absent. Action: fire only the probe for the absent component. Do not re-ask what has been answered, which reads as inattention and reduces effort.
- **Off-topic**: answers a different question. Action: one neutral reframe of the original question, without pointing out the mismatch, then move on if it recurs.
- **Evasive or deflecting**: answers in generalities about people in general, or redirects to a safer topic. Action: one probe anchored to a specific occasion, since concrete anchoring is the standard response to abstraction, then accept and record. Do not press twice: persistence after deflection reads as interrogation and the design cannot judge whether the deflection is discomfort or simple lack of interest.
- **Contradictory**: conflicts with something said earlier. Action: surface it neutrally and without implying error ("earlier you mentioned X, and just now Y, can you help me understand how those fit together"), once. Contradiction is frequently the most valuable material in a transcript and it must not be smoothed away.
- **Over-disclosing or distressed**: see Step 8, which overrides everything in this list.
- **Non-responsive or refusing**: action, accept and move on, no repeat.

A correct result is a table a developer could implement and a researcher could audit, with no state falling through to a default of "continue".

**Step 7. Order the sections for an interview with no one reading the room.** A moderator adjusts order live. An AI cannot, so ordering carries more weight and the same contamination rules apply more strictly. Unaided before aided, always, since a stimulus or a named brand cannot be unshown. General before specific, so that the overall account is not shaped by a diagnostic. Behaviour before attitude, so that reasons are attached to something that actually happened. Sensitive material late, so that a break-off costs the core sections nothing. Add one rule specific to unmoderated work: **put the section with the highest analytic value second, not first and not last.** First is where respondents are still calibrating how much effort to give, and last is where fatigue and drop-off concentrate.

**Step 8. Design the safeguarding path, and make it the highest-priority rule in the system.** Specify: the categories of disclosure that trigger it (risk of harm to self or others, disclosure of abuse, acute distress, safeguarding concern about a third party, and any statement suggesting the respondent believes they are speaking to a person who can help them); what the AI does at that moment, which is to stop probing entirely, acknowledge briefly and without therapeutic language, and offer the named route; what the named route is, which must be a real human who is actually available in the fieldwork window, plus signposting to appropriate external support; whether the interview terminates or offers continuation at the respondent's choice; and what is logged, notified to whom, and how quickly. Three prohibitions apply absolutely: the AI never continues probing the distressing material, never offers advice, counselling or reassurance about the substance, and never makes a judgement about severity in order to decide whether to escalate. Under-escalation is the failure that matters here; a false positive costs an interview, and a false negative costs something that is not recoverable. This design is reviewed and signed off by a human before fielding (**K5 §2.4**) and, where the population is vulnerable or the topic sensitive, it is built with **13.05**.

**Step 9. Write the consent and AI disclosure, and place it before the first question.** The respondent is told, in plain language and before consenting: that the interview is conducted by an AI system and not a person; that a person will not be present during it; what is recorded and retained, and whether transcripts are read by humans; how their words will be used and whether verbatim quotes may be reported; that they can stop at any time and skip any question; and how to reach a human, both for distress and for questions about the study. Where the interaction is voice, the disclosure is spoken, not buried in a link. Disclosure is not merely an ethical requirement, it is a methodological variable: how it is worded is likely to affect what respondents say, so it is held constant across the study and reproduced verbatim in the methodology so that anyone comparing studies can see it.

**Step 10. Budget length against unmoderated attention.** Estimate from the section count, the probe budget per section and the mode, then treat the estimate as provisional and calibrate it in pilot. The specific risk in unmoderated work is that the social pressure that keeps a respondent in a moderated session is absent, so abandonment is easier and its cost is total: an interview abandoned at section four yields nothing for sections five and six, and abandonment correlates with the respondents who found the questions hardest, which is not a random loss. Design for this by ordering sections so the most important is reached early, by making the probe budget tighter in later sections, and by allowing the interview to end gracefully at a section boundary rather than mid-probe.

**Step 11. Pilot against real answers and rebuild the tree.** The pilot is not a test that the system works. It is the calibration of the sufficiency definitions and the probe tree against how people actually answer, and it will change both. Run a small number of interviews and read every transcript in full against four questions: did each section meet its sufficiency definition, and if not which component was consistently missing; did any probe lead, judged by whether the respondent's next answer adopted the probe's vocabulary; were there answer states the classification does not cover; and where did probing stop too early or run on. Fix the tree, then run again. A pilot that produces no changes to the probe tree means nobody read the transcripts closely enough. Where a parallel human-moderated arm is affordable even at very small n, it is worth more than a larger AI-only pilot, because it shows what was not asked.

**Step 12. Assess transcript quality before analysis, and disclose the assessment.** Score every transcript, not a sample, on: sufficiency achieved per section; probe count actually fired; evidence of respondent-adopted probe vocabulary (the leading signal); length and elapsed time; abandonment point; and any escalation triggered. Set an inclusion threshold in advance, exclude transcripts below it, and log exclusions with reasons and counts (**K4 §4.4**). Carry the per-section sufficiency record into analysis, because a theme built mainly from sections that did not meet sufficiency is a weaker theme and the analyst needs to know that. This assessment travels with the dataset to **07.03** and its summary belongs in the methodology.

## 8. Analytical framework

Every turn of an AI-moderated interview runs the same loop, and the design is the specification of each element in it:

    Section objective → Sufficiency definition → Opening question → Answer → Answer state classification
        → Probe selection (component or deepening) → Sufficiency re-test → Exit or next probe

Two properties make it work. **The sufficiency definition is the only thing that authorises an exit.** Not respondent length, not the model's sense that the answer was good, not politeness. If the loop can exit on anything else, the transcripts will be thin and the reason will be invisible.

**Classification precedes selection.** The commonest design error is a system that generates a follow-up directly from the answer text, which is exactly how leading probes are produced: the model finds the most salient content and asks about it, which tells the respondent what it found salient. Forcing an explicit classification step between the answer and the probe is what keeps the probe attached to the missing component rather than to the model's reading.

Read backwards from a transcript, the same chain is the diagnostic. A thin transcript is traced to whichever element failed: an opening question that invites a summary rather than an account, a sufficiency definition too loose to bite, a classification that read thin answers as sufficient, or a probe budget of one.

## 9. Output format

**1. Design summary.** Objectives, population, mode, section count, estimated length, and the Step 1 statement of what AI moderation gains and loses in this study, at stated confidence.

**2. Consent and AI disclosure text**, in final respondent-facing wording, in every language the study runs in.

**3. Section specification.** One block per section:

| Element | Content |
|---|---|
| Objective served | |
| Sufficiency definition (components required) | |
| Opening question, verbatim | |
| Component probes, by missing component | |
| Deepening probes | |
| Probe budget | |
| Minimum probing (yes/no) | |
| Stop conditions | |
| Exit rule | |

**4. Answer state handling table.** Every state from Step 6, with the action, the probe class fired, and the maximum attempts. No state defaults to "continue".

**5. Safeguarding specification.** Trigger categories, the AI's immediate behaviour, the escalation route with a named role and availability, external signposting, termination rule, logging and notification. Marked **RESEARCHER SIGN-OFF REQUIRED** per **K5 §3.1**.

**6. Prohibited behaviours list**, written as explicit constraints: no evaluation of answers, no candidate reasons, no paraphrase-and-confirm, no advice, no clinical language, no continuation after refusal, no more than the budgeted probes.

**7. Pilot plan and results**, showing what changed in the tree and why. A pilot section reporting no changes is a defect.

**8. Transcript quality specification**: the per-transcript scoring fields, the inclusion threshold, and how exclusions are logged.

**9. Methodology disclosure paragraph**, in final wording, stating that interviews were AI-moderated, what the moderator could and could not do, the disclosure given to respondents, the probe budget, the escalation provision, and the transcript inclusion rule.

**10. Review points** per **K5 §3**, consolidated at the front and repeated at the point of decision.

**When the inputs are thin**, a section with no agreed sufficiency definition is written as `[sufficiency not defined]` and marked as not ready to field, rather than given a plausible one. Where no human escalation route has been confirmed, the output says the study cannot field, not that escalation is recommended.

## 10. Quality checks

Run before the specification is handed to build. These sit on top of **K4 §8**.

1. Every section has a written sufficiency definition that two researchers would apply the same way.
2. Every probe is traceable to a missing component or to a named deepening purpose. No probe exists because it seemed like a good question.
3. Every probe passes the inversion test: the opposite answer is as easy to give as the expected one.
4. No probe paraphrases the respondent, offers a candidate reason, or evaluates the answer.
5. Every section has a bounded probe budget and cannot loop.
6. All four stop conditions are implemented, including two consecutive non-responsive answers and immediate honouring of refusal.
7. Every answer state has a specified action and a maximum attempt count, and none defaults to continuing.
8. The safeguarding path names a real human with stated availability inside the fieldwork window, and has been signed off by a person.
9. Unaided sections precede any aided or stimulus section, with no exception.
10. The highest-value section is not first and not last.
11. AI disclosure appears before consent, in the respondent's language, and is identical across the study.
12. Length has been calibrated in pilot, not only estimated, and the estimate is labelled as a planning convention until it has been.
13. The pilot changed the probe tree, and the changes are documented.
14. The transcript inclusion threshold was set before fieldwork, not after seeing the results.
15. The methodology paragraph states the probe budget and the escalation provision, so a reader can judge the depth the design permitted.

## 11. Common failure modes

| Failure | How to recognise it | How to prevent it |
|---|---|---|
| **The polite moderator** | Transcripts are wide and empty; almost every section exits after one answer | Sufficiency definitions with named components, and a minimum-probe rule on the key section |
| **Fluent leading** | The respondent's later vocabulary matches the probe's, not their own opening answer | Reflect only the respondent's words; classification before probe selection; inversion test on every probe |
| **Interrogation** | Four or five probes on the same point; respondents shorten answers as the section proceeds | Bounded probe budget, escalating specificity, and hard stops on non-responsiveness |
| **The unbounded tree** | Probes generated freely at runtime with no budget | Every node has a defined exit and a maximum; the tree is a designed artefact, not a prompt |
| **Sufficiency by length** | Long answers exit and short ones get probed, regardless of content | Sufficiency is defined by components, never by word count |
| **Contradiction smoothing** | The AI resolves an inconsistency into the more coherent reading and moves on | Explicit contradictory state with a neutral surfacing probe; **K5 §2.6** flags ambiguity for a human |
| **Escalation as decoration** | A safeguarding path exists on paper with no named, available human | No confirmed route, no fielding. Test the route before launch |
| **Under-escalation** | Distress is probed as content because it was on topic | Safeguarding overrides all other rules; the AI never judges severity to decide whether to escalate |
| **Silent mode mismatch** | The design assumes typed answers and the population answers in three words | Pilot in the actual mode and device profile, and set sufficiency against what that mode realistically yields |
| **AI: invented respondent meaning** | The transcript summary contains conclusions the respondent did not state | **K4 §2.3**. Analysis works from verbatim only; the moderator's own paraphrases are never treated as respondent data |
| **AI: counting the qualitative** | Themes reported as percentages because n is large | Prevalence is reported as counts of participants with the base stated (**K2 §4.2**), and the design's non-probability nature is disclosed |
| **AI: overclaiming equivalence** | The methodology says the interviews were equivalent to moderated depths | State the trade from Step 1 at the confidence the evidence supports, and name what the format could not do |
| **Comparability drift** | Wave 2 uses a revised tree and the depth difference is read as a change in respondents | Version the probe tree, log changes, and treat a tree change as a break in comparability |

## 12. AI guardrails

Skill-specific only. **K4** applies in full and is not repeated here.

1. **Never generate a follow-up directly from answer text without the classification step.** The classification is what keeps the probe attached to a missing component rather than to whatever the model found salient.
2. **Never paraphrase a respondent back to them and ask for confirmation.** Quote their words or ask openly. A confirmed paraphrase is the model's hypothesis wearing the respondent's voice, and it is indistinguishable from evidence downstream.
3. **Never offer a candidate answer, reason or example inside a probe**, including as a clarification of what the question means. If the question needs an example, it is the wrong question.
4. **Never evaluate, praise or thank an answer mid-section.** Approval shapes subsequent answers and the effect cannot be removed at analysis.
5. **Never continue probing after a refusal, a distress signal, or two consecutive non-responsive answers**, whatever the probe budget still permits.
6. **Never make a severity judgement in order to decide whether to escalate.** The trigger categories are met or not met. Escalate on the category.
7. **Never conceal or soften the AI disclosure**, including by using first-person language that implies a person, and never claim capabilities the system does not have (that a person is listening, that help will be arranged).
8. **Never report AI-moderated interviews as equivalent in depth to human-moderated ones**, and never state what AI moderation does to disclosure or rapport as established fact. The evidence base is young; label these as hypotheses with what would test them (**K3 §4.3**).
9. **Never treat the moderator's own generated text as data.** Only respondent turns are evidence. Summaries, reflections and confirmations produced by the moderator are excluded from analysis and from quote extraction.
10. **Never field a specification whose safeguarding route has not been confirmed with a named available human**, regardless of deadline (**K5 §2.4**).

## 13. Best-practice principles

1. **The sufficiency definition is the design.** Everything else in an AI interview specification is machinery serving it. A study with vague sufficiency definitions produces thin transcripts no matter how good the probes are.
2. **Specificity is the antidote to abstraction.** When a respondent generalises, the reliable move is to anchor to one occasion. This is the highest-yielding probe class in qualitative work and it is fully specifiable in advance, which makes it the natural strength of an AI moderator.
3. **Probe the gap, not the content.** A human moderator probes what interests them. An AI moderator should probe what is missing. The discipline is narrower and, for a system, more reliable.
4. **Silence is a tool a text interface does not have.** A human moderator's most effective probe is often saying nothing. There is no equivalent, so the design has to compensate with an explicit expansion probe rather than pretending the tool exists.
5. **A probe that a respondent could answer "yes" to is not a probe.** Anything closed, confirmatory or binary belongs in a questionnaire.
6. **Design for the interview being abandoned.** Unmoderated conversation loses the social contract that keeps people in the room. Order sections so that a partial transcript is still worth something.
7. **The disclosure is a research instrument.** How you tell respondents they are talking to an AI plausibly changes what they say. Hold it constant, publish it, and treat any change to it as a design change.
8. **Consistency is the real advantage, and it is worth designing for deliberately.** Identical probing across 300 interviews is something no team of moderators can deliver, and it makes cross-market and cross-segment comparison of qualitative material more defensible than usual.
9. **Read whole transcripts, not extracts, during design.** Thin probing is invisible in a quote and obvious in a full transcript. Anyone reviewing an AI interview design must read some end to end.
10. **The trade is against the realistic alternative.** An AI interview is usually worse than a skilled human depth and much better than an unprobed open end. Say which comparison the study is making, in the methodology, so the reader can price the evidence correctly.
11. **Contradiction is signal.** A moderator who smooths it loses the finding. Design an explicit contradiction handler and route ambiguous cases to a human (**K5 §2.6**).
12. **Assume you are running a methodological experiment as well as a study.** The field is young. Log what you tried, keep the pilot transcripts, and expect to revise the tree between waves rather than treating version one as settled.

## 14. Worked example

**INPUT**

A fictional international development NGO, the Harrow Foundation, wants to understand why monthly donors cancel. It has behavioural data on when cancellations happen and none on why. Objectives from **02.02**: reconstruct the cancellation decision, identify what preceded it, and understand what if anything would have changed it. Sample: 250 lapsed monthly donors, recruited from the organisation's own records. Mode: typed, resumable, target 12 minutes. Human moderation is unaffordable at this sample size; the realistic alternative is three open-end questions on a survey.

**PROCESS**

*Step 1, the trade.* Recorded: against a human depth, this design loses the ability to follow an unanticipated line and to build enough rapport for a donor to admit financial difficulty. Against the realistic alternative (three survey open ends) it gains reconstructed episodes and follow-up on every answer. The comparison being made is the second one, and the methodology will say so. Noted as a hypothesis at low to moderate confidence: donors may find it easier to say "I could not afford it" to a system than to a person from the charity, and the study will look for evidence either way rather than assume it.

*Step 2, sufficiency.* Three sections. Section 2 (the decision) requires four components: when the cancellation happened relative to anything else going on; what prompted it, as a specific occurrence rather than a general state; whether anything else was considered (reducing the amount, pausing, switching cause); and how the cancellation was actually carried out. Section 1 (the giving relationship before it) and Section 3 (what would have changed it) get their own definitions, with Section 3 deliberately looser because hypothetical answers are weak evidence and the design does not want to probe hard for them.

*Step 3 and 4, the tree.* Section 2 opens: "Think back to when you stopped your monthly donation. Can you tell me what was going on around that time?" Component probes: "When was that, roughly?", "Was there something in particular that happened around then?", "Did you consider anything other than stopping altogether?", "How did you go about stopping it?". Deepening probe on the prompt component: "You mentioned [respondent's own word]. What was it about that?"

*The judgement call.* The first draft of the deepening probe was "Was that mainly about money, or about how you felt about the charity?" It was cut. It offers two candidate reasons, and both are the client's hypotheses. Worse, it makes the financially difficult answer available in the same breath as a socially easier one, which is precisely the choice a leading probe should not create. Replaced with the respondent's-own-word reflection above. Recorded in the design rationale, because the client asked for the money-versus-trust split and needed to see why the probe that appeared to deliver it would have manufactured it.

*Step 5, depth rules.* Probe budget of three in Section 2, two in Sections 1 and 3. Minimum probing on Section 2 only. Escalating specificity throughout. All four stop conditions implemented.

*Step 6, states.* Deflection was anticipated: donors explaining what "people" do rather than what they did. Handler specified as one anchoring probe to a specific occasion, then accept, because a second press on a donor who is uncomfortable discussing money is both unproductive and, given who the sponsor is, not appropriate.

*Step 8, safeguarding.* Financial hardship was foreseeable, so the trigger list includes explicit statements of acute financial distress alongside the standard categories. On trigger, the AI stops probing, does not acknowledge the substance beyond one neutral line, does not offer advice, and offers the named research contact plus signposted independent financial support. Escalation route staffed weekdays during the four-week window, with a stated response time. Signed off by the NGO's safeguarding lead before launch.

*Step 11, pilot.* Twelve interviews. Two findings changed the design. Sufficiency on the "considered anything else" component failed in nine of twelve, because respondents read the opening as asking only about stopping, so a component probe was promoted to always-fire. And in three transcripts the respondent adopted the word "affordability" after the probe used it, which the probe had picked up from an earlier turn but reused in a summarising way; the reflection probe was tightened to quote rather than reuse.

**OUTPUT**

A three-section specification with component-level sufficiency definitions, a bounded tree with 3/2/2 probe budgets, an eight-state answer handling table, a signed-off safeguarding path with a named contact, disclosure text placed before consent, a documented pilot that changed two probes, a transcript inclusion threshold of two of three sections meeting sufficiency, and three review points: whether Section 3's hypothetical answers should be reported at all, whether the sponsor identity suppresses financial disclosure enough to warrant an independent-branding test, and whether the disclosure-effect hypothesis should be tested formally in wave 2.

## 15. Advanced usage

**Hybrid designs.** Use AI moderation for the structured, coverage-driven sections and a human for the section where the analytic value is highest, either as a follow-up interview with a subset or as a live handover at a defined point. This buys scale where scale is cheap and judgement where judgement matters, and it makes the comparison between the two visible within one study.

**Adaptive routing between respondents rather than within an answer.** Rather than trying to make the moderator improvise, branch at the section level on classified answers: a respondent who described a service failure goes down one section path, one who described a price decision goes down another. This is specifiable, auditable and reproducible, and it delivers much of the perceived benefit of live adaptation without the leading risk.

**Deliberate probe experiments.** Randomly assign two probe formulations across respondents on the same question and compare the sufficiency rate and the vocabulary adoption rate. Because the moderator is consistent, an AI-moderated study is an unusually clean environment for testing probe design, and the result generalises to human guides too. Plan it at design stage: it costs comparability within the study and buys evidence for the next one.

**Multi-language work.** Do not translate the probe tree. Rebuild the sufficiency definitions in each language with a native-speaking qualitative researcher, since what counts as a specific, sufficient account is partly a cultural convention, and back-translated probes routinely lose the openness that makes them non-leading. Then treat cross-market comparison of depth as suspect until the sufficiency rates by market have been compared (**K5 §2.2**).

**When the standard approach does not fit.** Where the population is unlikely to sustain a typed exchange, consider a short voice format with fewer sections and a smaller probe budget rather than a shortened text interview, and pilot the drop-off separately. Where the topic turns out in pilot to be too sensitive for the format, the correct response is to change the format, not to soften the probes: softened probes on a sensitive topic produce a transcript that looks complete and contains nothing.

## 16. Skill chain

**Recommended previous skills**
- **02.02 Discussion Guide Design.** Hands over objectives, section purposes, primary questions, order rationale and the time budget. It decides what the interview is about; this skill decides what happens after each answer.
- **02.03 Interview Question Development.** Supplies the wording of the opening question in each section, which the probe tree is built against.
- **03.01 Recruitment and Sample Sourcing.** Establishes whether the population can be reached and is suited to an unmoderated format at all.

**Recommended next skills**
- **03.04 Fieldwork Monitoring and Response Quality.** Watches sufficiency rates, abandonment points and escalation counts while fieldwork is live, when the tree can still be fixed.
- **07.03 Interview and Transcript Analysis.** Receives the transcripts, the per-section sufficiency record and the exclusion log, and inherits the rule that only respondent turns are evidence.
- **07.04 Quote and Evidence Extraction.** Takes verified respondent turns and assembles attributed evidence, inheriting the rule that only respondent text is quotable.

**Runs well alongside**
- **13.05 Research Ethics and Consent Design**, which owns the safeguarding and consent design wherever the topic or population is sensitive.
- **13.06 AI Research Governance**, which owns the disclosure of AI involvement in the deliverable and the retention and review policy.
- **02.04 Question Bias Detection**, which audits probes for leading independently of the person who wrote them.

---
A Yazi Supplied Skill and resource.