AI content labeling in the EU 2026: when should you indicate that text or an image was created with AI?

From 2 August 2026, the European Union has introduced AI Act Section 50 Transparency Rules. These include chat systems, generative AI content, deepfake-type images, certain AI-generated publications, and technical AI content tagging. European Commission guidelines explained in more detail how these obligations should be applied in practice.

However, a much simplified explanation can still be found online: if the text was written by ChatGPT or the image was created by generative AI, it is mandatory to write "Created with AI".

That's not the case.

The AI Act does not require a visible label to be placed on every article, product description, advertising copy, illustration, or book cover that was created with AI. The requirements vary depending on the the type of content, its purpose, the context of publication, and the human involvement in the final result. The European Commission also clearly states, that not all AI-generated or modified content needs to be visibly marked.

Therefore, the most important thing for businesses to understand is not „was AI used?“, but what the AI created, how the content will be used, and whether a human actually controlled the final publication.

What has changed since August 2, 2026?

Article 50 of the AI Act Several different transparency obligations are provided for. The European Commission approved, that these transparency rules have been in effect since 2 August 2026.

AI system providers, such as developers of generative AI platforms, are subject to requirements for the detection and labeling of technical AI content. At the same time, companies, organizations or professionals who use AI systems in their activities are under certain obligations. inform a person about the use of AI or clearly mark specific content. European Commission FAQ on Article 50 of the AI Act discusses these obligations of suppliers and users separately.

In practice, two cases are most relevant for business:

• AI created or modified deepfake content;
• AI created or significantly modified text on matters of public interest, if it has not undergone appropriate human review or editorial control.

The limits of these requirements are explained in more detail The European Commission's guidelines on transparency of AI-generated content.

This is a much narrower requirement than "label everything AI contributed to.".

Do I need to mark up texts created in ChatGPT?

Not everyone.

According to AI Act Section 50(4) The obligation of visible marking applies to AI-generated or manipulated text when it is published for the purpose of informing the public on matters of public interest. In the European Commission's explanations discusses in more detail how to determine whether a particular publication falls into this category.

European Commission Such topics include, for example, politics and democratic processes, public administration, law enforcement and justice, fundamental rights, public security, public health, environmental protection, consumer safety, and economic, financial, scientific or cultural developments that may be significant issues for public debate.

So just because a company has written a product description, commercial email, or promotional text with ChatGPT, does not in itself create a general obligation to add the tag "AI generated" next to it„.

What if AI helped you write an article or a book?

Here is one of the most important exceptions to the AI Act.

Even if the publication deals with a matter of public interest, its no need to mark as AI generated, if the text has undergone actual human review or editorial control and a natural or legal person assumes editorial responsibility for the final publication. This exception is explained by European Commission FAQ on Article 50 of the AI Act.

However, just reading the text or correcting the grammar is not enough.

European Commission defines human review as an assessment of the substance of the content by a person with sufficient knowledge and professional judgment. Editorial control means the real possibility of approve, modify or reject content, check facts and assess the reliability of sources. Merely formal correction of language or spelling is not considered such control.

The practical difference is very clear.

Insufficient control:

ChatGPT generates the article → corrects a few grammatical errors → publishes the text.

Real editorial control:

ChatGPT helps prepare a draft → a person checks information and sources → evaluates arguments → corrects inaccuracies → changes structure and wording → decides what gets published → takes responsibility for the final content.

In the second case, we are not talking about automatically published AI text, but about a human-edited publication that used AI in the creation process.

What common business texts do not usually require the tag?

Simply using AI to create content does not mean that visible AI markup automatically becomes mandatory.

For example, AI can help prepare:

• Product description;
• A draft of the service page;
• Email;
• Advertising text;
• Social media post;
• FAQ;
• The structure of the article;
• A draft of a book or guide.

There is no general AI Act rule for such content, requiring each publication to state „Created with AI.“ Articles on topics of public interest require additional consideration of human review, editorial control, and editorial accountability. This is discussed in more detail in In the European Commission's official explanation of Article 50 of the AI Act.

This is an important difference between AI-assisted the work process and the situation where AI-generated content is published essentially without human control.

Do I need to tag every image created by AI?

No.

Obligation for businesses to display visible markings according to AI Act Section 50(4) most relevant when the image, sound or video created or modified by AI corresponds to deepfake definition.

European Commission Deepfake is defined as AI-generated or manipulated image, audio, or video content that resembles a real or potential person, object, place, entity, or event and can mislead a person into believing that the content is authentic or real.

Therefore, it is customary to:

abstract illustration, infographic, book cover, icons, stylized workplace, illustrative AI robot or decorative product composition

just because it was created with generative AI usually doesn't make it a deepfake.

The European Commission directly highlights: Not all AI-generated or modified content needs to be visibly marked.

When should AI visuals be labeled?

The situation changes when the content can be considered an authentic representation of reality.

For example:

A purported photo of a real event is generated, even though the event did not take place.

A real person is depicted saying or doing something they did not actually do.

The image of a real place is altered in such a way that the audience may believe that this is how it actually looks.

An authentic photograph is replaced by AI in such a way that the essential information it conveys changes.

When assessing whether content can be considered deepfake, European Commission guidelines It is suggested to consider not only visual similarity, but also the message being conveyed, the context of publication, and what a specific audience reasonably expects.

Such content needs to be disclosed clearly and intelligibly to a person no later than the first time he or she encounters the content. A technical tag in the file metadata alone is not enough. More about providing information is explained in European Commission.

A simple formulation could be:

„"The image was created using artificial intelligence."“

or:

„"The image has been partially altered using artificial intelligence."“

The AI Act does not establish one universal phrase that everyone is required to use.

What about AI-generated art and creative illustrations?

Article 50 of the AI Act a special rule is provided for obviously artistic, creative, satirical, fictional and similar works.

If a deepfake-type element is used in such a work, the obligation to inform remains, but the disclosure may be provided in such a way that would not interfere with the display or use of the work. Practical principles of labeling are provided by European Commission AI content labelling guidelines.

This is another reason why it is not right to automatically apply the same large label to every AI visual.

What is the difference between visible marking and machine-detectable marking?

These are two different responsibilities that are often confused even in articles about the AI Act.

Machine marking

According to AI Act Section 50(2) generative AI systems suppliers must ensure that the synthetic text, sound, image and video they generate or manipulate are marked up in a machine-detectable format, to the extent technically feasible.

These may include technical mechanisms for determining provenance, metadata, watermarks or other detection technologies. European Commission guidelines This obligation of providers is separate from the disclosure of visible content to users.

This is first and foremost AI system supplier, and not the responsibility of the ordinary user.

Visible awareness

Another responsibility for a business or other user of an AI system is to, in certain cases, to inform clearly, that the content you see or hear has been created or modified by AI.

When such visible disclosure is required, a person must be able to understand it without a special technical tool. Therefore, file metadata alone is not sufficient. This is explained in more detail The European Commission's FAQ on Article 50 of the AI Act.

The EU already has AI content flagging icons

The European Union has announced an official set of AI content tagging icons for 2026. It includes a generic AI tag and variants to indicate content that is fully AI-generated or partially AI-modified.

However, there is an important difference:

The information obligation established by the AI Act is mandatory in certain cases, while the use of icons created in the EU is optional.

The European Commission clearly states, that the use of icons is optional. Their mere use does not in itself guarantee legal compliance.

If used, the Commission recommends that the label be clearly visible at the first contact with the content, not obscured by other elements and in plain language. It is also recommended to ensure its accessibility for people using assistive technologies. These recommendations are provided in on the official EU AI content icons page.

What is the AI-generated Content Code of Practice?

In 2026, the European Commission announced Code of Practice on Transparency of AI-generated Content – a practical code of transparency for AI-generated content.

It is designed to help providers and users of AI systems implement the requirements of section 50 of the AI Act regarding technical labelling of AI content, disclosure of deepfakes and certain AI-generated text publications. European Commission points out that the code covers specifically the obligations provided for in sections 2, 4 and 5 of Article 50 of the AI Act.

Joining the code is voluntary, but the requirements of section 50 of the AI Act themselves are not voluntary. The European Commission and the AI Board recognized the Code as an appropriate measure to justify compliance with these requirements.

Businesses that simply use ChatGPT to produce content do not automatically need to become signatories to the code. However, its principles are useful for developing their own internal rules for using and tagging AI content.

What should businesses do in practice?

A complex labeling system is not necessary for most small businesses. It is much more important to have a clear internal procedure.

The practical process might look like this:

1. Determine whether AI only assisted or created a substantial portion of the content.

AI grammar correction and a full AI-generated article are not the same. The limits of human review and editorial control are discussed The European Commission's explanations of Article 50 of the AI Act.

2. Evaluate whether the text informs the public about a matter of public interest.

If not, the special rule for marking texts in the public interest in section 50 of the AI Act does not generally apply. The limits of the application of publications in the public interest are explained by European Commission.

3. If the topic falls into the area of public interest, evaluate human control.

Did the person check the facts and sources? Could they have changed or rejected the text? Does the person or organization take editorial responsibility?

If so, the human review and editorial control exception described in In the official explanation of the European Commission.

4. For visuals, assess whether they may be misleading about reality.

If it is an obvious illustration, diagram, or fictional composition, the situation is one. If the AI image looks like an authentic photograph of a real person, place, object, or event, the rules of deepfake need to be considered. The European Commission provides explanations on the use of deepfake here.

5. When marking is required, present it clearly to the person.

Don't hide important information only in metadata or on a hard-to-find page. The AI origin of relevant content should be understood at first encounter. Practical labeling principles are provided by European Commission recommendations for AI content labeling.

A simple AI content markup table for business

Situation	Does section 50 of the AI Act generally require a visible AI tag?
ChatGPT helped prepare the product description	Not automatically
AI helped write advertising copy	Not automatically
AI-generated article on a topic of public interest, published without real human review	Yes
AI text has been substantially reviewed by the editor, fact-checked, and editorial responsibility assumed	There is an exception to this text marking obligation
AI-generated abstract book cover or diagram	Not automatically
AI-generated realistic photography of a supposed real event	Could be a deepfake – needs to be assessed and, if it meets the criteria, marked
Real person's face AI inserted into an event he didn't attend	Likely deepfake – mark
AI altered real photography so that it can be misleading about reality	You may need to mark

This table is a practical guide. The specific legal qualification depends on the content, its context and the way it is used. The main criteria can be checked European Commission FAQ on Article 50 of the AI Act and in official transparency guidelines.

Do I need to re-tag old AI content?

The European Commission's explanations of Article 50 of the AI Act The application of the rules to content and systems during the transitional period is also discussed.

Therefore, a company should not automatically conclude that, simply because of the new requirements, all articles or illustrations previously produced with AI must now be mass-labeled. It is important to assess the scope of the specific AI Act obligation on a case-by-case basis.

What does this mean for books, guides, and other expert content?

AI can be used for research, information structuring, text drafting, editing variations or visualisation. The mere fact of using AI does not automatically make the final book „AI content“ that must be visibly labelled under the AI Act.

The most important question is who controls the final outcome.

If the author or publisher checks sources, assesses the substance of the content, corrects errors, makes editorial decisions, can reject or change the text, and takes responsibility for publication, this is a completely different situation than the automatic publication of the AI result. European Commission It is this kind of real human review and editorial responsibility that distinguishes it as an important exception to AI text markup.

In practice, it is also worth preserving an internal trace of the editorial process: sources, drafts, revised versions, and the date of final review. The AI Act does not establish this as a universal mandatory documentation format for every publication, but it helps to show that human control was not just formal.

FAQ

Will all AI texts have to be marked up from August 2, 2026?

No. A specific obligation to provide visible attribution applies to AI-generated or modified text when it is published for the purpose of informing the public on matters of public interest. There is also an exception where the text has undergone appropriate human review or editorial control and editorial responsibility has been assumed. This is explained in The European Commission's official AI Act Article 50 FAQ.

Should I say "Created with AI" in the product description?

Just because ChatGPT or other generative AI was used to draft the product description, there is no general obligation under the AI Act. The limits of application of section 50 of the AI Act can be checked European Commission guidelines.

Does every image created by AI need to be labeled?

No. The requirement for visible disclosure under Article 50(4) primarily applies to AI-generated or altered image, audio or video content that meets the criteria for deepfake. The European Commission explains it here.

Is human grammar correction considered editorial control?

No. European Commission points out that superficial or formal proofreading, such as simply correcting spelling or grammar, is not the same as real content review and editorial control.

Is it mandatory to use the EU AI icon?

No. Official EU AI content icons are voluntary. If a labelling obligation applies to specific content, the obligation itself remains, but it does not have to be implemented specifically with the EU pictogram.

Finally

The new EU rules do not imply a principle:

„"You used AI → mark everything."“

It would be much more accurate to follow another question:

„"In this case, can AI-generated content mislead a human or inform the public about a sensitive issue of public interest without actual human control?"“

This allows businesses to use AI to create text, visuals, and other content without adding unnecessary labels to each file, but at the same time clearly identifies situations where transparency is already a legal obligation.

In practice, the best rule is simple: AI can help with creation, but a person or organization must be clearly responsible for the accuracy, context, and final decision of the content published publicly.

The article is based on Section 50 of the AI Act, The European Commission's 2026 guidelines on AI Act transparency requirements and Code of Practice on Transparency of AI-generated Content. This is general information and not individual legal advice.

Want to learn more about using AI in your business?

At AIhelp you will find practical guides that will help you independently apply AI to daily business operations, content creation, website visibility and other processes. They provide clear methodologies, ready-made AI queries, and specific actions that can be applied without complex technical knowledge.