Starting August 2, 2026, AI labeling requirements will become mandatory under Article 50 of the EU AI Act. To help businesses prepare, the EU Commission set up two working groups to develop a "Code of Practice on Transparency of AI-Generated Content."
For the first time, this code provides practical, real-world guidance on how to label AI-generated content in a legally compliant way. In this post, we will explain what these recommendations are, who they apply to, and what your marketing team needs to do right now.
Key takeaways
- The deadline is firm: Starting August 2, 2026, the transparency requirements under Article 50 of the EU AI Act will apply. This is true whether you sign the voluntary Code of Practice or not.
- Marketing teams are "deployers": As a marketing or communications team, you are generally classified as a deployer. This means you are directly responsible for labeling your content.
- What needs labeling: Requirements primarily target deepfakes and published text on matters of public interest that bypass editorial oversight.
- The visual test matters: Compliance does not depend on the specific tool you use. It depends on whether the content looks photo-realistic and plausible to the average viewer.
- Official assets are coming: The EU is providing three official icons. When in doubt, the safest approach is to label your content.
On June 10, 2026, the EU published a voluntary code of conduct for Article 50 of the EU AI Act. This Code of Practice translates the law's abstract transparency obligations into actionable guidelines. It details how to mark, detect, and label AI-generated content.
If you work with synthetic images, video, audio, or text, this document provides your first concrete roadmap for compliance.
A crucial distinction: The Code of Practice itself is voluntary. However, the underlying legal obligation in Article 50 is not. It takes effect on August 2, 2026. In Germany, the Federal Network Agency (Bundesnetzagentur) will serve as the market surveillance authority.
This also impacts businesses outside the EU. If you target consumers or run AI systems (like chatbots) within the EU, you must comply. This is due to the market-of-destination principle.
Want to build AI literacy in your team? Our AI literacy training prepares your marketing and communication teams for the realities of Article 50.
The Code of Practice is a voluntary compliance framework. Following it does not grant automatic immunity. However, it gives you a structured, standardized way to demonstrate compliance with Article 50(2), (4), and (5) of the AI Act to regulators.
The code is technology-neutral and uses three levels of commitment:
"Will" (mandatory for compliance)
"Encouraged" (recommended best practice)
"May" (optional)
If you sign the code, you must implement everything marked as "will."
A smart procurement tip:
When purchasing marketing software, ask vendors if they have signed the Code of Practice. Actively request proof of their AI-detection and metadata-marking capabilities.
If your software does not embed these invisible marks "under the hood," your team cannot easily meet its own labeling duties. Documenting these tools early also prepares you for stricter reporting rules if your use cases ever shift into "high-risk" categories.
The Code of Practice is split into two sections, targeting different players in the AI ecosystem:
Section |
Target audience |
Core obligation |
|---|---|---|
| Section 1 | Providers of generative AI systems (Art. 50 (2), (5)) | Must embed machine-readable metadata and watermarks into AI outputs. |
| Section 2 | Providers of generative AI systems (Art. 50 (2), (5)) | Must visibly label deepfakes and published text for human audiences. |
These two layers work together to ensure transparency for consumers.
The first layer focuses on technical watermarking. This is the software provider's job.
The second layer focuses on the end-user. This is where marketing and communication teams operate. If you use an AI tool to create content, you are a deployer. You are legally responsible for visible labeling under Article 50. You cannot delegate this duty, even if platforms like Instagram or LinkedIn automatically detect and label your content for you.
If you use AI to create deepfakes (realistic images, audio, or video of people, places, or events) or to publish unedited articles on public interest topics, you must disclose it.
The EU provides three official icon variations. User testing shows that labels with clear text (like "GENERATED" or "MODIFIED") perform much better than icons alone. If you design your own labels, they must still meet the EU's layout and placement rules.
Icon type |
When to use |
Example |
|---|---|---|
|
Standard Icon
|
Use when AI helped create a deepfake or public-interest text, especially alongside custom text labels. | A deepfake video accompanied by the text "Voice generated with..." followed by the icon. |
|
Fully AI-Generated
|
Use when a deepfake or text was created entirely by AI without human editing (beyond initial prompting). | Fully AI-generated deepfake videos, AI-composed music, or automated news summaries. |
|
Partially AI-generated
|
Use when AI was used to modify real, human-made content to create a deepfake or alter public-interest text. | Use when AI was used to modify real, human-made content to create a deepfake or alter public-interest text. |
Labels must be visible instantly upon first contact. Users should not have to search for them, and the text must remain readable against any background.
For artistic, satirical, or fictional works, you only need to disclose AI use in a way that does not ruin the creative integrity of the work (for example, in the credits or description). If a text goes through a formal human editorial process with clear editorial accountability, the labeling rule for published text does not apply. You must document this process carefully.
The Code of Practice expects you to build internal processes. This includes documentation, team training, and a review system to catch missing or incorrect labels. Your compliance program should be scaled to the size of your business.
Format |
Recommended labeling approach |
|---|---|
| Image/video | Embed the EU icon visibly (e.g., top right). For video, show it at the start, at set intervals, and after any interruptions (like ad breaks). |
| Audio only | Use a short, clear spoken disclaimer at the beginning. Repeat at intervals for long-form or live audio. You can also use short audio cues ("ear cons"). |
| Audio with a screen | Combine the spoken disclaimer with a visual icon on the screen (such as a car or phone display). |
| Published text | Place the icon next to the headline or in the author bio/colophon. For very short texts, a contextual note in the user interface works. |
| Artistic/ creative work | Place the label discreetly to protect the art (e.g., in the description, caption, end credits, or at the point of sale). |
Note on accessibility: The code requires all labels to be fully accessible (supporting high contrast, screen readers, and tactile alternatives) in line with the European Accessibility Act.y Directive.
Neither the EU AI Act nor the Code of Practice provides a definitive list of exactly what must be labeled. However, the accompanying commission guidelines from May 2026 offer useful guidance. The decisive test for images, audio, and video is the definition of “deepfake” (Art. 3 (60)), which includes two criteria:
Could this content actually exist in the real world? A photo-realistic portrait of a fictional person is a deepfake. An elephant driving a sports car or a sphinx flying over the Eiffel Tower is not. These scenarios obviously go against the laws of physics and so you don't need to label them.
Did the AI alter the actual meaning and substance of the media, or just the technical quality? Standard adjustments like color grading, noise reduction, lighting fixes, or cropping do not turn an image into a deepfake.
The Commission's interpretation serves as a rule of thumb:
We have already addressed some of these questions together with our customers. This should help clarify many of your questions about gray areas:
Your question |
Our initial assessment |
|---|---|
| Do we need to label AI content on our internal intranet? | Strictly internal content with no public reach is generally exempt. However, if the content impacts employee decisions (like performance reviews), or if your intranet is accessed by a massive audience, rules can apply. Best practice: Label it anyway. |
| Do we have to label Photoshop edits now? | It is not about the tool; it is about the edit. Standard AI-assisted retouching (lighting, cropping) is fine. If you make a major, photo-realistic change to the actual content using AI, you need to label it, even if you did it in Photoshop. |
| We place real products in front of AI-generated backgrounds. Do we need to label this? | It depends on whether the final image looks realistic and plausible. If a real product sits in front of a photo-realistic AI background, a consumer might assume the entire photo is real. This counts as a deepfake. Highly stylized or obviously fake backgrounds (like those in online product configurators) are much safer. |
| Photoshop background vs. Midjourney background? | The tool does not matter; the output does. A manual composite built in Photoshop that looks obviously artificial does not fall under Article 50. A photo-realistic background generated in Midjourney makes the final image look like a real photo. |
| Do we need to label every organic social media post made with AI? | No. The trigger is not whether the post is an ad, but whether it is a deepfake. If it is photo-realistic and plausible, label it. If it is clearly illustrative or stylized, you generally do not need to. Your intent to deceive (or lack thereof) does not change this rule. |
Even if you do not sign the voluntary Code of Practice, it is currently the clearest guide we have for compliant AI labeling. For marketing teams, the message is clear: audit your content pipeline, identify where you act as a "deployer," and set up your labeling workflows now.
By defining standard use cases and setting internal rules early, you can avoid bottleneck debates down the road. Structured AI governance protects your brand from regulatory risk and signals transparency and trustworthiness to your customers.
We expect many gray areas to persist until courts begin interpreting the AI Act. Until then, use this simple rule of thumb: If you are debating whether to label an asset, label it.
How FELD M can help you
We help marketing and communication teams align their content workflows with Article 50. From tool audits and AI governance setups to hands-on team training, we make compliance practical.
Note: This article provides professional insights and is not a substitute for legal counsel. Compliance strategies should be verified against the final text of the EU AI Act and official European Commission guidelines.