The EU Commission's Draft Code of Practice on AI Content Transparency: What It Contains and What It Means for Organizations

Practical analysis of the draft code of practice for transparency around AI-generated content

Why This Code Exists: Making Article 50 AI Act Concrete

Article 50 of the AI Act requires, in specific situations, that it be made clear that content was created or manipulated by AI. In broad terms, this concerns two worlds that intersect:

Providers of AI systems must ensure that AI content, where technically feasible, can be marked and is detectable. Think of including provenance information or building in signals that can later be used to determine that an image, audio, or video was generated or modified by an AI system.

Deployers of AI content must, in certain cases, visibly label content for the public. Think of deepfakes, manipulated images, or AI-generated text shared in a context of public interest.

This Is a Draft, But It Is Directional

Because it is a draft, you cannot read it as a definitive set of requirements. You can read it as a clear signal about the direction Europe is choosing:

• Technology and Organization: Transparency is elaborated as a combination of marking/detection and labeling/governance/accountability
• Interoperability: Not every company with its own label and definitions, but shared agreements as much as possible
• Realistic Limitations: Not every modality can be marked equally "hard," and markings can often be removed—this is translated into a layered approach

Core 1: Providers Must Move Toward "Layered Marking" and Provable Detection

For providers, the central idea is that one technique is rarely enough. The draft code therefore steers toward a multi-layered approach:

Detectability as a Service

Notably, the code doesn't stop at "mark it." It also steers toward detectability as a service. Providers are pushed toward a free or publicly accessible way to verify content, for example via a web interface or API with confidence scores.

Practically, this means providers must not only build a technical solution but also think about:

• How do you scale verification without it becoming a cost or security problem?
• How do you handle false positives and false negatives?
• How do you provide transparency without immediately revealing model secrets or abuse information?

Core 2: Deployers Get a Labeling Obligation That Goes Beyond a Text Line

For deployers, the code strongly emphasizes a recognizable and consistent labeling approach. This is not just about "put 'made with AI' somewhere." The draft code works toward a shared vocabulary, a recognizable icon, and agreements about where and when a label should be visible.

Taxonomy for AI Content

This distinction seems simple, but in practice, this is exactly where discussions arise. A marketing department that has a photo "just slightly" adjusted by generative AI often feels this is minor. From a transparency and trust perspective, the public may experience it differently.

Modality-Specific Implementation

The code also works toward a (temporary) label icon and later a broader EU icon that can also be interactive. Audio requires different measures than images. Think of a podcast fragment or voice-over: a label in a description is often insufficient when people are only listening. Therefore, disclosure that also returns "in" the experience is discussed, for example repeated during longer audio.

Core 3: There Will Be Exceptions, But They Require Discipline

Transparency in the AI Act has exceptions and nuances. The draft code elaborates on these further. Two stand out:

What Does This Mean for Your Organization If You're Not an AI Provider?

Many organizations are not providers of AI systems but are intensive users. Think of municipalities publishing images, educational institutions with communications departments, HR teams creating recruitment materials, or legal departments generating summaries.

For these organizations, the core message is: transparency becomes a workflow requirement.

You will need to know in your processes:

• Where AI is used in the chain (text, image, audio, video, translation, editing)
• Which output goes external and which stays internal
• In which cases you need to label, and how you handle exceptions
• How you stay consistent across channels: website, social media, newsletters, presentations

Practical Example

An organization has a video made for a campaign. The images are partly real, partly AI-generated, and the voice-over is synthetic. Without agreements, fragmentation occurs: one channel labels, another doesn't. With a consistent taxonomy, a standard label, and a simple content checklist, it becomes manageable.

What Does This Mean for Providers and Product Teams?

For providers and product teams, the draft code has a second effect: it transforms transparency into a product feature that customers will start asking about.

When procurement asks: "Can we detect, demonstrate, and label AI output?", you need more than a policy document. You need technical building blocks that fit into the ecosystem:

• Provenance metadata
• Watermarks
• Verification interfaces
• Logging

Product Choices

For product teams, this also means you must make choices about:

• Default settings: Marking on by default, or opt-in?
• User experience: How do you provide transparency without paralyzing the workflow?
• Integrations: How do you connect to CMS systems, DAM systems, social publishing tools, and archives?

What You Can Do Now, Without Waiting for the Final Version

You don't need to make everything "AI Act-proof" today. You can already take actions that will definitely hold value later.

📚 Deepen Your Knowledge: Check out the Complete EU AI Act Guide for a full overview of all aspects of the AI legislation.