Do AI-Generated Images Need Labels in the EU Now? What Article 50 Actually Requires

No: the EU AI Act does not require every AI-generated image to carry a visible “AI” label.

The rule that started applying on August 2, 2026 is more specific. Article 50 separates two jobs that are easy to confuse:

• AI system providers generally have to make synthetic audio, image, video and text outputs machine-readable and detectable as AI-generated or manipulated.
• Professional deployers — businesses, organisations and people using AI professionally — have visible disclosure duties for deepfakes and for certain AI-generated or manipulated public-interest text.

That distinction matters. A provider-side provenance mark hidden in metadata is not the same thing as a visible label a viewer can understand. And a professional designer creating an obviously fictional illustration is not automatically in the same situation as a company publishing a realistic fake video of a real chief executive.

Rules check: August 9, 2026. The European Commission’s Article 50 transparency guidelines were published on July 20, and the obligations began applying on August 2. A limited transition until December 2, 2026 applies to the provider-side marking and detection obligation for qualifying AI systems placed on the market before August 2. Content generated before August 2 does not need to be labelled retroactively. This article is general information, not legal advice.

The quickest way to understand the rule

Think of Article 50 as two different layers.

The European Commission explicitly says deployers cannot rely only on the provider’s machine-readable mark when a deepfake needs disclosure. The viewer must receive the disclosure clearly, no later than first exposure.

So “the model added metadata” and “the audience was told this is AI-generated” solve different compliance problems.

First question: are you the provider or the deployer?

This is the most useful classification to make before thinking about labels.

Provider

A provider develops an AI system, or has one developed, and places it on the EU market or puts it into service under its own name or trademark.

For generative systems, Article 50(2) generally requires providers to make synthetic outputs — audio, images, video and text — machine-readable and detectable as AI-generated or manipulated.

The Commission says that requirement should use techniques that are effective, interoperable, robust and reliable as far as technically feasible, taking into account the type of content, implementation cost and the state of the art.

Deployer

A deployer uses an AI system under its authority. For Article 50, this normally means professional use by a company, organisation, public body, freelancer or other professional actor.

The Commission gives an important organisational example: when employees use AI under a company’s authority, the company is the deployer, rather than every designer, marketer or journalist becoming a separate deployer.

The deployer is the role most ordinary publishers, agencies, ecommerce businesses and professional creators need to understand, because this is where the visible disclosure duties appear.

Personal use is treated differently

The Commission’s FAQ says a natural person using an AI system in a personal, non-professional capacity is excluded from the deployer definition.

Its example is a person creating deepfakes and sharing them on social media as a personal activity. If the activity becomes professional — for example, it regularly produces economic benefit or forms part of a business, trade, occupation or freelance activity — the analysis changes.

This does not mean personal deepfakes are automatically lawful or platform-compliant. Other privacy, defamation, intellectual-property, election, consumer-protection, criminal and platform rules can still apply. It only explains the scope of this particular AI Act deployer obligation.

Which images, audio and videos need a visible disclosure?

For professional deployers, the key category is a deepfake.

The AI Act definition covers AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and could falsely appear to a person to be authentic or truthful.

The Commission breaks that into three cumulative questions:

1. Resemblance: does the content closely resemble the subject?
2. Existing or plausibly existing subject: is it depicting a person, object, place, entity or event that exists, could plausibly exist, or could plausibly have existed?
3. False appearance of authenticity: could the content mislead the audience about whether it is real or truthful?

That third question prevents “AI-generated” and “deepfake” from becoming synonyms.

An obviously fantastical illustration may be generated entirely by AI without falsely appearing to document a real event. A photorealistic clip of a public figure appearing to say words they never said is a much clearer deepfake case.

The intended audience and context matter too. The Commission says the assessment can consider resemblance, the substantive message, the deployment context and what the audience reasonably expects.

A practical decision table for creators and businesses

The table below is a screening framework, not a substitute for legal advice in a difficult case.

The most common mistake is to ask only “Was AI involved?” The better questions are what did AI change, how will the audience interpret it, who is publishing it, and what role does the publisher have under the Act?

Public-interest text has a separate rule

Images, audio and video use the deepfake test. Text has a different trigger.

Article 50(4) requires deployers to clearly label AI-generated or manipulated text when all three of these conditions are met:

1. the text is published;
2. it is intended to inform the public;
3. it concerns a matter of public interest.

The Commission lists examples including politics and democratic processes, public administration, justice and law enforcement, fundamental rights, public security, public health, environmental protection, consumer safety, and economic, financial, political, scientific or cultural developments relevant to public debate.

That is much narrower than “all AI-written text on the internet.”

A synthetic product description, fictional story or internal draft is not automatically covered by this particular public-interest-text rule simply because a model wrote it.

Human review can remove the public-interest text label requirement — but a skim is not enough

Article 50 contains an important exemption for qualifying public-interest text that has undergone human review or editorial control and where a natural or legal person holds editorial responsibility.

The Commission’s guidance makes the threshold more concrete.

Human review means deliberate examination of the substance by people with relevant knowledge and professional judgement. Editorial control means a responsible editor or entity has real authority to approve, alter or reject the substance, including checking information and source trustworthiness.

The Commission specifically says superficial, formal or procedural checks such as spelling or grammar correction are not enough.

That creates a useful workflow rule for publishers:

If the process is “AI writes → human glances → publish,” do not treat the human as a magic exemption.

A stronger process is:

AI draft → source verification → factual edits → substantive editorial decision → named responsibility → publication

Keeping a simple record of that review is useful even when no visible AI label is ultimately required.

What does a compliant visible disclosure look like?

For deepfakes, the Commission says disclosure should be clear and distinguishable, understandable and perceivable without special technical tools, and provided no later than first exposure.

That means an invisible provenance field alone is not enough for the deployer obligation.

Depending on the format, a disclosure might be:

• a visible text label near an image;
• an on-screen notice in a video;
• an audible disclosure for audio;
• a platform overlay that remains clear at first exposure.

The EU has also released a set of optional icons for AI-generated and AI-modified content. The icon set includes basic, fully AI-generated and partially AI-modified variants.

Using the official icon is not mandatory, and the Commission says the icon by itself does not establish legal compliance. It is a practical option for making disclosure more consistent.

The Commission’s icon guidance recommends that labels be clearly visible, avoid jargon, support accessibility, and remain associated with content when it is reshared or downloaded where the Code of Practice applies.

Creative and satirical work gets a more flexible treatment

A deepfake can still exist inside a film, satire, artwork or fictional programme. But the AI Act does not require disclosure to wreck the experience.

For deepfakes that are part of an evidently artistic, creative, satirical, fictional or analogous work, the transparency obligation is limited to disclosure in an appropriate manner that does not hamper the display or enjoyment of the work.

This is another reason the slogan “every AI image needs a watermark” is misleading.

A commercial fake testimonial designed to look authentic and a clearly fictional visual effect may both involve generative AI, but the audience context and disclosure treatment are not the same.

Provider marking has its own exceptions

The provider-side requirement is broad, but it is not unlimited.

The Commission says some outputs fall outside the marking obligation, including:

• short sequences of numbers, symbols or letters;
• source code;
• machine-to-machine outputs never exposed to humans;
• outputs used only inside closed-loop industrial or product-development environments, unless they become the final output.

The marking obligation also does not apply when the AI system is performing an assistive function for standard editing.

The guidelines additionally describe a narrow business-to-business or industrial-context exemption when specified conditions are met.

For a normal creator using a third-party image generator, these provider obligations are mostly something the tool maker must implement. The creator’s own question is usually whether their professional publication creates a separate human-facing disclosure duty.

The December 2 transition is narrower than it sounds

There is a limited grace period, but it should not be read as “AI labelling is delayed until December.”

The Commission’s current FAQ says:

• Article 50 applies from August 2, 2026.
• AI systems placed on the market before August 2 receive a limited transition until December 2, 2026 for the provider-side marking and detection obligation under Article 50(2).
• Content generated and already made available before August 2 does not need to be labelled retroactively.

The transition therefore does not erase the August 2 start date for the other transparency duties.

For businesses publishing new content now, the safer operational approach is to classify the content under the current rules rather than assume December is a blanket deadline.

A seven-step publication checklist

For a small business, agency, publisher or professional creator, the following process is more useful than applying an “AI” badge to everything.

1. Record whether the use is professional

If the content is created for a business, client, publication, monetised account or professional activity, treat deployer analysis as relevant.

2. Record what AI actually did

Was the output fully generated, was a real asset materially manipulated, or did AI only perform routine editing?

Do not rely on memory a month later. Store the model/tool name and a one-line description of the transformation with the project.

3. For image, audio or video, run the deepfake test

Ask whether the output resembles a real or plausibly real person, object, place, entity or event and could falsely appear authentic or truthful to the intended audience.

If the answer is clearly yes, plan a visible disclosure.

4. For published text, run the public-interest test

Ask whether the text is being published to inform the public about a matter of public interest.

If yes, determine whether there has been genuine substantive human review, editorial control and editorial responsibility.

5. Separate visible disclosure from provenance

Preserve machine-readable provenance or content credentials when the tool provides them. But do not assume hidden metadata satisfies a human-facing disclosure obligation.

6. Put the disclosure where the audience encounters the content

For qualifying deepfakes, the Commission says the disclosure should arrive no later than first exposure. Do not bury it on an unrelated legal page.

7. Check the second rulebook: the platform and sector

The AI Act is not the only source of labelling duties. Advertising rules, marketplace policies, election rules, professional standards, platform terms and sector-specific law can demand additional disclosure even when Article 50 does not.

This is especially important for advertising, property, health, finance and political communication.

A simple internal content log can prevent most confusion

Small teams do not necessarily need a complex compliance system. A lightweight table is enough to make the decision reproducible:

The value of the log is not bureaucracy for its own sake. It stops a team from making inconsistent decisions based on whether somebody happens to remember that AI was used.

The rule is about deception risk, not punishing AI use

Article 50 makes more sense when viewed through the audience rather than the tool.

The strongest disclosure duties appear where a person could reasonably misread synthetic material as authentic: a fake recording, manipulated real-world image, or unreviewed AI publication on a public-interest subject.

Other uses receive narrower requirements or exceptions because the audience is less likely to be deceived, because a responsible human has substantively reviewed the information, or because the AI is acting only as an editing aid.

That produces a better default than “label everything”:

Preserve provenance broadly; use clear visible disclosure where the content could mislead or where Article 50 specifically requires it.

What to watch next

The practical implementation is still young. Three developments will matter over the next few months.

Whether platforms standardise the EU labels

The Commission has published optional icons, but platforms may use their own labels, overlays and metadata systems. Compatibility between those approaches will determine how well disclosure survives reposting.

How machine-readable marking holds up after editing

Provider marks are supposed to be robust and interoperable as far as technically feasible. Real-world crops, screenshots, re-encoding and social-media processing will test how durable those signals are.

How authorities interpret borderline professional content

Freelancers, monetised creators, small publishers and mixed personal/professional accounts can sit near the line between personal activity and professional deployment. Where the stakes are meaningful, the Commission’s guidelines and national authority guidance are better sources than social-media summaries.

Bottom line

The EU’s new transparency rules do not mean every AI-generated image needs a visible badge.

The practical split is this:

• providers generally handle machine-readable marking of synthetic outputs;
• professional deployers visibly disclose qualifying deepfakes;
• public-interest AI text needs labelling unless it has genuine human review/editorial control and editorial responsibility;
• obviously creative or fictional work can receive a less intrusive disclosure treatment;
• standard editing and several technical contexts have specific exceptions;
• personal, non-professional use sits outside the deployer definition in Article 50;
• the limited December 2 grace period applies to the provider-side marking requirement for certain pre-August systems, not to every transparency duty.

For creators and small businesses, the low-regret workflow is to classify the content, preserve provenance, document real human review, and make the disclosure visible when the audience could otherwise mistake synthetic content for reality.

This article provides general information about the EU AI Act and the European Commission’s current guidance as checked on August 9, 2026. It is not legal advice. Specific campaigns, regulated sectors and cross-border publishing can involve additional rules.

Sources

• European Commission — Transparency obligations under Article 50 of the AI Act (Q&A)
• European Commission — Guidelines on transparency obligations for providers and deployers of AI systems
• European Commission — Code of Practice on Transparency of AI-generated Content
• European Commission — EU icons for labelling AI-generated content
• European Commission — Commission publishes the Article 50 transparency guidelines, July 20, 2026
• EUR-Lex — Regulation (EU) 2026/1744 amending the AI Act implementation timeline