AI content labelling in 2026: a news website uses a synthetic image, a company publishes an avatar, and a portal automatically distributes an AI-written news story. Do all three need the same label? No. From 2 August 2026, the transparency obligations in Article 50 of the EU AI Act apply, but they differ according to who provides the system, who uses it and what type of content reaches the public.
The key distinction is between technical marking that software can detect and information visible to the reader. The first obligation principally concerns the provider of the AI system; the second may fall on the party using it professionally to create a deepfake or text on matters of public interest. There is also an important exception for text substantially reviewed by a human or subject to editorial control, with someone assuming responsibility for publication.
This is a practical analysis for newsrooms and businesses: 12 examples, verification methods, wording templates and editorial-control steps. For members of the public who want to check a suspicious clip, InfoBuz.ro already has a guide to AI labels and checking deepfakes. The question here is different: what do you do before you click “Publish”?
AI content labelling in 2026: the four obligations under Article 50
The core text is Article 50 of the AI Act, supplemented by guidelines published by the European Commission on 20 July 2026. The provider of an AI system is known as the provider, while the organisation or individual using it professionally under their own authority is the deployer. A publisher using an external service does not automatically become the provider of that system.
| Legal basis | Who is principally covered | What it entails |
|---|---|---|
| Art. 50(1) | Providers of interactive AI systems | Systems must be designed to tell people when they are interacting with AI, unless this is obvious from the circumstances and context of use. |
| Art. 50(2) | Providers of generative systems | Audio, image, video and text outputs must be marked in a machine-readable and detectable format, within the limits of and subject to the exceptions in the law. |
| Art. 50(3) | Users of emotion-recognition or biometric-categorisation systems | Exposed persons are informed that the system is operating, while personal-data rules must also be observed. |
| Art. 50(4)–(5) | Professional users of AI publishing certain materials | Deepfakes and relevant AI-generated text on matters of public interest must be disclosed in the cases provided for; the information must be clear and available at the first exposure. |
Not every AI-generated photo is a deepfake
The legal definition of a deepfake concerns AI-generated or manipulated images, audio or video that resemble existing or plausible people, objects, places, entities or events, and could falsely appear authentic or truthful. The context in which it is presented is decisive. Visibly fantastical artwork is not, in itself, equivalent to a video simulating a statement by a real person.
A generic AI image of a house with photovoltaic panels does not automatically fall into the deepfake category. But if the same image is presented as evidence that a solar park has been built at a real address, the assessment changes. In news publications, the caption “AI-generated illustration — does not depict the actual location” remains a prudent practice, even where no specific obligation to apply a visible label can be asserted.
Twelve practical situations: applicable rule and editorial decision
| Hypothetical situation | Possible interpretation | Recommended action |
|---|---|---|
| 1. An AI image of an invented accident presented as real | May be a deepfake of an event | Do not present it as evidence; disclose it clearly and verify the claim. |
| 2. A generic AI illustration for an article on energy | Not automatically a deepfake | Voluntarily add a clear caption explaining its illustrative nature. |
| 3. A real photograph with a minor lighting correction | May constitute standard editing | Keep the original and check that the image’s meaning has not changed. |
| 4. A crowd added with AI to a protest photograph | May create the appearance of a false event | Assess it as a possible deepfake; do not present it as journalistic evidence. |
| 5. A cloned voice of a real doctor used for an invented statement | May be an audio deepfake | Disclose the use of AI and check the person’s rights. |
| 6. A generic synthetic voice reads verified text | Not every AI voice imitates an existing person | Voluntary disclosure is useful; do not automatically assume it is a deepfake. |
| 7. AI-created public-interest news published automatically | Text potentially subject to disclosure under Art. 50(4) | Display the applicable notice and introduce human control before publication. |
| 8. AI text checked by a responsible editor | May fall within the editorial-control exception | Document substantive review, not merely proofreading. |
| 9. AI correction of diacritics or spelling | Not automatically substantial generation | Check whether claims or meaning have changed. |
| 10. Obvious parody using an AI reconstruction of a politician | If it is a deepfake, disclosure adapted to the creative work applies | Provide appropriate information without unnecessarily compromising the presentation. |
| 11. A chatbot conversing with website visitors | Interactive-system design obligation under Art. 50(1) | Ensure an obvious notice appears at the beginning of the conversation. |
| 12. An AI avatar presented as a real customer in an advertisement | Not every avatar is a deepfake; it may constitute misleading commercial practice | Do not invent testimonials, and check consumer-protection rules. |
A small business does not automatically become the technology provider
For a website editor using an existing AI tool, the usual role is that of professional user. If the editor develops or markets an AI system in its own name, provider obligations may also arise. The distinction matters: responsibility for a technical marker built into a system is not the same as responsibility for how a publication presents it to readers.
Moreover, content that does not meet the definition of a deepfake may still raise separate issues: image rights, copyright, misleading advertising or data protection. A label does not turn a false message into accurate information.
Public-interest text: what genuine human editorial control means
For certain AI-generated or manipulated text published to inform the public on matters of public interest, the AI Act provides for a disclosure obligation. However, Article 50(4) includes an exception for material that has undergone human review or editorial control and for which a natural or legal person assumes editorial responsibility for publication. This is not permission to publish any text automatically simply by adding an editor’s name.
In its FAQ on transparency obligations, the European Commission explains that substantive review differs from a spell check or a formal procedure. The editor must be able to verify the factual accuracy, the coherence of conclusions and the soundness of sources; they must be able to amend or reject the material.
When an editor checks only diacritics
Suppose a tool drafts a news item about a medical report and the article appears automatically after an operator corrects punctuation. In such a case, it cannot be asserted without further evidence that substantial editorial control took place. Simple proofreading is no substitute for source verification or a proper understanding of a study.
When an editor checks sources, recalculates figures and assumes responsibility for publication
A second example: the journalist opens the primary sources, checks every claim, removes unconfirmed passages, recalculates proportions, seeks explanations where appropriate and signs off on final approval. In such a case, the text exception may be relevant. Documenting the review is recommended editorial practice, even though Article 50 does not impose a universal duty to archive every prompt.
Technical marking versus a reader label: why they are not the same
Under its conditions of application, Article 50(2) requires providers of systems generating synthetic text, sound, images or video to design outputs that are marked in a machine-detectable format. Article 50(4) focuses on disclosure to the public for deepfakes and relevant text. The Commission clarifies that an invisible marker is not, by itself, sufficient information for people.
| Mechanism | Role | Limitation |
|---|---|---|
| Software-detectable marker | Technical indication of generation or manipulation | Does not guarantee visible disclosure and does not verify the truth of the account |
| Visible label | Tells the public that it is seeing or hearing synthetic content | Does not factually confirm what the content claims |
| Editorial control | Checks claims, context, data and sources | Cannot be replaced by a symbol or watermark |
Content Credentials and C2PA: provenance is not a certificate of truth
The C2PA standard describes mechanisms through which an application can record verifiable information about a file’s origin and modifications. A photograph may have authentic provenance yet still be shared with a false description. Conversely, an image may contain no verifiable metadata after compression or re-uploading, without that absence proving it was artificially generated.
For newsrooms, good practice is to retain the original file and provenance information where possible, without treating metadata as a verdict. Readers wishing to investigate a suspicious image can also consult InfoBuz’s guide to checking fake photographs.
EU icons are optional; legal transparency is not
The Commission has launched a set of icons for AI-generated or AI-manipulated content. They may provide a common visual language, but their use is voluntary. Where a disclosure obligation exists, it must be met regardless of whether a publisher uses the EU icons or another clear and accessible solution.
On mobile, a very small symbol or information hidden behind several menus may go unnoticed. The practical recommendation is a legible notice beside the relevant content, along with a description appropriate for people using screen readers. Explanatory text may be more useful than an icon whose meaning is not known.
Three editorial wording options for labelling AI content
The examples below are InfoBuz proposals for transparency. They are not legally mandated word-for-word formulas. They should be adapted to the truth about how the material was produced and verified.
| Situation | Illustrative wording |
|---|---|
| Generic illustration | “Illustration generated with artificial intelligence; it does not depict a photograph of an identified real place or event.” |
| Realistic synthetic audio/video reconstruction | “This sequence was created or modified with AI and is not an authentic recording of the event.” |
| AI-created public-interest news without editorial control | “This text was generated with AI and was not reviewed by a human editor before publication.” |
A 10-step editorial checklist for publishers and businesses
- 1. Identify the type of material: news, opinion, image, advertisement, video or chatbot.
- 2. Establish the legal role: system provider, professional user or another relevant category.
- 3. Define the AI intervention: proofreading, substantive drafting, summarisation, voice cloning or visual manipulation.
- 4. Check the appearance of authenticity of any person, location or event depicted.
- 5. Check primary sources: documents, figures, quotes, chronology and the limitations of studies.
- 6. Ensure substantive human review where the editorial exception for text is invoked.
- 7. Decide whether disclosure is required in relation to Article 50 and the applicable exceptions.
- 8. Test the label on mobile, in the article and in standalone previews.
- 9. Retain file provenance and, where appropriate, documentation of the review.
- 10. Take responsibility for published claims and correct errors transparently.
This protocol is an editorial tool, not a legal certification. It is particularly helpful to small publications that use several tools without a dedicated legal department. For the ethical implications of human control, see InfoBuz’s analysis, Human in the Loop.
AI Act timetable: 2 August and 2 December 2026 do not mean the same thing
The transparency obligations in Article 50 apply from 2 August 2026. Providers of generative systems placed on the market before that date have a limited transition period until 2 December 2026 for the technical marking and detectability requirement in Article 50(2). This deadline is not a general postponement of disclosure for deepfakes distributed thereafter.
According to the Commission’s clarifications, content generated and already made available before 2 August 2026 does not need to be labelled retroactively solely on the basis of Article 50, though authorities encourage voluntary transparency. In complex cases, the date of generation and the date of publication must be assessed separately.
For more on AI content, see also The internet after the internet: how much AI content is already online and what happens when machines write for other machines, which provides complementary detail on the subject.
Fines: maximum levels do not mean automatic penalties
Under the AI Act regime, certain breaches of obligations can result in penalties of up to €15 million or 3% of worldwide annual turnover, under the conditions set by the regulation. This is a maximum relevant to the category of infringements, not an automatic fine for every website using an AI illustration. Enforcement depends on the obligation actually breached, the authority’s procedure, proportionality and the operator’s circumstances.
In each specific case, the competent institution, the operator’s role and the applicable rule must be established. Some issues — such as misleading advertising or image rights — may also be addressed under regulations other than the AI Act.
Why this matters for independent Romanian websites
A local website may use AI to structure articles, translate, create images, produce subtitles or make summaries. The value of a responsible process is not the number of pictograms, but the newsroom’s ability to answer this question: which part of this claim has been verified, and which part is illustrative?
A generated image that purports to show a real building but alters its architecture can destroy trust, even if it looks striking. For documenting a real location, authentic and verified photography remains preferable. For conceptual analysis, an illustration clearly identified as such may be appropriate. Truthfulness in representation and transparency are editorial decisions, not merely technical ones.
Frequently asked questions about AI content labelling
Does every AI-generated photo have to be visibly labelled?
No. There is no single, identical obligation for every synthetic image. Providers have technical obligations, while for certain deepfakes the professional user must make a visible disclosure. A voluntary caption is often editorially advisable.
Is a label required for an article initially created with AI but checked by an editor?
Article 50 provides an exception for text subject to genuine human review or editorial control, with responsibility for publication assumed. Simple spelling correction is not equivalent to such a review.
What do C2PA markers prove?
They can provide information about a file’s provenance and history. They do not, by themselves, prove the truth of the image or its accompanying explanation.
Must the official EU icons be used?
No. The icon set is optional, but applicable disclosure obligations remain mandatory.
Is it legal to publish a cloned voice for satirical purposes?
For evidently artistic, satirical or fictional works that constitute deepfakes, the regulation provides for a disclosure approach adapted to the work. That does not automatically remove other obligations or rights of the person imitated.
If a label is missing, does that mean an image is real?
No. The file may have lost its metadata, been re-uploaded or been distributed without compliance with the rules. Its origin and the facts require separate verification.
Official documents and methodology
- AI Act — Article 50, transparency obligations.
- European Commission — questions and answers on Article 50.
- European Commission — guidelines for providers and deployers, 20 July 2026.
- European Commission — voluntary code of practice for AI transparency.
- European Commission — icons for labelling AI content.
- C2PA — specifications on file provenance.
Editorial note: This analysis reflects information and official documents consulted up to 14 November 2026. The 12 scenarios are hypothetical; they are examples for assessing risks, not definitive legal decisions. The precise application may depend on the actor’s role, the time of publication, the context and related legislation. This material does not replace legal advice for a specific situation.








Discuția cititorilor
Ai o opinie sau o întrebare despre acest articol? Spune-ne ce crezi.
Fii primul care deschide discuția.
Te rugăm să păstrezi un limbaj civilizat și să discuți la subiect. Comentariile spam sunt filtrate automat.