Guides
8 min read

AI Labeling Requirements: What Businesses Must Label in Images, Text and Chatbots

Four colorful structures made of toy building blocks with pink flags, next to a magnifying glass made of blocks
Image: AI-generated
In short

Since August 2, 2026, you have to label deepfakes, unreviewed AI text on public interest topics and chatbots. Clearly illustrative images, standard retouching and texts a human reviews and takes responsibility for need no label. The extension to December 2, 2026 only applies to the companies that build AI tools, not to you.

What do you have to label since August 2, 2026?

You have to label three things: deepfakes, AI-generated text on matters of public interest that nobody takes editorial responsibility for, and chatbots where it isn't clear that an AI is answering. These rules come from Article 50 of the EU AI Act, which according to the European Commission has applied since August 2, 2026.

The key is who plays which role:

  • Providers are the companies that build AI systems, such as OpenAI, Google or Midjourney. They have to mark their output in a machine-readable way, meaning invisible watermarks or metadata that software can use to detect AI content.
  • Deployers are businesses that use these systems professionally. That's you if your marketing team works with ChatGPT, Gemini or an image generator. You have to visibly disclose when you publish a deepfake or an unreviewed AI text on a public interest topic.

The Digital Omnibus, the package the EU used this summer to amend the AI Act, changed little here. Regulation (EU) 2026/1744 of July 8, 2026 mainly postpones the obligations for high-risk systems. For Article 50, it only gives providers until December 2, 2026 for machine-readable marking, and only if their system was already on the market before August 2. As a deployer, you get no extension.

Which content is actually covered?

What matters is not whether AI was involved, but whether content looks deceptively real or informs the public without review. Under the regulation's definition, a deepfake is an AI-generated image, video or audio that resembles existing persons, objects, places or events and would falsely appear to be authentic.

Content Do you have to label it? What to do
Photorealistic AI image of your office, product or team Yes, it looks like a real photo Visible "AI" label in the image
Cloned voice in an ad or podcast Yes, audio deepfake Spoken disclaimer at the start
Video with an AI avatar of a real person Yes Label at the start of the video and after interruptions
Clearly illustrative AI image (3D graphic, comic) Usually not, it doesn't look real A voluntary note doesn't hurt
Retouching, cutouts, color correction with AI Usually not, standard editing Nothing, as long as nothing new is created
Deepfake in satire or an obviously fictional campaign Yes, but lighter Note in the description or credits, for example
AI-written expert article, reviewed and owned by a person No Editorial policy (see below)
AI text on business or politics, published unreviewed Yes Label at the headline
Internal training videos with deepfakes Yes, but simplified Note upfront, for example in the learning platform

What "public interest" means for text is described in the Commission's FAQ, which lists politics, public health and environmental protection along with economic, financial, scientific or cultural developments. Plain product copy isn't explicitly mentioned. Our take: guides and industry articles on your website can fall under it. Rather than assessing every text individually, an editorial sign-off is the simpler route.

No retroactive duty: according to the Commission, you don't have to label content created before August 2, 2026. It does recommend doing so where possible.

What does a correct label look like?

The label is a clearly visible "AI", ideally the official EU icon. The law doesn't prescribe a format, but the final Code of Practice, published by the EU on June 10, 2026, is becoming the benchmark. The Commission and the AI Board have confirmed it as an adequate voluntary tool, and by the end of July about 190 organizations had signed it. If you don't use it, you have to show compliance through other, equivalently adequate means, according to the Commission.

The Code's requirements in short:

  • Icon: a capitalized "AI" as the main element. The EU offers three free icons: "AI + GENERATED" for fully generated content, "AI + MODIFIED" for partly altered content, and a basic icon. In user tests, the versions with added text performed significantly better on noticeability and clarity.
  • Images and video: the label sits directly in the content, in a spot without overlays, such as the top right corner. In videos it appears at the start, at regular intervals where possible and at least after every interruption, so it stays visible in clips and screenshots.
  • Audio: a short spoken disclaimer right at the beginning. If there's a screen, add the icon.
  • Text: label above the text, near the headline or in the colophon at the top. If only part of the text is AI-generated, you can label just that part.
  • Websites and apps: for art and satire, the label can sit next to the image or video, for example as a note that explains more on mouse hover. It has to be visible without an extra click.
  • Accessibility: high contrast and readable by screen readers. Article 50(5) already requires this, and it fits the requirements we covered in our web design trends for 2026. So the note belongs in the alt text too.

The label has to be there at the latest on first contact with the content, and noticeable without any action from the viewer. A line in the footer or your terms and conditions doesn't cut it.

How can you skip labeling AI-written text?

With a short editorial policy: if a human reviews the text before publication and a person or your company holds editorial responsibility, the text label is not required. Commitment 4 of the Code names two minimum elements:

  1. Who is responsible? Name, role and contact details of the responsible person or body. The contact details should be public. On most German websites they already appear in the Impressum (the legal notice German law requires on business sites).
  2. How is content reviewed? An overview of the people and processes that ensure review before publication. For example: the managing director reviews expert articles, the head of marketing reviews social posts, nothing goes live automatically.

The Code explicitly does not require logs of every single review. A page in your internal wiki is enough, as long as your team actually follows it.

Be careful with automations: a workflow that publishes AI text straight to your website, blog or social media has no human review. If you build something like that with Make or n8n, add an approval step or put a label on every post. One easy option: the workflow creates a draft that only goes live after a click.

What applies to chatbots on your website?

A chatbot has to state clearly that it's an AI, right at the start of the conversation. The Commission reads the "it's obvious" exception narrowly. Our take: a bot called "Lisa" with a photo is not obviously a machine.

Formally, this duty falls on the provider, who has to design the system that way. If you use an off-the-shelf chat widget, check whether it shows the notice. If you build the bot yourself and run it under your own name, then by our reading of the definition you may count as a provider yourself. Either way, the effort is small: a fixed line in the greeting such as "Hi, I'm the AI assistant of Muster GmbH".

What happens if you don't label?

Fines can reach up to 15 million euros or 3 percent of worldwide annual turnover, and according to the Commission, proportionality should be taken into account for small and medium-sized businesses. In Germany, the Bundesnetzagentur (the Federal Network Agency) has been the market surveillance authority and complaints point for the AI Act since July 29, 2026. It also accepts complaints about violations. Within its remit it monitors the transparency obligations as well, while the Landesmedienanstalten (the state media authorities) remain responsible for media.

The fine isn't the only risk. The Code makes clear that a label doesn't exempt you from other laws, such as getting consent from the people depicted or from rights holders. A labeled deepfake of a real person made without their consent is still a problem. For realistic intimate depictions of identifiable people without explicit consent, the AI Act even adds a separate ban from December 2, 2026.

What does this mean for your business?

For most mid-sized companies, the effort is manageable. This checklist is enough to get started:

  1. Take stock: where do you use AI images, videos, voices, texts and chatbots? Website, social media, newsletters, presentations, training.
  2. Label anything photorealistic: everything that looks like a real photo, video or voice gets the EU icon, placed according to the rules above.
  3. Write an editorial policy: responsible person, review process, one page. That takes reviewed texts out of scope.
  4. Check your automations: every workflow that publishes without approval gets an approval step or an automatic label.
  5. Update your chatbot greeting: one sentence that discloses the AI.
  6. Brief your team: the Code asks you to make your team aware of the obligations in proportion to your company size. Ten minutes in a team meeting with the table above is a good start.
  7. Ask your vendors: agencies and freelancers who create content for you should know the same rules.

Our take: the main effort isn't the label, it's knowing everywhere AI is involved. If a human already reviews your content before it goes out, nothing changes for your texts. An open "AI" label on images can even build trust, because nobody feels deceived. How AI content performs in search engines is covered in our article on AI SEO. And if you want to build approvals into your content workflows, talk to us.

This article is not legal advice. For borderline cases, such as campaigns featuring real people, it's worth having a lawyer take a look.

Frequently asked questions

Do I have to label AI images on social media differently than on my website?

The obligation is the same, only the execution differs. For deepfakes, the label belongs directly in the image or video, because posts get shared and passed around as screenshots. You can use the platforms' own AI labels on top.

Is a general note in our legal notice that we use AI enough?

No. The label has to be noticeable at the latest on first contact with the specific content. Nobody sees a blanket note in the Impressum or the terms and conditions at the moment they look at the image or video.

Can I use the EU icon freely?

Yes. The EU provides the icons in the annex of the Code of Practice for free use, without having to credit the Commission or the AI Office. You can adjust size, color and contrast as long as the label stays clearly recognizable.

Do I have to label texts that were translated or rewritten with AI?

Only if the text covers a matter of public interest and nobody reviews it and takes responsibility before publication. With an editorial policy your team actually follows, translated texts fall under the exception too.

Does the duty also apply to AI images in internal presentations?

For deepfakes, yes, internally as well. In a closed professional setting, though, the Code allows a simplified note upfront, for example in the training platform or at the start of the presentation.

Sources

  1. EUR-Lex: Regulation (EU) 2024/1689 (AI Act), Article 50
  2. European Commission: Final Code of Practice on Transparency of AI-generated Content
  3. European Commission: FAQ on transparency obligations under Article 50
  4. EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI)
  5. Bundesnetzagentur: Key role in implementation of AI Act
This article was created with AI.
Got a project in mind?

Tell us what you're planning. You'll get an honest assessment within 24 hours.

Got it! We will get back to you asap!
Whoopsie! Something went totally wrong here... Maybe just send an email to hi@blocksandcolors.de? Sorry!
Or email me directly: hi@blocksandcolors.de