Do you have to tell participants they are talking to an AI?

Yes. Professional research codes covering consent and participant welfare require that people know who or what is conducting the interview, what is recorded, and that they can stop. Most AI interviewers are voice-only and fluent enough that participants may not otherwise be able to tell, which is exactly why explicit disclosure matters.

Does telling people it is an AI change their answers?

Yes, and the direction depends on the topic. Disclosure moves the interview toward a self-administered feel, which tends to raise candor on sensitive subjects and can reduce engagement on relationship-led ones. Keep the wording identical across the study and record it, so the effect is a known constant.

Does the disclosure have to match word for word across languages?

Not word for word, because a literal translation often reads worse than the original. What stays constant is the substance and the placement: the same facts, in the same position in the flow, at a reading level each audience actually uses. Keep a master version, have every translation reviewed by a native speaker, and record which version ran in which market.

Which research topics should not use an AI moderator?

Anything where distress is foreseeable and a participant might need a human response rather than a next question: health conditions, financial distress, bereavement, discrimination, family conflict and any research involving minors. No automated moderator has been validated as a distress detector, whatever signals it reads, so none should be relied on to notice that it should stop.

What should happen if a participant becomes distressed mid-interview?

The study should have designed a stop rather than relied on detection. The defensible pattern is a standing skip option on every question, a named human escalation route stated in the introduction, and debrief contacts that outlive the chat window. Participants use exits shown at the start far more than ones they have to request.

How do you handle consent in languages other than English?

Have the consent statement written for comprehension in the interview language rather than translated literally, and reviewed by a native speaker. Where literacy varies, use spoken or voice-note consent instead of text. Confirm understanding with a short check rather than accepting a tick box as evidence.

Consent and Disclosure in AI-Moderated Research

Last updated: 19 August 2026

Participants in an AI-moderated study must be told they are speaking with an AI. And the way you tell them changes how they answer. Disclosure is the statement explaining who or what is conducting the interview and how the recording will be used, and treating it as a compliance checkbox rather than a design decision is the most common mistake in this area.

The reason is straightforward. Disclosure sets the participant's model of who is listening, and people calibrate what they say to that model. Change the wording and you change the data. Before the first question is asked.

The Nielsen Norman Group makes the point directly in its work on methodological problems built into research tools: how a tool words its AI disclosure is itself a design choice that affects awareness and behavior.

What Are You Actually Required to Disclose?

Three things, and they are established practice rather than novel AI rules. The ESOMAR code and guidelines govern consent, participant welfare and data handling across the profession, with parallel standards from the Insights Association and AAPOR.

Who is conducting the interview. That it is an automated system rather than a person, stated plainly and before the interview begins.
What is being recorded and retained, in what form, and for how long.
How to stop. That participation is voluntary and withdrawal is possible at any point without consequence.

None of that is new. What is new is that the participant may not otherwise be able to tell. Most AI interviewers are voice-only and reasonably fluent, and independent testing found participants sometimes felt genuinely heard because the system summarized their thoughts back to them well. Fluency is precisely why explicit disclosure matters.

Informed consent has always rested on one test: did the person understand what they were agreeing to. A fluent synthetic voice does not change the standard. It raises the effort needed to meet it, because the cue people normally use to work out who they are talking to has been removed.

Does Disclosure Change the Answers?

Yes, and in both directions depending on the topic. This is the part that makes it a design decision.

Survey methodology has measured the underlying effect for decades. Pew Research Center's work on mode of interview effects notes respondents may feel a need to present themselves in a more positive light to an interviewer, which inflates socially desirable answers. Research on social desirability bias and sensitive questions finds interviewer-administered instruments draw more socially desirable responses than self-administered ones.

Telling someone they are talking to a machine moves the interview toward the self-administered end of that spectrum. On a sensitive subject that can raise candor. On a relationship-led subject it can lower engagement, producing shorter and flatter answers.

The effect is topic-dependent rather than universal. Pew found few mode effects when asking about news consumption habits.

So the practical instruction is not "minimize disclosure". It is: disclose fully, keep the wording constant across the whole study, and record it in the method note so the effect is a known constant rather than an uncontrolled variable.

That last point matters most for continuous work. If disclosure wording changes between waves of a tracking program, any shift you measure could be a change in the market or a change in your own preamble, and nothing in the data separates the two.

How Should Disclosure Be Worded?

Plainly, early, and identically for everyone. The goal is comprehension, not coverage.

Element	Good practice	Warning sign
Timing	Before the first substantive question	Buried in a pre-screen wall of text
Agent	"An automated interviewer will ask the questions"	"Our smart research assistant"
Recording	What is captured, kept, and for how long	"Standard terms apply"
Withdrawal	Stated, with a concrete way to stop	Implied but not explained
Register	Language and reading level of the participant	Legal English for a non-English speaker
Consistency	Word-for-word identical across the sample	Varies by market or platform default

The register row is the one most often failed. Consent that a participant cannot read is not consent, and translating a legal paragraph literally usually produces something less comprehensible than the original.

The timing row runs it close. Disclosure placed inside a pre-screen wall of terms is technically present and practically invisible, and a participant who scrolled past it has not been informed of anything. If the disclosure only appears somewhere a determined reader could find it, treat it as absent.

Which Topics Should Not Go to an AI Moderator?

Ones where a participant may need a human response rather than a next question. This is a welfare question before it is a data-quality question.

That same work concluded these tools are not suited to messy problem spaces, high-stakes decisions or studies requiring deep domain knowledge and real-time judgment, and that they supplement rather than replace human moderation. Applied to sensitive research, that translates into a fairly clear boundary:

Route to a human: health conditions, financial distress, bereavement, discrimination, family conflict, anything involving minors, and any topic where distress is foreseeable.
Proceed with care and a stated exit: money habits, debt, body image, workplace dissatisfaction.
Fine for AI moderation: product experience, category habits, concept reaction, service friction.

The operative test is not how personal the topic sounds. It is whether a participant could become distressed and, if they did, whether anything in your study would notice. An automated moderator should not be trusted to. What signals a system reads varies by tool and mode, and none has been validated as a distress detector, so the design must not depend on it noticing.

That is a design requirement rather than a reason to abandon the method. A study can route sensitive sections to a human wave while the AI-moderated portion covers everything else, and participants can be given a stated route to a person if they want one. What is not defensible is running a foreseeable-distress topic entirely through a system that cannot recognize distress.

Design the stop instead of hoping for detection: a standing skip option on every question, a named human escalation path in the introduction, and debrief contacts that outlive the chat window. Participants use exits shown at the start far more than ones they must request. Where a human moderator still wins is worked through in AI against human moderated interviews.

What Changes in Multilingual Fieldwork?

The hard problem stops being disclosure and becomes comprehension. A consent statement translated for legal accuracy rather than for how it will be understood is a common and quiet failure.

Reach compounds it. Pew Research Center's mobile technology fact sheet reports 16 percent of US adults as smartphone-only internet users, rising to 34 percent in households under $30,000 a year, and the ITU's Facts and Figures 2025 shows a steeper gradient again across markets. Studies reaching those populations are exactly the ones where written English consent works least well.

How Do You Make Consent Comprehensible?

Three practices help:

Consent in the interview language, at the register the participant actually reads, reviewed by a native speaker rather than a translation tool. Published transcription and translation protocols for cross-language research exist for exactly this, and they treat translation quality as a study-validity question rather than an administrative one.
Spoken consent where literacy varies. A voice-note or read-aloud statement reaches people a text wall does not.
Confirm understanding rather than receipt. A short comprehension check beats a tick box.

This is one of the clearest arguments for managed fieldwork over handing over software. Consent quality in an unfamiliar market is an operational problem: someone has to know which languages the sample actually reads, what incentive framing reads as coercive locally, and who to escalate to. A platform does not solve that, and a checkbox certainly does not.

Alchemic runs interviews natively inside WhatsApp with no link and no app to install, and AI phone interviews to any working number, with moderation across 57+ languages including Spanish, Hindi, Tamil, Bangla, Arabic and Indonesian.

Fieldwork is managed across fourteen markets spanning the USA and the UK as well as South and Southeast Asia, the Gulf and Africa. Those modes reach populations a browser study cannot. That raises the comprehension bar rather than lowering it, because the wider the frame, the more varied the reading levels inside it. Testing a vendor's language claim is covered in multilingual AI-moderated interviewing.

How Do You Check Your Own Study?

Five checks, none of which requires legal review.

Read the disclosure as a participant would, in the interview language, out loud, at the pace someone half-paying-attention would use.
Time it. If disclosure runs longer than a minute, people will skip it, and comprehension drops faster than length increases.
Confirm it is identical across every market and platform. Vendor defaults vary and quietly introduce a variable.
Ask what happens if someone becomes distressed. If the answer is nothing, that topic needs a human.
Record the exact wording in the method note, so a later reader can judge the mode effect.

Ask your vendor to show you the disclosure text and the point in the flow where it appears, rather than describing it. It is a two-minute request that surfaces a surprising amount.

Document what you find, even when it is fine. Consent practice is one of the few parts of a study that a client, a regulator or an ethics reviewer may ask about years later, and reconstructing it from memory is unpleasant. A short method note recording the wording, its placement and the languages it ran in costs minutes now and settles the question permanently.

Where This Guidance Stops

This is practice, not legal advice. Data protection duties vary by jurisdiction and by client contract, and your legal team owns those.
Standards are still settling. Professional bodies are actively revising guidance for AI-moderated methods, so check the current version rather than a cached summary.
Full disclosure has a cost. It can reduce engagement on relationship-led topics, and pretending otherwise is how teams end up quietly weakening it.
Disclosure does not make a sensitive topic safe. Some studies need a human regardless of how well the consent is written.
Comprehension is not the same as consent. Someone can understand perfectly and still feel unable to decline, which is why voluntariness has to be stated rather than assumed.

Several of these decisions are made for you by tool defaults unless you override them deliberately. That is the practical reason to read your own disclosure text end to end rather than confirm that one exists.