What Is Informed Consent in AI-Moderated Research?
Qualitati Research Team · 2026-08-15 · 11 min read
Short answer. Informed consent in AI-moderated research is the participant's voluntary, documented agreement to take part after being told, in plain language, that an AI system will conduct the session, what data it collects, how that data is processed and retained, whether it trains a model, and what the AI can get wrong. Since 2 August 2026, EU law also requires disclosing the AI itself.
What is informed consent in AI-moderated research?
Informed consent in AI-moderated research is the same ethical instrument researchers have always used — voluntary, competent, documented agreement to participate — extended to cover facts that only exist when a machine is the moderator. Traditional consent forms tell participants who is asking, why, and what happens to their answers. AI-moderated consent must additionally tell them what is asking, how the transcript is processed, and where the automation is fallible.
Two developments in 2026 made this concrete rather than aspirational. On 29 January 2026, Rutgers Office for Research published AI in Human Subjects Research: Rutgers Consent Guide, giving IRB-facing plain-language templates across eight sections including procedures, risks, privacy protections, commercial use, and data retention. On 2 August 2026, the transparency obligations in Article 50 of the EU AI Act took effect, making AI disclosure a legal duty in the EU and not only an ethical norm.
Key takeaways
- Consent for AI-moderated research adds four disclosures to a standard form: the AI moderator's identity, how transcripts are processed, whether data trains any model, and the AI's known limitations.
- EU AI Act Article 50(1) requires that people be informed they are interacting with an AI system, at the latest at the time of the first interaction, unless it is obvious to a reasonably well-informed person.
- Article 50(3) separately requires informing people exposed to an emotion recognition system — relevant to anyone running affect inference on interview audio.
- Disclosure must be legible before the session starts, not buried in a terms-of-service link at the end.
- Consent is a research-design decision, not a compliance afterthought: what you promise about retention and model training constrains how you can analyze the data later.
- This article is research-practice guidance, not legal advice. Confirm obligations with your IRB, ethics board, or counsel.
Why AI moderation changes the consent conversation
A human-moderated interview has one obvious processor: the interviewer. Participants intuitively model what happens next — someone listens, someone writes it up. An AI-moderated session breaks that intuition in three places.
The counterpart is not a person. Participants who believe a human is listening may disclose differently than participants who know it is a system. That is not a nuisance to be managed; it is a fact about the data. Concealing it corrupts both the ethics and the interpretation.
Processing is invisible and multi-stage. A single interview may pass through speech recognition, a language model that generates follow-up questions, a separate analysis pipeline that codes the transcript, and a storage layer. "Your interview will be recorded" no longer describes what happens.
Inference goes beyond what was said. Automated analysis can surface patterns participants did not intend to communicate — hesitation, sentiment, acoustic markers. Honest consent names this possibility rather than letting it arrive as a surprise in a report.
What Article 50 actually requires (as of 15 August 2026)
Article 50 of the EU AI Act sets transparency obligations for certain AI systems. The two paragraphs that matter most to research teams:
| Provision | Requirement (paraphrased from the published text) | Research implication |
|---|
| Article 50(1) | Providers must design AI systems intended to interact directly with people so that those people are informed they are interacting with an AI system. Not required where this is obvious to a reasonably well-informed person. | Name the AI moderator on the consent screen. Do not rely on the "obvious" carve-out for a voice interview that sounds human. |
| Article 50(3) | Deployers of an emotion recognition or biometric categorisation system must inform the people exposed to it. | If you run affect or emotion inference on interview audio, that is a separate disclosure — a general "we use AI" line does not cover it. |
| Timing | Information must be given in a clear and distinguishable manner at the latest at the time of the first interaction or exposure. | Consent before the first question, not a debrief afterwards. |
The obligations became applicable on 2 August 2026, and the European Commission has published guidelines for providers and deployers alongside them, with the full text of the provision available at Article 50. Scope, enforcement timing, and the provider/deployer split are genuinely technical questions — the point here is not to settle them but to note that the disclosure a sound research ethic already recommended is now also a legal expectation in the EU.
The AI-Moderated Research Consent Disclosure Checklist
This is a Qualitati-built checklist, written to be readable by a participant rather than by a lawyer. Work through it before you send a single interview link. Each item should be answerable in one sentence on the consent screen.
- Who is conducting the session. State plainly that an AI system will ask the questions, and name the organization behind it.
- Whether a human will see the data. Say whether researchers read transcripts, and roughly when.
- What is captured. Text, audio, video, timing, or all of these — be specific rather than saying "your responses."
- What processing happens. Transcription, automated follow-up questions, automated coding or thematic analysis, any acoustic or emotion inference.
- Whether any of it trains a model. A yes or no in plain words. If no, say so; participants overwhelmingly assume yes.
- Who else touches the data. Named categories of subprocessors, and whether data leaves the participant's jurisdiction.
- How long it is kept, and what deletion means. A retention period, and whether deletion covers derived artifacts such as embeddings and codes.
- Whether it is commercial. Whether findings feed a commercial product decision, a client deliverable, or published research.
- Limitations and fallibility. The AI may misunderstand, may miss a cue a human would catch, and may infer patterns the participant did not intend to convey.
- How to stop, skip, or withdraw. An explicit exit route mid-session, and a contact for withdrawal after the fact.
A two-layer consent pattern
Long consent text reduces comprehension, which defeats the purpose. A workable structure separates what participants must read from what must exist on record:
| Layer | Contains | Where it lives |
|---|
| Layer 1 — Key information | Items 1, 2, 5, 9, 10: it is an AI, humans do or do not read it, training yes or no, it can be wrong, how to stop. | The consent screen itself, before the first question. Six sentences or fewer. |
| Layer 2 — Full notice | Items 3, 4, 6, 7, 8, plus jurisdiction, legal basis, and contacts. | Expandable panel or linked notice, retained with the study record. |
The test for Layer 1 is simple: could a participant who reads only that layer accurately describe, afterwards, what they agreed to? If not, something belongs one layer up.
Where Qualitati fits
Qualitati is an AI user research platform for product, UX, and customer insights teams. It runs AI-moderated interviews in text and voice, AI-moderated and synthetic focus groups, conversational surveys with AI-driven follow-ups, and ThemeLens AI thematic analysis across up to 100 transcripts, in ten languages.
Consent is where a platform's design choices become visible to participants. Three points are worth stating precisely, because vagueness here is exactly the problem this article is about:
- Voice Analytics is a separate disclosure. Qualitati's Voice Analytics extracts acoustic features from interview audio — pitch, loudness variability, speech rate, voice quality. If you enable it, treat it as its own line in the consent text, not as a footnote to "we record audio."
- Automated analysis is part of "procedures." ThemeLens and QDA Workspace code and synthesize transcripts. Participants should know their words will be machine-coded, not only read.
- The AI moderator is disclosable by design. An AI-moderated interview is conducted by an AI interviewer you configure; naming that on the consent screen costs one sentence and resolves Article 50(1) squarely rather than relying on an "obvious" argument.
For the specifics of your own deployment — retention periods, subprocessors, jurisdiction, and whether any data is used for model training — verify current terms on the pricing and terms pages rather than reproducing them from a blog post. Those are contractual facts and should come from the contract.
Limitations and trade-offs
Disclosure changes behavior. Telling participants the moderator is an AI is non-negotiable, but it is also a measurable intervention. Some participants disclose more to a machine; some disclose less; some perform for it. Treat the disclosure as a documented condition of your study, and be cautious comparing AI-moderated data against older human-moderated data collected under different consent.
Comprehension is not the same as signature. A checkbox proves a click. Whether the participant understood that an automated pipeline would code their words is a separate question, and the honest answer for most consent flows is that we do not measure it.
Withdrawal is harder than it sounds. Once a transcript has been coded, summarized into a theme, and quoted in a report, "delete my data" is a multi-artifact operation. If you promise withdrawal, confirm your pipeline can actually honor it downstream, not just delete the source row.
Regulatory scope is unsettled. Whether a given research deployment makes you a provider, a deployer, or neither under the AI Act depends on facts this article cannot know. The checklist above is written to be defensible under a strict reading; it is not a compliance determination.
Who this is for, and when not to use this approach
Who this is for: UX researchers, product managers, insights leads, and academic researchers running AI-moderated interviews, focus groups, or conversational surveys, and anyone preparing an IRB or ethics submission that mentions AI.
When not to use AI moderation at all: studies with participants who cannot meaningfully consent to automated processing; topics where a distressed participant needs a human who can recognize and respond to distress; contexts where the population has well-founded reasons to distrust automated data collection, and where disclosure would make participation feel coercive rather than voluntary. In those cases the right answer is a human moderator, and consent language is not the fix.
FAQ
Do I legally have to tell participants the moderator is an AI?
In the EU, Article 50(1) of the AI Act requires that people interacting directly with an AI system be informed of that, at the latest at the time of first interaction, unless it is obvious to a reasonably well-informed person. The obligations became applicable on 2 August 2026. Outside the EU, disclosure is generally an ethics-board expectation rather than a statute. Either way, disclose. Verify your own obligations with counsel.
Does a standard IRB consent template cover AI moderation?
Usually not without amendment. Standard templates describe human-collected data. The Rutgers consent guide published on 29 January 2026 exists precisely because the AI-specific elements — processing, retention, commercial use, model limitations — were not covered by existing language.
Do I need separate consent for automated analysis of the transcript?
Not necessarily a separate form, but a separate statement. Machine coding and thematic synthesis are procedures, and procedures belong in the consent. Emotion or acoustic inference is a stronger case: Article 50(3) addresses emotion recognition systems specifically.
What should I say about model training?
Whatever is true, in one sentence, without hedging. Participants default to assuming their data trains a model. If it does not, saying so plainly is one of the highest-value sentences on the form.
How long should the consent screen be?
Layer 1 should be readable in under a minute — roughly six sentences. Everything else belongs in an expandable full notice. Length is not rigor; comprehension is.
Can participants withdraw after an AI-moderated session?
They should be able to, and you should confirm your pipeline can honor it across derived artifacts — transcripts, codes, embeddings, and any quotes already pulled into a report — before you promise it.
Bottom line
Informed consent in AI-moderated research is not a longer form. It is four honest additions to a familiar one: say it is an AI, say what happens to the recording, say whether it trains anything, and say what the system can get wrong. The EU made the first of those a legal duty on 2 August 2026; the other three were always the ethical baseline, and they are what actually earns a participant's trust.
Qualitati runs AI-moderated interviews, focus groups, conversational surveys, and AI thematic analysis with transparent, published pricing. Start free with 30 credits — no credit card required — or view transparent pricing. If you are designing the study itself, our guides on ethical risks in AI-assisted interviewing, anonymizing interview transcripts, and reflexivity in qualitative research cover the neighboring decisions.
Last updated: 15 August 2026. This article is an independent editorial summary of publicly available sources and is research-practice guidance, not legal advice. Regulatory obligations depend on your jurisdiction and deployment — confirm them with your IRB, ethics board, or counsel.