Who has to say it was AI?
Transparency and editorial responsibility under Article 50(4) of the EU AI Act
On 2 August 2026 a new EU transparency rule began to apply to AI generated text published to inform the public. It does not ban newsrooms from using AI, and it does not require a label on everything AI touches. It ties the duty to disclose to how editorial oversight and responsibility are organised.
That makes it a question about how journalism is run, not about technology. We are studying how European news organisations read the new rule, what they tell their audiences, and what they say they do inside the newsroom.
Two different obligations, often confused
Article 50 of the EU AI Act became applicable on 2 August 2026. Two of its provisions get mixed up in coverage, and they fall on different parties.
Falls on providers
Companies that build generative AI systems must mark outputs in a machine readable way so they can be detected as artificially generated. This is the technical watermarking layer, and it is mostly not what news publishers are subject to.
Falls on deployers
Organisations using an AI system under their own authority. Where they use it to generate or manipulate text published to inform the public on a matter of public interest, the artificial origin must be disclosed, subject to an exception. This is the provision that reaches newsrooms.
Article 50(4) also covers deepfakes. This project examines only its second limb, on public interest text.
The label is the last question, not the first
Three conditions decide whether a disclosure is required. Each one has to be satisfied before the next is reached.
Choose an answer at each stage to follow the path.
A journalist writing unaided, or a tool used outside the organisation’s control, is outside the provision.
The Commission reads this broadly: politics and democratic process, public services, justice, fundamental rights, public security, health, environment, consumer safety, and economic, financial, scientific and cultural matters relevant to public debate.
Review means a real examination of the substance by someone with the judgement to do it. Editorial control means authority to approve, change or reject. Fixing spelling does not count. Editorial responsibility means someone ultimately carries legal responsibility. Both conditions are required.
The question, and four parts of it
How are European news organisations interpreting and implementing Article 50(4)’s transparency rules for AI generated public interest text, and how visible are those arrangements to the public?
- SQ 1
What do news organisations publicly disclose about their use of generative AI, human review, editorial control and editorial responsibility?
- SQ 2
How do newsrooms themselves say they are implementing the rule, including when they label AI generated text and when they rely on review or editorial control?
- SQ 3
How do these practices vary across ownership, editorial governance model and national media environment?
- SQ 4
How do practices change during the first year the rule applies?
Three sources of evidence, because one is not enough
A website audit on its own can tell you what an organisation publishes about itself. It cannot tell you what happens before an article goes up. The comparison between the two is where the interesting findings sit.
Legal and regulatory mapping
The environment each country’s newsrooms operate in.
- Which authority supervises this, and whether one has been designated at all
- Whether national guidance exists
- How a member of the public would complain
- Press council or industry guidance on AI
Public facing audit
What each outlet publishes about itself, recorded with evidence.
- Public AI policy, and what it permits
- What it says about review and responsibility
- Disclosures observed, with wording and placement
- Reader facing AI features
Newsroom survey and interviews
What the newsroom says it does, reported as self described.
- Where generative AI is used, and by whom
- What happens before publication, and who can reject
- Who carries responsibility
- When the outlet considers the exception to apply
Every observation carries its own evidence
Country, type, ownership, reach, State Media Monitor classification, parent organisation.
Whether a policy could be found, where, when it was published, and whether the AI Act or the transparency Code of Practice is mentioned.
What the policy says AI may be used for: drafting, summaries, translation, headlines, images, data work.
Whether review or editorial control is described, what kind, and whether a role or body is identified.
Disclosures observed in sampled material: wording, position, visibility.
Summaries, chatbots, translation tools, recommendation features.
URL, archived snapshot or local capture, whether archiving succeeded, and the date of coding.
Cannot determine, conflicting information, not located. How much of this we record is itself a finding about how legible these organisations are to their readers.
We are not detecting AI, and we are not judging compliance
We use no detection tools and draw no conclusions from how a piece of writing sounds. Detection of AI text is unreliable, and a study resting on it would deserve to be ignored.
That has consequences we would rather state plainly than bury in a footnote. Five rules bind everything we publish.
- The law does not require a newsroom to publish an AI policy. Not having one is a fact about transparency, not a violation.
- The law does not appear to require an outlet to name its responsible editor publicly. Not naming one is not a violation either.
- An article with no disclosure may simply have been written by a person, or fall outside the rule, or fall within the exception.
- A visible label is not an admission. Some outlets will label generously by choice.
- We write that no disclosure was observed in the material we sampled. We never write that AI material was unlabelled, unless the outlet has said the material was AI generated.
What we can tell you is how visible and specific these arrangements are in public, what newsrooms say they do internally, where those two things diverge, and how it all varies across ownership types, media systems and countries that have or have not built any capacity to enforce this.
Three countries first, then a decision
We begin in Hungary, Romania and Germany, a small group of outlets in each, every one coded independently by two researchers. That first stage decides whether the wider design is workable, which questions researchers can actually answer, and how far it can stretch. The ten country frame below is a proposal until then, not a commitment.
Under way, first report in November
The rule is new and practice is still forming, so the study runs in stages rather than arriving at one answer. We publish what we can stand behind at each stage, and say what remains open.
Building the instrument and testing it on a small group of outlets across three countries, with every one checked independently by two researchers. Mapping the regulatory environment country by country. Opening the first conversations with newsrooms.
First report. What we learned from the first stage, the regulatory picture across the countries in scope, and early findings on what outlets are publishing about their own arrangements. Published with the instrument itself.
The wider audit, the newsroom survey and interviews, and follow up rounds on the same outlets. This is where change over the first year becomes visible, and where the comparison between public transparency and internal practice can be made.
Instruments first, findings second
- Nov 2026First reportThe regulatory picture country by country, early findings on what outlets publish about their own arrangements, and an honest account of what the method turned out to be able to see.
- Nov 2026The instrumentPublished alongside the report, together with the drafts it replaces, so the revisions are visible and anyone can repeat or challenge the work.
- 2027Comparative report and open datasetThe wider sample, with every observation carrying its own archived evidence and a record of whether archiving succeeded.
- 2027Methods paperOn measuring public transparency arrangements when the underlying legal ideas, substantive review and editorial responsibility, sit inside organisations and cannot be observed from outside.
- 2027Country notes and follow upsWritten with the researcher who did the fieldwork, in the national language where they prefer, with repeat rounds on the same outlets.
The full description of the study: what Article 50(4) requires, the research questions, the three sources of evidence, the limits of what the research can establish, and the countries and outlets in scope.
What this study does not say. We record what outlets publish about their arrangements and what newsrooms tell us they do. Whether an outlet complies with Article 50(4) depends on what happens inside it, which cannot be established from outside. We do not allege that any outlet is in breach, we make no claim about whether any particular article was written by AI, and anything a newsroom tells us is reported as self described rather than verified.
Media and Journalism Research Center · mjrc@journalismresearch.org
Invest in independent media research and join a community of practice.
MJRC takes no government money — so the questions we chase are set by the field, not by a funder. A contribution keeps our research and global analysis independent, and buys you a seat closer to the work: early access to new findings, invitations to small-group briefings, a place in our Supporters Circle, and the option to be listed on our Supporters Page.