EU AI Act Transparency Rules for Market Research: What Article 50 Means for Research & Insights Teams

Updated: August 2026. This article reflects the European Commission’s final Article 50 transparency guidance published in July 2026.

Sept 02, 2026 11:51
All articles

At DataTile, we spend a lot of time thinking about what happens between raw research data and the final output a client actually uses — analysis, crosstabs, dashboards, commentary and reports. AI is quickly becoming part of that process too.


So when the EU AI Act’s transparency rules became applicable in August 2026, we wanted to understand what they actually mean for research teams. Not at the level of generic AI compliance or another headline about “AI-generated content”, but for very practical questions: if AI helps code open ends, draft dashboard commentary or interview a respondent, what actually changes?


We went back to Article 50, the European Commission’s final guidance and the current MRS and ICC/ESOMAR standards and tried to separate what the law actually says from interpretation and good practice.


We are not lawyers, and we are certainly not claiming that our reading is the final word. Some parts of Article 50 still depend on context and interpretation. If you spot something we have missed, think we have interpreted a provision differently from you, or have a better example from real research practice, please tell us. We would rather update the article than pretend the rules are clearer than they are.




The EU AI Act’s new transparency rules are now in effect. But for market research teams, the change is more specific than many headlines suggest: it does not mean every report, chart, dashboard comment or sentence touched by AI needs an “AI-generated” label.


The useful question is what the AI actually did and where its output went. Was it working behind the scenes or talking to a respondent? Are you using someone else’s system or providing one? Is the output internal, client-only or public? And did a researcher actually review the substance?


Article 50 contains several different transparency duties, triggered by different things. There is no single rule for “AI-generated content”.


For market research agencies and insight teams, there is another layer too: professional standards such as MRS and ICC/ESOMAR can require more transparency around AI use than Article 50 itself.


So this guide follows AI through a typical market research workflow — from questionnaire design and fieldwork to analysis, dashboards and reporting — and gives a practical answer for the common cases while flagging the smaller number of genuine grey areas.


This article provides a practical overview for research teams and is not legal advice.

What actually changed on 2 August 2026?

The EU AI Act did not come into force on 2 August 2026. The Regulation entered into force earlier and has been applied in stages.

On 2 August 2026, the transparency obligations in Article 50 became applicable. The European Commission’s current guidance confirms that providers and deployers falling within Article 50 must comply from that date, subject to a limited transition for some provider-side machine-readable marking requirements.


For research teams, the most relevant parts cover:


  • AI systems that interact directly with people;

  • machine-readable marking of certain AI-generated or manipulated content by AI-system providers;

  • disclosure of deepfakes by organisations using AI systems;

  • disclosure of certain AI-generated or manipulated text published to inform the public on matters of public interest.

These are separate rules, triggered by different things. Article 50 does not say: “If AI helped create something, put an AI label on it.”



Before asking “Do we need a label?”, ask whether you are the provider or the deployer


Before applying the rules to a research workflow, it helps to know which role your organisation is playing.


Under the AI Act, a provider is broadly the organisation that develops an AI system — or has it developed — and places it on the market or puts it into service under its own name or trademark.


A deployer is the organisation using an AI system under its authority in a professional context.


Research scenario

Likely role

An agency uses ChatGPT to help analyse survey responses

Usually deployer

A brand insights team uses an external AI analysis tool

Usually deployer

A research-tech company develops and sells its own AI research assistant

Provider

An agency builds or white-labels a client-facing AI system

Potentially provider (and possibly deployer); assess how the system is branded, supplied and operated

The same company can be a deployer in one workflow and a provider in another.


That distinction matters because some duties described loosely as “AI labelling” sit with the provider, not automatically with every organisation using the output.


For example, Article 50(2) puts the machine-readable marking duty for certain synthetic text, image, audio and video outputs on the provider of the relevant AI system.


So an agency using a third-party tool such as ChatGPT does not automatically inherit that marking duty. Its own visible-disclosure obligations arise only when a deployer-side trigger applies — for example, a qualifying deepfake or certain public-interest text.

AI in research design and fieldwork: when respondent disclosure starts to matter

A simple comparison shows why the role of AI matters more than the fact that AI was used.

A researcher asks an AI tool to suggest five follow-up questions about customer loyalty, chooses one, edits it and puts it into a conventional questionnaire.


Now compare that with an AI interviewer that reads a respondent’s answer, interprets it in real time and independently asks a new follow-up question.


In the first case, AI is helping the researcher behind the scenes; the respondent is not interacting with the AI system.

In the second, if the AI itself is holding a genuine two-way conversation with the respondent, Article 50(1) is likely to apply. The system should be designed to tell the participant that they are interacting with AI from the first interaction, unless that is already obvious.


That distinction matters as research platforms move beyond fixed questionnaires into conversational surveys, AI moderators and automated interviewing.




When AI talks to respondents


The European Commission’s final guidance says the direct-interaction rule is aimed at genuine AI-human exchanges, not simple data collection or predetermined automated responses.


Where the rule applies, the provider must design the system so that people are informed that they are interacting with AI, unless this is already obvious in context.


AI chatbots and respondent support. If a chatbot generates context-aware answers and holds a genuine two-way exchange with a participant, Article 50(1) is likely to apply and the participant should be told they are interacting with AI. A static FAQ or predetermined autoresponder would normally sit outside this specific rule.


Conversational surveys. A survey using branching logic or pre-written automated responses would normally sit outside the genuine two-way exchange test. If the system interprets free-text answers and generates its own context-aware replies or follow-up questions, Article 50(1) is much more likely to apply.


AI interviewers and moderators. If the AI is actually conducting the interview, the practical answer is fairly clear: participants should be told they are talking to AI from the first interaction, unless that is already obvious.


For respondent-facing AI, transparency is therefore a fieldwork design issue, not something to add to the final report afterwards.

AI in analysis: coding, summarising and interpreting research data


A large share of AI use in research happens after fieldwork: coding open ends, summarising interviews, identifying themes and drafting commentary.

Here the “AI used = label required” shortcut breaks down quickly.

Coding open-ended responses


Suppose an AI system helps classify 20,000 open-ended responses into themes.

For a typical internal coding workflow, you would not normally need a visible Article 50 label just because AI was used. The more immediate questions are whether the coding is validated, whether client or professional disclosure is required, and what data was sent to the AI system.

In practice, the action is usually not “label the coded data”; it is “document the AI use, validate the coding and disclose it to the client where the applicable professional standard requires it.”

Summarising interviews and identifying themes


If an AI-generated summary or theme analysis remains an internal working document, a visible Article 50 label would not normally be required for that reason alone.

If it becomes part of a private client deliverable, the Article 50(4) public-interest text rule is still unlikely to apply because the material is not being published to inform the public. MRS or ICC/ESOMAR client-disclosure requirements can still apply.

If the text is published publicly on a matter of public interest, the position changes: unreviewed AI-generated text is more likely to require disclosure, while qualifying substantive human review or editorial control can remove that specific labelling requirement.

AI-generated crosstab commentary and interpretations

Take AI-generated crosstab commentary. The answer is easier to see when you separate three common uses:


  • Internal analyst suggestion: no visible Article 50 label would normally be needed. It is internal analysis, not text published to inform the public.
  • Private client dashboard after analyst review: again, an Article 50(4) label would normally not be needed if the dashboard is genuinely restricted to the client. MRS/ICC-ESOMAR client-transparency rules may still require the agency to explain how AI was used.
  • Automatically published on an open website: if the commentary is intended to inform the public on a matter of public interest and has not undergone qualifying human review or editorial control, Article 50(4) is likely to require disclosure. If the topic is not a matter of public interest, that specific deployer-side text rule is unlikely to apply.

The provider’s machine-readable marking duty under Article 50(2) is separate from the question of whether the research organisation itself needs a visible label.

This is also where the question becomes particularly relevant to us at DataTile.


A large part of the work our users do happens between the raw survey data and the final client output: building crosstabs, exploring differences between groups, creating dashboards and turning those findings into something a client can actually understand and use.

As AI starts highlighting differences in crosstabs, suggesting explanations or drafting dashboard commentary, it is no longer just helping with the mechanics of analysis — it is starting to influence the story being told about the data. That makes it especially important to understand what came from the system, what was checked or changed by a researcher, and what ultimately reached the client.


Reports and dashboards: private client output is not the same as public publication

For agencies, one distinction matters more than almost any other: a private client deliverable is not the same thing as a public publication.

Article 50(4)’s text disclosure rule applies where AI-generated or manipulated text is published to inform the public on a matter of public interest.


Private client reports


Suppose AI drafts part of an executive summary for a brand-tracking study, and a researcher checks the data, corrects an interpretation and sends the final report privately to the client.

For that typical private-report scenario, an Article 50(4) label would generally not be required simply because AI helped draft the text: the report is not being published to inform the public.

But “no Article 50 label” is not the same as “no disclosure at all”. MRS/ICC-ESOMAR requirements, client contracts, confidentiality and data-protection rules can still matter.

Client dashboards


Dashboards are where “private” becomes less tidy, so it helps to split them into practical buckets.

  • Internal dashboard used only by the research team: no Article 50(4) public-text label would normally be needed.
  • Restricted client dashboard or portal: again, normally no Article 50(4) label if access is genuinely limited to the client. Professional client-disclosure requirements can still apply.
  • Public interactive dashboard: if AI-generated commentary informs the public on a matter of public interest and has not undergone qualifying human review or editorial control, disclosure is likely required.
  • Large gated or subscription portal: this is a real grey area. The current guidance does not give a market-research-specific rule for when a gated audience becomes “the public”, so a login should not be treated as an automatic exemption.

Public research reports, whitepapers and articles

For public reports, the test is more useful if you turn it into a decision rule.

If a report is public and deals with a topic clearly within the Commission’s examples — such as public health, consumer safety or economic developments relevant to public debate — assume Article 50(4) may be in scope.

If the AI-generated or manipulated text has not undergone qualifying human review or editorial control, the deployer is likely to need to disclose its artificial origin.

The Commission’s examples also include politics, public administration, fundamental rights, public security, environmental protection and scientific or cultural developments relevant to public debate.

If the text has undergone substantive human review or editorial control and a person or organisation holds editorial responsibility for publication, Article 50(4) says that this text does not need that AI label.

Routine commercial brand or product research is less likely to qualify as a “matter of public interest” merely because it is published. But there is no market-research carve-out: a commercial study can move closer to scope if it deals with consumer safety, public health or another issue relevant to public debate.

A useful shorthand is: public + public-interest subject + no qualifying review = likely disclosure.


Human review matters — but not in the way many people think

The human-review exception is important, but it is narrower than “a person looked at the output”


The Commission’s guidance says qualifying public-interest text does not need the Article 50(4) label if it has undergone the relevant human review or editorial control and someone holds editorial responsibility for publication.


What counts as substantive human review?

The Commission describes human review as deliberate examination of the substance by people with relevant knowledge and professional judgement. Editorial control means having real authority to approve, change or reject substantive content.


A spelling or grammar check is not enough.


In a research workflow, meaningful review could include:


  • checking AI-generated claims against the underlying data;

  • validating statistical conclusions;

  • checking whether the methodology supports the interpretation;

  • identifying and correcting hallucinated or unsupported conclusions;

  • changing or rejecting AI-generated commentary;

  • taking responsibility for the final interpretation.

That is our application of the Commission’s standard to research practice, not an official market-research checklist.


If qualifying public-interest text has gone through this level of review and someone carries editorial responsibility, Article 50(4) says it does not need that AI label.


But human review is not a universal exemption. It only affects this specific Article 50(4) text rule. It does not remove other duties — for example, reviewing an interview transcript later does not undo the requirement to tell a respondent they are speaking with an AI interviewer.




AI-generated images, audio, video and synthetic respondents

Research teams are also experimenting with synthetic personas, simulated respondents, generated voices and video.


Synthetic respondents and personas


Article 50 has no special legal category called a “synthetic respondent”. The practical answer depends on what the synthetic element is doing.


  • Synthetic persona used only to simulate responses internally: there is no special Article 50 “synthetic respondent” label. MRS/ICC-ESOMAR rules can still require client transparency when synthetic methods affect analysis or outputs.
  • Synthetic persona collecting data from a real participant: ICC/ESOMAR requires the participant to be told at the beginning of the research. If the persona also works as a genuine conversational AI system, Article 50(1) is likely relevant as well.
  • Synthetic data or personas used in published findings: ICC/ESOMAR requires disclosure where AI or synthetic methods played a significant role, together with the extent of human oversight.


Synthetic video and audio: when deepfake rules may matter


Not every AI-generated image, voice or video is a deepfake.


The Commission’s test focuses on content that resembles something existing or plausibly existing and falsely appears authentic or truthful.


A synthetic “respondent video” designed to look like a genuine interview can therefore meet the deepfake test. If it does, the deployer should disclose — visibly or audibly — that the content was artificially generated or manipulated when the viewer is first exposed to it.


An obviously fictional or illustrative stimulus is less likely to qualify as a deepfake because the audience is not being led to believe it is authentic. Provider-side machine-readable marking remains a separate issue.

DataTile - EU AI Act Transparency Rules for Market Research - when deepfake rules may matter
when deepfake rules may matter

What about UK research agencies?

Being based in the UK does not automatically put a research agency outside the EU AI Act.

Article 2 gives the Act extra-territorial reach in several situations, including where AI output produced by a provider or deployer outside the EU is used in the Union.


For agencies, three scenarios are useful:


  • UK-only internal use, with no AI system placed on the EU market and no AI output used in the EU: those main Article 2 scope hooks are less likely to apply to that use.
  • UK agency produces AI output that is used by an EU client or team: Article 2 creates an EU scope hook. You still need to identify the specific Article 50 trigger — for example, public-interest text or a deepfake.
  • UK research-tech provider makes an AI interviewing product available in the EU: this is a much clearer in-scope provider scenario, and Article 50(1) is likely relevant if the product directly interacts with people.

So “we have an EU client” is not the legal test. EU-market availability and use of AI output in the EU are the more concrete scope questions.

The AI Act is not the only transparency standard research teams need to consider

For agencies, this is probably the part worth paying most attention to.


Article 50 is only one layer. No Article 50 label does not automatically mean that AI use can stay invisible.


Teams often mix up four different things:


  • the AI Act: specific AI transparency duties;
  • GDPR / UK GDPR: how personal data is processed and protected;
  • client contracts: what data can be sent to which external tools, and whether AI may be used at all;
  • professional research standards: broader expectations around transparency, quality and human responsibility.



MRS and ICC/ESOMAR can require more transparency than Article 50


The MRS Guidance on Using AI and Related Technologies applies to MRS members and Company Partners and sets expectations around transparency, responsibility, confidentiality, human oversight, traceability and reporting.


MRS requires practitioners to keep enough records for AI-related work to be audited, replicated and traced, and to maintain human oversight when assessing AI-generated research outputs and insights.


For reporting, MRS requires practitioners to record and communicate which research outputs are generated by AI or related technologies — including synthetic data — and which are attributable to human insight.


The 2025 ICC/ESOMAR International Code takes a similar approach: clients must be informed when AI or other emerging technologies are used in compiling datasets, analysis, reporting or interpretation, and the extent of human oversight must be stated.


Question

Article 50

MRS / ICC/ESOMAR

AI was used internally in analysis

Visible Article 50 label: generally no, if the output remains internal

Client transparency and human-oversight requirements can still apply when AI affects outputs

AI helped create a private client report

Article 50(4) label: generally no, if the report is delivered privately rather than published to the public

MRS/ICC-ESOMAR can require the client to be told how AI was used

Synthetic persona used in research

No special Article 50 'synthetic respondent' category; Article 50(1) can apply if it directly converses with a person

ICC/ESOMAR requires disclosure at the start when a synthetic persona is used for data collection

Human oversight

For qualifying public-interest text, substantive review/editorial control plus editorial responsibility can remove the Article 50(4) label

Broader human oversight is expected across the research process

The practical point is simple: “no Article 50 label” and “no need to disclose AI to the client” are not the same conclusion.


A practical Article 50 map for the research workflow

Try it and give us your opinion if it works


The table below gives the practical answer for the common version of each scenario. The genuine edge cases are the ones where role, audience or publication model needs closer review.


Research use case

Main Article 50 question

Practical takeaway

AI helps draft survey questions

Is AI directly interacting with the respondent?

Most likely no Article 50 disclosure to the respondent solely because AI helped draft the questions

AI suggests follow-up questions to a researcher

Who actually communicates with the participant?

If suggestions go only to the researcher: most likely no interaction disclosure. If AI asks the participant directly: Article 50(1) is likely relevant

AI chatbot communicates with respondents

Is there direct AI-human interaction?

Most likely yes: disclose the AI interaction from the first interaction unless it is obvious

AI interviewer or moderator conducts an interview

Is the AI holding a genuine two-way exchange?

Most likely yes: Article 50(1) is likely relevant and the participant should be informed they are interacting with AI

AI codes open-ended responses

Is this internal analysis or a public output?

Internal use: most likely no visible Article 50 label. Professional client-disclosure and validation duties can still apply

AI summarises qualitative interviews

Where does the summary go next?

Internal: most likely no label. Private client deliverable: usually no Article 50(4) label, but professional disclosure can apply. Public-interest publication: assess Article 50(4)

AI generates crosstab commentary

Who sees the commentary and how is it used?

Internal/private client dashboard: usually no Article 50(4) label. Public-interest website text without qualifying review: likely disclosure

AI drafts a private client report

Is the text being published to inform the public?

Most likely no Article 50(4) visible label if delivered only to the client; MRS/ICC-ESOMAR disclosure can still be required

AI contributes to a public research report

Is it public-interest text and was it substantively reviewed?

Public-interest + no qualifying review: likely disclosure. Qualifying review/editorial responsibility: no Article 50(4) text label

Researcher substantively reviews AI-generated text

Does the specific Article 50(4) review exception apply?

For qualifying public-interest text, substantive review plus editorial responsibility can remove the Article 50(4) label; other Article 50 duties remain

Synthetic persona is used

What is it doing and who is interacting with it?

No special Article 50 category. Direct conversational use can trigger Article 50(1); ICC/ESOMAR requires participant disclosure when used for data collection

Synthetic respondent video is shown as authentic

Could it meet the deepfake criteria?

If it looks like a genuine interview and meets the deepfake test: likely visible/audible disclosure. Obviously fictional stimuli are less likely to qualify

UK agency uses AI in an EU-facing project

What is the territorial connection and the agency’s role?

If AI output is used in the EU, Article 2 can bring a third-country provider/deployer into scope; then apply the specific Article 50 trigger


What should research teams do now?

The practical response is not to put “AI-generated” stickers everywhere. Start by understanding your own workflow.


  1. Map where AI is used. Cover questionnaire design, recruitment, fieldwork, respondent interaction, analysis, coding, interpretation, dashboards and reporting.
  2. A simple AI-use register can capture: use case, tool/provider, data submitted, respondent interaction, human reviewer, final owner, client/participant disclosure and the reason for that disclosure decision.
  3. Know your role. Are you using a third-party system, or supplying an AI system to clients or respondents?
  4. Separate respondent-facing AI from internal assistance. An AI interviewer is not the same regulatory scenario as a researcher using ChatGPT to improve questionnaire wording.
  5. Know what data goes into third-party AI systems. Document whether client, participant, personal or confidential data is submitted, and under what contractual and data-protection arrangements.
  6. MRS also requires protection of employer, client and participant data when AI is used, including appropriate controls on sharing confidential or personal information with systems used for training.
  7. Define meaningful human review. Decide who checks AI-generated analysis, what they must validate and who takes responsibility for the final research conclusion.
  8. Keep legal requirements and professional standards separate. Ask both what Article 50 requires and what MRS/ICC-ESOMAR, contracts and data-protection obligations require.
  9. Be ready to explain your AI use to clients. Know which tools are used, where they are used, what data they receive, how outputs are checked and what remains the researcher’s responsibility.

In an AI-enabled research workflow, “we use AI” is no longer a sufficient explanation.



Frequently asked questions


Not automatically. A report sent only to a client would not normally need the Article 50(4) public-interest text label. A public-interest report without qualifying human review or editorial control is much more likely to require disclosure.

Usually yes if the chatbot holds a genuine direct two-way AI conversation with the respondent. A conventional survey, branching logic or predetermined autoresponse would normally sit outside this specific interaction rule.

No. Human review is relevant to a specific exception for qualifying public-interest text under Article 50(4); it does not remove the other Article 50 transparency duties.


Usually not for an internal analyst suggestion or a genuinely private client dashboard. If the commentary is published to the public on a matter of public interest without qualifying human review or editorial control, disclosure is much more likely to be required.


Yes, in some scenarios. Article 2 can bring UK organisations into scope where an AI system is placed on the EU market or AI output is used in the EU. Having an EU client alone is not the legal test.


Final takeaway

For research teams, the practical question is not simply whether AI appeared somewhere in the process. It is what the AI did, who interacted with it, what was produced, where the output went and who reviewed it.


Article 50 creates real transparency duties around direct AI-human interaction, certain synthetic content and qualifying public-interest publications. It does not require every AI-assisted research output to carry a visible label.


And Article 50 is only one layer. MRS, ICC/ESOMAR, data-protection rules and client contracts can create broader expectations.

The strongest position is not “label everything” or “say nothing unless the law forces us to”.


It is being able to explain where AI was used, how the data was handled, what a researcher checked and who stands behind the final insight.



Key sources


A note on AI used for this article:
AI helped us research this topic, compare sources, test interpretations and draft parts of the article. We then reviewed the underlying guidance, challenged and changed conclusions, rewrote the text and take editorial responsibility for what you are reading.
Based on our reading of Article 50, that level of human review and editorial control means we do not believe this article requires an AI-generated label. We are telling you anyway — because an article about AI transparency seems like a particularly odd place to hide how AI was used.
If you think we have interpreted something incorrectly, please let us know. We will happily revisit it.