Do you need to label AI-generated content? This rule has been in effect since August 2, 2026.

You're having ChatGPT help you write a LinkedIn post. Do you have to include a disclaimer stating that AI was used?

No.

New transparency rules under Article 50 of the EU AI Act have been in effect since August 2, 2026. But that doesn’t mean you’ll have to label every text, image, video, or social media post with a big  “Made with AI” label from now on.

The rules are much more specific. They primarily apply to chatbots, deepfakes, and certain texts on topics of public interest. And when it comes to texts, human editorial oversight plays a particularly important role.

It’s time to make the AI jungle a little more manageable again.

TL;DR

• As of August 2, 2026, the transparency rules under Article 50 of the EU AI Act have taken effect.
• Not all AI-generated content needs to be labeled.
• Deepfakes and some AI-generated texts on topics of public interest may, however, be subject to the transparency rules.
• For AI-generated text, human review or editorial oversight can be an exception, but only if someone also bears editorial responsibility.
• With chatbots, it must be clear that the user is communicating with AI.
• A technical AI marker is not the same as a label visible to the user.

When should you label AI-generated content?

For organizations that use AI for marketing, communication, and content, it basically comes down to this:

Bottom line: using AI doesn’t automatically mean you have to label your content. It all depends on how you use it, what the content is, and how you handle the process around it.

When is an AI-generated image or video a deepfake?

When you hear “deepfake,” you might immediately think of a fake video of a politician saying something he never actually said.

And rightly so. But the definition is broader.

The AI Act defines a deepfake as AI-generated or manipulated image, audio, or video content that closely resembles a person, object, place, organization, or event that exists, could plausibly exist, or could have existed. In addition, the content must be capable of misleading someone into believing that what they see or hear is authentic or true.

The European Commission considers three factors in this regard:

1. How closely does the AI content resemble the subject it imitates?
2. Does it concern something that exists, could plausibly exist, or could have existed?
3. Could someone mistakenly believe that the content is authentic or truthful?

Context also plays a role. A dragon flying over Amsterdam is unlikely to be seen as real news, but a realistic photo of a major fire on Dam Square might be.

You must make a deepfake clearly recognizable

Does an image, audio, or video fall under the definition of a deepfake? If so, it must be clear upon first exposure that the content is artificially generated or manipulated.

Simply embedding metadata somewhere in the file is not sufficient for this purpose. The recipient must be able to see or hear the indication without special software or additional steps.

For example:

• This image was generated using AI.
• This video contains AI-manipulated images.
• The voice in this video was generated using AI.

The EU doesn’t prescribe the exact wording, as long as the notice is clear, recognizable, and accessible.

What about creative, fictional, and satirical content?

There’s an important nuance here.

A clearly fictional image isn’t automatically a deepfake. If no one would reasonably expect the image or video to be authentic, it may fall outside the definition of a deepfake.

But does creative, artistic, satirical, or fictional content still qualify as a deepfake? In that case, the transparency requirement doesn’t suddenly disappear just because you label it “satire.”

You still need to indicate that AI was used to generate or manipulate the content in question, and that disclosure should only be presented in a way that does not unnecessarily disrupt the experience of the creative work.

Does text written with ChatGPT, Claude, or Sterc.ONE need to be labeled?

Here, too, the answer is: not automatically.

To be subject to the visible labeling requirement under Article 50(4), an AI-generated or manipulated text must meet three conditions. The text must be:

1. published;
2. intended to inform the public;
3. focused on a topic of public interest.

Examples of topics of public interest include politics, democratic processes, public services, the administration of justice, fundamental rights, public safety, public health, the environment, consumer safety, and economic, financial, scientific, or cultural developments that may be relevant to public debate.

An automatically generated news article about new legislation may therefore fall under this category.

A product description for a new office chair usually does not.

But note: the platform does not determine this.

A commercial LinkedIn post is not automatically excluded. If you use LinkedIn to inform people about, for example, new legislation, public health, or an important social development, then that same post may indeed address a topic of general interest.

Human editorial review makes a big difference.

Now it gets interesting.

An AI-generated text on a topic of public interest doesn't need a visible label if it has undergone human review or editorial oversight and a person or organization takes editorial responsibility for the publication.

These are three distinct things:

Human review
In this process, one or more people consciously assess the content of the text. They must have sufficient relevant knowledge of the subject and be able to professionally judge whether the content is accurate and can be published responsibly.

Editorial review
In this process, a responsible editor or editorial body has actual control over the content. That person must be able to approve, modify, or reject the text based on its content. Examples include fact-checking information and verifying the reliability of sources.

Editorial responsibility
It must also be clear who ultimately bears legal responsibility for the publication. This can be an individual or an organization.

And no, simply clicking “check spelling” does not count.

A superficial language check, grammar check, or other purely formal review is explicitly not considered by the European Commission to be human review or editorial oversight.

So AI is perfectly fine for doing the heavy lifting, but “ChatGPT said it, so let’s publish it” is not an editorial process.

And yes, that applies to this article as well

An article about the EU AI Act may be intended to inform the public about a topic of general interest. If you use AI to create such an article, the way you verify the content is therefore important here as well.

For this article, for example, that means we don’t just check whether it reads well. The content must be carefully evaluated against official sources, and any incorrect or insufficiently substantiated claims must be corrected. In addition, there must be someone within Sterc who can actually approve, modify, or reject the content, and it must be clear who bears ultimate responsibility for the publication.

Does the process meet the conditions for human review or editorial oversight and editorial responsibility? If so, the exception in Article 50(4) applies, and under that provision, the text does not need to be labeled as AI-generated.

Quite appropriate, isn’t it?

What do you need to disclose when using a chatbot?

Article 50 also contains rules for AI systems that communicate directly with people. Think of chatbots, AI assistants, AI agents, and interactive avatars.

People must know from the very beginning of the first interaction that they are communicating with AI, unless this is already abundantly clear to a reasonably informed and attentive user. The European Commission notes that this exception should be applied with caution.

So don’t make things difficult for yourself:

You’re chatting with our AI assistant.

Done.

And definitely don’t hide that notice somewhere on page 37 of your terms and conditions.

Who’s responsible for this?

This is where things get a bit more technical from a legal standpoint.

The obligation to design an interactive AI system in such a way that users are informed about the AI interaction lies with the provider of the AI system.

Does your organization use an external chatbot solution? If so, make sure the supplier has handled this properly.

Are you developing an AI system yourself, or having it developed for you, and then marketing or putting it into use under your own name or trademark? In that case, under the AI Act, you are considered the provider of the system in many instances.

A visible AI label is different from a technical AI marker

Article 50 includes a second form of transparency that is easily confused with visible labels.

Providers of generative AI systems must design their systems so that synthetic text, images, audio, and video can, in principle, be marked in a machine-readable format and be detectable as AI content.

Think of metadata, watermarks, or other technical methods.

There are specific exceptions to this rule, such as for certain forms of standard editing and some technical or industrial applications.

But what’s most important for you is this:

In the case of a deepfake, a technical marker does not automatically replace the visible or audible notification to the user.

Metadata that no one sees is useful for the machine, but people need to understand it too.

Do you have to use the official EU AI icons?

No.

The European Union has now developed official icons that you can use to indicate that content has been fully generated by AI or partially modified by AI.

Their use is voluntary.

Does your content need to be labeled under Article 50? In that case, of course, the requirement is not voluntary. However, you may also communicate in another clear way that AI was used.

In fact, simply using an EU icon does not automatically mean you’re complying with all requirements.

A clear text label can therefore work just fine.

What about content created before August 2, 2026?

Good news for anyone who was about to panic and start going through their entire content archive.

Content generated before August 2, 2026, that was already available does not need to be labeled retroactively. The European Commission does, however, encourage voluntary labeling where possible.

In addition, a limited transition period applies to one specific technical requirement: Generative AI systems placed on the market before August 2, 2026, have until December 2, 2026, to comply with the technical labeling and detection requirement set forth in Article 50(2).

So this is not a general postponement of the transparency rules.

Common scenarios

• ChatGPT writes product descriptions for your online store → normally, no visible AI label is required under Article 50(4).
• ChatGPT writes a blog post about a new tax rule and no one checks the content → a label is required if the text falls under the conditions for information of general interest.
• That same blog post is fact-checked, and the organization bears editorial responsibility, so the exception may apply.
• You generate a clearly fictional illustration of a purple elephant in an office, generally not a deepfake if it is not reasonably perceived as authentic.
• You generate a photorealistic vacation photo of an existing destination and add a nonexistent swimming pool or palm trees, so assess it as a potential deepfake, precisely because a visitor might perceive it as a genuine representation of that location.
• You use Canva solely to crop a real photo, perform color correction, or resize it → this does not suddenly require labeling just because Canva has AI features.
• You use Generative Fill to add a significant, realistic element to a photo, reassess based on the deepfake criteria.
• A website chatbot answers customer questions, so the user must be informed that they are communicating with AI, unless this is already obvious.

And if you do nothing?

The transparency requirements are not just non-binding suggestions.

Violations under the AI Act can result in administrative fines of up to 15 million euros or 3% of the company’s total global annual revenue from the previous fiscal year. Proportionality rules apply to smaller companies and SMEs.

That’s no reason to start slapping “AI” everywhere in huge letters just yet.

It is, however, a good reason to make sure your processes are in order.

Checklist: do I need to label this AI content?

Go through these questions before publishing AI content to determine whether you can publish it with or without a label:

• Does an AI system communicate directly with customers, visitors, or employees?
• If so: is it clear from the start that they are communicating with AI?
• Are you publishing an AI-generated or manipulated image, video, or audio clip?
• Could someone reasonably mistake that content for authentic or true?
• Does this mean the content could potentially qualify as a deepfake?
• Is there a clear, visible, or audible notice accompanying a deepfake?
• Are you publishing AI-generated or manipulated text to inform the public about a topic of public interest?
• Has someone with relevant knowledge and professional judgment deliberately reviewed the content, or has a responsible editor been able to approve, edit, or reject it?
• Have the facts and the reliability of sources been verified where necessary?
• Is it clear which person or organization is ultimately responsible for the publication?
• Have you established internal guidelines for when and how AI-generated content is labeled?

That last point may be the most important of all.

Because if every team member has to figure out for themselves what the EU AI Act says about that one Instagram post today, you don’t have a process. You have a weekly knowledge puzzle.

Still unsure? Use the decision guide

The rules depend on the type of content, how realistic the output is, and what you’re using it for. Answer a few questions and get an initial assessment for your situation.

Labeling AI content is part of your AI policy

Ultimately, labeling is just one aspect of responsible AI use.

You’ll also want to establish guidelines within your organization regarding:

• which AI tools employees are allowed to use;
• what data they are permitted to process using those tools;
• who evaluates new applications;
• when human oversight is required;
• who is responsible for AI output;
• how to record AI usage;
• when transparency and labeling are required.

That's precisely why an AI policy isn't a legal document that should be left to gather dust somewhere at the bottom of SharePoint. It's the set of guidelines for how your organization works with AI.

Be sure to read our article on why your organization needs an AI policy.

Within our Sterc.ONE AI roadmap, defining these frameworks is therefore one of the first steps. First, clarity on policy, roles, and responsibilities. Only then should AI be scaled up further into processes, assistants, and agents.

Don’t just slap a label on everything, but do think things through thoroughly.

The most important conclusion?

As of August 2, 2026, you must make certain AI content and AI interactions transparent. Not everything that has ever been touched by ChatGPT, Gemini, Claude, or an image generator automatically gets a label.

For organizations, three situations are particularly important:

• people need to know when they’re communicating directly with an AI system;
• deepfakes must be clearly identifiable as artificially generated or manipulated;
• AI-generated texts on topics of public interest must be labeled, unless they have undergone human review or editorial oversight and someone bears editorial responsibility for the publication.

And with that, one thing becomes more important than ever: your AI process.

Because you can try to check off each new rule individually. Or you can make sure that AI use, human oversight, accountability, and transparency become a structural part of how your organization operates.

We’re fans of option two. It saves a lot of labeling work.

How’s your AI policy shaping up?

Want to know which AI regulations are relevant to your organization and how to practically incorporate them into policies, processes, and everyday AI tools?

With our AI approach, we don’t just identify the risks, we focus on what you need to organize differently starting tomorrow.

It’stime to not only use AI but also manage it effectively.

Sources

This article draws on sources including the EU AI Act, the European Commission’s final guidelines of July 20, 2026, the official FAQ on Article 50, the Code of Practice on Transparency of AI-Generated Content, the EU guidelines for AI labels, and the Digital Omnibus on AI.

This article provides a practical explanation of Article 50 of the EU AI Act and does not constitute individual legal advice.