Marking AI content from 2 August 2026: what the AP expects from your workflow

From 2 August 2026 the Dutch Data Protection Authority, together with the European AI Office, supervises the transparency obligations of the EU AI Act. You must then mark AI-generated image, audio, video and certain text content both machine-readably and visibly, and be able to demonstrate per workflow which content is AI content.

In its EU AI Act dossier, the Dutch Data Protection Authority (AP) describes that from 2 August 2026 it will supervise the transparency obligations for certain AI systems. The core of this: citizens must be able to see when they are communicating with an AI system and when content has been artificially generated or manipulated. For organisations that deploy generative AI in sensitive environments, this means that the distinction between AI content and 'real' content is no longer an interface choice, but a testable obligation.

What exactly is the Dutch Data Protection Authority announcing for 2 August 2026?

The AP positions itself as the national supervisor of the EU AI Act, working in cooperation with the European AI Office. According to the AP dossier, supervision of the transparency rules starts on 2 August 2026: systems that generate artificial content must clearly mark that output, and people must know when they are dealing with AI or AI-generated content.

The Dutch-language trade outlet Governance Web reports that the AP advises organisations to sign the European code of practice, which among other things provides EU icons for 'AI', 'AI generated' and 'AI modified'. On the basis of the EU AI Act, the AP supervises this transparency; the broader privacy protection for the personal data involved additionally remains subject to the General Data Protection Regulation (GDPR), the European privacy law known in Dutch as the Algemene verordening gegevensbescherming (AVG). In our assessment the practical message is that the AP forms a concrete trigger: from that date Dutch organisations will be held to account not only on principles, but on demonstrable implementation.

Which types of AI content must I mark and label?

The legal analysis by Greenberg Traurig of the European Commission's guidelines on Article 50 summarises which categories fall under the obligation. In brief, these are:

• chatbots and other AI systems that communicate directly with people;
• synthetic image, audio and video content, including deepfakes;
• certain AI-generated texts that inform the public on matters of general interest.

Importantly, the intention to mislead is not a precondition. According to the Greenberg Traurig analysis, content without a misleading purpose can also fall under the deepfake definition. 'Well-intentioned' AI applications must therefore also be recognisable as AI content.

What is the difference between machine-readable marking and visible labels?

The European Commission's Quick Facts distinguish two layers that exist alongside each other. That distinction is decisive for how you set up your workflow:

• Machine-readable marking. Providers of generative systems must technically mark the output, for example via watermarks or metadata, so that detection is possible.
• Visible or audible labelling. Users of deepfake systems must state visibly or audibly that the content was created or manipulated by AI, for example with the EU icons, disclaimers or captions.

The Quick Facts also mention a short transition period until December 2026 for marking generative systems that were already on the market before 2 August 2026. In our assessment this means you should not see marking as a single label, but as a chain of technical traceability and visible distinction for the end user.

When does the exception for human editorial control apply?

Not all AI-generated text needs to be labelled. The Bird & Bird analysis of the final guidelines on Article 50 explains that the labelling obligation for texts on matters of general interest applies unless there is demonstrable editorial control with clear responsibility. The same analysis emphasises that retroactive marking of synthetic content generated before 2 August 2026 is excluded, and that breaches are punishable with fines of up to 15 million euros or 3% of worldwide annual turnover.

In our assessment the exception for editorial control is above all a design choice. You must record where AI only prepares content under human supervision, and where AI is the primary publication channel. In the latter case you label systematically. The line between the two must be demonstrable, not merely plausible. See on this point also the audit rights and evidence obligations in AI contracts that the EU AI Act requires elsewhere.

How do I record marking and editorial control verifiably per workflow?

The practical consequence is that labels and icons are not sufficient without underlying registration. In our assessment the AP's supervision calls for a reconstructable chain. A workable approach per high-trust workflow:

1. Identify where in your processes AI-generated content arises.
2. Choose the right marking technique per channel: machine-readable (watermark, metadata) and visible (icon, disclaimer, caption).
3. Record which content went live under which label, with logging that can be checked later.
4. Determine and document per workflow whether human editorial control applies.
5. Retain evidence that is usable in audits, complaints or investigations.

These steps also touch on broader choices about data minimisation in generative AI workflows and on the fines under the EU AI Act from 2 August 2026. An overview of related obligations can be found in the topic hub on the EU AI Act and compliance.

For professionals in law, healthcare, supervision, finance and government, this is above all a documentation question. A verification layer such as Vera can support this by making visible per workflow which verification steps have been taken and what evidence is available; the choice of which content is marked as AI content and the final judgement remain with the professional. Vera is not a chatbot and not its own language model, and offers no guarantee of full compliance. The obligation to demonstrably separate AI content from real content lies with your organisation.

Sources: The article draws on the EU AI Act dossier of the Dutch Data Protection Authority, the Quick Facts and Article 50 guidelines of the European Commission, and legal analyses by Greenberg Traurig and Bird & Bird.

← All articles in this topic ← All articles