The European Commission is soon expected to circulate the first draft of the Code of Practice on marking and labeling AI-generated content . Drafted by a group of experts, the Code will help companies demonstrate compliance with their obligations under Article 50(2) and Article 50(4) of the EU AI Act, which cover the marking and detection of AI-generated content, as well as the labelling of deepfakes and certain AI-generated publications.
The Code seeks to address important issues that can help with increasing trust and transparency around generative AI. At the same time, it needs to account for the nascent nature of marking and detection techniques, the diverse transparency needs across AI use cases, and the lack of standards.
As a member of the stakeholder group supporting the drafting of the Code, 91快活林 has been making the case for a balanced and future-proof Code. Our priority? The Code must leave flexibility to find the right marking and labelling approaches that work for different products, audiences and types of content. Only by taking this approach can the Code be workable and effective.
Here is what industry seeks from policymakers as they finalize the first draft of the Code:
1. Avoid prescribing specific marking and detection techniques for article 50(2)
Companies are investing in the development of different tools to mark AI-generated content: from applying invisible watermarks to images and videos, to developing robust content provenance and metadata standards such as . Despite significant innovation in this space, no single solution is suitable for all types of generative AI systems. This is why AI providers should have the flexibility to choose the most appropriate approach to comply with Article 50(2)– depending on the use case, the type of content (e.g., text vs audio or images), and the level of risk involved. In many cases, a combination of approaches may be the most effective solution. If the Code takes a prescriptive approach – for example by imposing one specific technology – it would become quickly outdated, as marking techniques will continue to evolve alongside the technology.
2. Apply a flexible approach to help acknowledge limitations of today’s tools
Despite significant efforts, today’s marking techniques are far from perfect. Specifically, some methods may be technically possible but still be impractical – too expensive, too slow, or too damaging to the user experience. Additionally, interoperability may not be yet possible for certain state-of-the-art, novel marking techniques. Pushing companies to adopt techniques that are commercially or operationally unworkable would chill research and development and slow progress on better solutions. A key challenge is also the marking of AI-generated text. Since text can be significantly edited by users, marking it is much more difficult than, for example, embedding a watermark in a picture.
While article 50(2) of the AI Act requires marking techniques be “effective, interoperable, robust and reliable,” the Code must recognize the limitations of the current state of the art and adequately balance between different trade-offs.
3. Make marking requirements risk-based
Generative AI is used in many different ways – from online services to creative or professional tools - and transparency needs vary accordingly. A realistic synthetic video raises different concerns than a small AI-assisted edit in a photo app - which should not require specific marking or labelling. Similarly, in the case of AI-generated Code or real-time audio, marking would significantly degrade performance, while the artificial nature of the content may be obvious.
To reflect these nuances, the Code should remain proportionate to the level of risk posed by different use cases. Small AI outputs or situations where it’s obvious that a piece of content is AI-generated should not be covered by the Code.
4. Ensure that labelling is meaningful, not overwhelming
Transparency only builds trust when it is meaningful. Users need labels they can notice and understand, and different audiences have different levels of AI literacy. For example, AI tools used by the general public would have different transparency needs than specialized tools for professionals.
In addition – labelling must be based on risk. If every minor or routine use of AI is labelled, users can quickly become overwhelmed, leading to “label fatigue” and even the false assumption that unlabeled content is more trustworthy.
To avoid this, the Code should take a targeted approach that focuses on situations where AI involvement genuinely matters or could mislead users. Realistic AI-generated content can be used in a variety of beneficial and low-risk cases, such as for advertisement, entertainment or internal communications – and these situations should not be covered by the AI Act’s labelling requirements.
Looking ahead
The upcoming Code of Practice is a chance for Europe to set clear and practical expectations for marking and labeling AI-generated content. By keeping the Code flexible, grounded in technical reality, and focused on meaningful transparency rather than overwhelming users, the EU can support trust and innovation as the AI Act is implemented. A balanced and thoughtful approach will help ensure the Code becomes a valuable tool for companies and a reliable source of clarity for people across the EU.