From 2 August 2026, newsrooms publishing automated public-interest text or synthetic media may have to tell audiences that AI was used. The rules do not require a label on every AI-assisted article, but they do require meaningful editorial oversight.
Author: Aleksandar Manasiev
If your newsroom uses AI to draft articles, generate images, clone voices or publish automated updates, new EU transparency rules may change what you need to tell your audience.
From 2 August 2026, Article 50 of the EU AI Act requires disclosure for certain AI-generated material. This does not mean that every article produced with the help of AI needs a label. What matters is what the system produced, whether a journalist or editor genuinely reviewed it, and who accepted responsibility for publication.
The rules cover AI-generated text, images, audio and video. Whether a disclosure is required will depend on how the material was produced, reviewed and published. For editors, the key distinction is between content that has passed through a genuine editorial process and content released with little or no human scrutiny.
On 31 July, the European Commission announced that about 190 organisations had signed the new Code of Practice on Transparency of AI-generated Content. The signatories include major AI providers such as Anthropic, Google, Meta, Microsoft, Mistral and OpenAI, as well as smaller technology companies and professional users of generative AI.
Signing the code is voluntary. Complying with the underlying transparency rules is not.
Providers and newsrooms have different responsibilities
The code separates the responsibilities of providers and deployers.
Providers are companies that develop AI systems or place them on the market. Where technically feasible, they must make AI-generated outputs detectable through machine-readable markings. These may include metadata, watermarking or other technologies that indicate that material was generated or manipulated by AI.
A newsroom using an AI system professionally will usually be treated as a deployer. Responsibility for reviewing, labelling and publishing AI-generated material rests with the newsroom, rather than with each journalist working under its authority.
According to the Commission’s guidance, the same principle applies when freelancers or contractors use the system on an organisation’s behalf and under its control.
A media organisation may also become a provider if it develops an AI tool, or commissions one, and offers it under its own name.
Not every AI-assisted article needs a label
Article 50 applies to AI-generated or manipulated text published to inform the public about matters of public interest.
The Commission interprets that category broadly. It includes politics, public administration, justice, public health, security, the environment, consumer safety, the economy, science and culture.
Journalism, however, is covered by an important exception.
AI-generated text does not require a mandatory label if it has undergone substantive human review or editorial control and a natural or legal person accepts editorial responsibility for its publication.
A quick read-through is not enough. Nor can an editor simply correct a few sentences and press publish.
The Commission’s Article 50 guidance says human review must involve a deliberate examination of the substance by someone with the relevant knowledge and professional judgement. Editorial control also requires the authority to amend or reject the text, verify facts and assess whether the sources are reliable.
Checks limited to spelling, grammar, formatting or other procedural matters do not meet that standard.
In practice, the workflow matters more than the prompt. An article drafted with AI may qualify for the exception if a journalist or editor checks the evidence, returns to the original sources, corrects unsupported claims and accepts responsibility for the published version.
An automatically generated report uploaded without that scrutiny is likely to require a label.
Newsrooms may still choose to disclose AI assistance under their own editorial policies, even when the law does not require it. The exception sets a legal minimum; it is not an instruction to conceal significant AI involvement from readers.
Synthetic images, voices and video are treated differently
The editorial-review exception applies to public-interest text. It does not remove the disclosure requirement for deepfakes.
Under the AI Act, a deepfake is an AI-generated or manipulated image, audio recording or video that resembles an existing person, object, place, organisation or event and could falsely appear authentic.
A newsroom publishing an AI-generated reconstruction of an event, a cloned voice, a face swap or fabricated footage must disclose that the material was artificially generated or altered if audiences could mistake it for reality.
Editorial approval does not remove that obligation.
The disclosure must be clear and distinguishable when the audience first encounters the content. A notice hidden at the bottom of an article, or included only in technical metadata, may not be sufficient.
The EU has released optional labels for fully AI-generated and partially AI-modified content. Newsrooms are not required to use those particular labels, but any notice should be visible, understandable and accessible.
Where possible, the disclosure should remain attached to the material when it is downloaded or shared on another platform. The EU’s labelling guidance recommends plain language and says labels should be compatible with assistive technologies.
Not every synthetic illustration will meet the legal definition of a deepfake. A clearly stylised image that no reasonable viewer would interpret as documentary evidence may fall outside the definition.
Even so, labelling synthetic visuals is generally the safer editorial choice, particularly when they depict real people, places or events.
Five common newsroom scenarios
1. AI corrects grammar or proposes headlines
A journalist writes the article, while an AI tool checks spelling, shortens sentences or suggests alternative headlines. An editor reviews and approves the final version.
A public AI label will generally not be required under Article 50.
The use of the tool should still comply with the newsroom’s internal policy, especially when journalists enter unpublished, confidential or sensitive information into an external service.
2. AI produces a first draft from verified material
A reporter supplies documents, notes or structured data. An AI system generates a draft. A journalist then checks each claim against the original material, reviews the sources, rewrites weak passages and submits the article for editorial approval.
In that situation, the exception for human review and editorial responsibility may apply.
The newsroom should record who reviewed the text and what checks were carried out, rather than relying on a general statement that a human was “in the loop”.
3. An automated system publishes public-interest briefs
A system produces weather warnings, election results, economic updates, public notices or municipal news and publishes them without substantive human review.
That material is likely to require a visible disclosure stating that it was generated or manipulated by AI.
An automated spelling or grammar check before publication does not amount to editorial control.
4. A report uses a synthetic image, cloned voice or reconstructed video
Synthetic media should be labelled when it could be mistaken for an authentic representation of a real person, place or event.
The label should appear on or next to the content when the audience first sees or hears it. For a video published on social media, this may require an on-screen notice rather than a disclosure placed only in the caption.
Editorial review does not cancel the labelling requirement.
5. A newsroom introduces an AI chatbot or virtual presenter
Users should be told from the beginning that they are interacting with AI, unless that fact is already obvious from the context.
Whether the main legal responsibility belongs to the technology provider or the newsroom will depend on how the system was developed, branded and offered to the public.
What newsrooms should do now
Newsrooms do not need to place an “AI-generated” label on every piece of content produced with the help of an automated tool.
They do need to identify where AI enters the publishing process and determine which uses require editorial review, public disclosure or both.
A practical response should cover the following areas.
1. Map current AI uses
Identify every point at which AI is used in the newsroom, including transcription, translation, summaries, headline generation, automated briefs, images, cloned voices and chatbots.
The review should also cover tools used by freelancers and contractors working on the organisation’s behalf.
2. Define substantive human review
An internal policy should state who is responsible for checking facts, quotations, sources, context and legal or ethical risks.
It should also identify who has the authority to change the material, delay publication or stop it altogether.
3. Prepare standard labels
Newsrooms should draft plain-language disclosures that can be used consistently across formats.
Examples include:
“This image was generated with AI.”
“This report was generated automatically and was not reviewed before publication.”
The wording should tell audiences what happened without relying on technical terminology.
4. Build disclosure into the publishing system
Labels should work across the newsroom’s website, content-management system, social platforms, video exports and downloadable files.
A disclosure that disappears when material is copied, downloaded or shared offers limited protection to the audience.
5. Check suppliers
Newsrooms should ask AI vendors whether their systems preserve machine-readable provenance information and support the EU’s marking requirements.
They should also check whether their own editing, export or publishing processes remove metadata or other provenance signals from files.
6. Keep a record of editorial review
For AI-assisted public-interest text, the newsroom should record who reviewed the material, which sources were checked and who gave final approval.
A record of substantive editorial work is more useful than a note saying only that a person clicked “publish”.
7. Review the code before signing
A newsroom may join Section 2 of the Code of Practice as a deployer, but it must accept the section as a whole.
Signing may provide a recognised way to demonstrate efforts to comply, but it is not an automatic defence against enforcement. The Commission’s assessment says adherence to the code is not conclusive evidence of legal compliance.
The rules can also reach beyond the EU
The territorial scope of the AI Act matters to newsrooms in the Western Balkans and other countries outside the European Union.
Article 2 can apply to providers and deployers established in third countries when the output produced by an AI system is used in the EU.
That does not bring every non-EU publisher within the Act in every circumstance. Newsrooms serving EU audiences, producing material for EU partners or distributing AI-generated content within the Union should nevertheless assess whether the rules apply to their work.
The potential penalties are significant. According to the Commission’s guidance, breaches of Article 50 can lead to fines of up to €15 million or 3 per cent of worldwide annual turnover. Proportionality and different limits for smaller companies must be taken into account.
The most important change for newsrooms, however, will be operational.
Organisations that can show where AI was used, who checked the output and who accepted responsibility for publication will be better placed than those relying on an informal instruction to take a quick look before pressing publish.
Under the new rules, meaningful human review is not a ceremonial final click. It is part of the editorial process.
(This text was written and reviewed by the editor with support from artificial intelligence tools for language editing and stylistic refinement. More on how NarativAi uses AI — Link)