EU AI Transparency Rules: Deepfakes Must Be Labeled
New European Union transparency obligations for AI-generated content took effect on August 2, 2026, requiring clear labeling of deepfakes, AI-generated images, videos, and chatbots across the bloc. The rules, enacted under Article 50 of the EU AI Act, aim to help citizens identify AI-generated content in their daily lives and combat the spread of deceptive synthetic media.
Context: The EU AI Act’s Phased Implementation
The AI Act (Regulation (EU) 2024/1689), the world’s first comprehensive legislation on artificial intelligence, entered into force on August 1, 2024, and has been implemented progressively. Following earlier phases covering AI literacy and prohibited practices in February 2025, and governance and general-purpose AI model rules in August 2025, the European Commission began enforcing the transparency obligations under Article 50 on August 2, 2026.
Henna Virkkunen, the European Commission’s Executive Vice-President for Tech Sovereignty, Security and Democracy, emphasized the significance of this milestone: “AI is a transformative technology that can bring extraordinary benefits to our people and businesses. But we are also seeing that harms can occur if AI is not properly designed and used and the most advanced models create risks on an entirely new scale. Europe anticipated this development.”
What the New Rules Require
The transparency guidelines distinguish between two categories of actors with different obligations: providers who develop AI systems (such as OpenAI, Google Gemini, and Anthropic Claude) and deployers who use these systems (including platforms like Instagram, Facebook, and TikTok).
Providers must:
- Design AI systems so individuals are explicitly informed whenever they interact directly with an AI system
- Add machine-readable marks (watermarks, metadata) to enable detection of AI-generated or manipulated content
Deployers must inform individuals when exposed to:
- Emotion recognition and biometric categorisation tools
- Deepfakes
- Text publications on matters of public interest without human review or editorial control
Trisha Meyer, a researcher in digital governance at the Brussels School of Governance (VUB), explained the practical implications: “Providers must have a marking or labelling system that can take the form of watermarks or metadata, for all types of content: text, video, image.” She noted that this will make it “much easier for fact-checkers, for example, to detect AI-generated and manipulated content.”
Defining Deepfakes
Under the AI Act, content qualifies as a deepfake when three cumulative criteria are met: resemblance (a high degree of similarity to existing subjects), existence (subjects that exist or could plausibly exist), and false appearance of authenticity or truthfulness. The European Commission’s FAQ clarifies these definitions and provides practical examples.
The EU has also created a set of free black-and-white icons that deployers may use to label AI-generated content, with variations for fully AI-generated versus partially AI-modified content.
Enforcement and Sanctions
Non-compliance with the transparency obligations can result in fines of up to EUR 15 million or 3% of a company’s worldwide annual turnover. Higher fines—up to EUR 35 million or 7% of turnover—apply for prohibited practices. The AI Office enforces rules for general-purpose AI models and systems integrated into very large online platforms, while national competent authorities oversee other AI systems.
More than 180 organisations have signed the Code of Practice on Transparency of AI-generated Content, drawn up by independent experts under the supervision of the EU AI Office. Signatories can rely on the code to demonstrate compliance with marking and labelling obligations.
Belgian Context: Supervisory Authorities Not Yet Designated
Belgium has not yet formally designated its supervisory authorities for the AI Act. According to Etienne Mignolet, spokesperson for the SPF Economie, the federal public service is preparing a draft law designating federal authorities, with the IBPT (Belgian Institute for Postal Services and Telecommunications) as central authority. “The 2 August 2026 deadline for designating these competent authorities proved difficult to meet,” Mignolet acknowledged.
The draft law will be submitted to the government in first reading after the parliamentary summer recess. In the meantime, citizens can contact the European Commission’s Service Desk, with requests handled by the AI Office.
Industry Concerns and Expert Skepticism
The tech industry has expressed concerns about the scope of the new rules. Boniface de Champris, AI policy lead at CCIA Europe, told The Guardian that the guidelines have expanded the definition of deepfake beyond the original AI Act: “Like cookie banners, once labels are everywhere, users stop noticing them. That doesn’t protect anyone, it just burdens commercial activity across the board.”
Laurence Dierickx, a researcher at the Université libre de Bruxelles (ULB), pointed to practical limitations of watermarking techniques: “All these marking techniques can be circumvented: you can crop, take a screenshot, add a bit of human input, or just remove the Gemini symbol.” She added that “real fraudsters, those with criminal intent who use fake videos to scam people, will always slip through the net. They don’t care about EU fines.”
Markus Reinisch, Meta’s vice-president for public policy in Europe, cautioned: “We must avoid a situation where online content is subject to a growing array of different labels and disclosures that end up overwhelming people.”
Grace Period and Forward Look
Existing AI systems on the market before August 2 have until December 2, 2026, to comply with marking and detection obligations. Content generated before August 2 does not need to be retroactively labelled. The AI Omnibus has postponed high-risk AI system rules to December 2, 2027, and high-risk AI systems in regulated products to August 2, 2028.
Sergey Lagodinsky, the Green MEP who helped negotiate the AI Act, framed the rules as essential for democratic integrity: “It is a matter not only of customer protection, it’s also a matter of democracy protection. Making transparent this information is something which we need to preserve our democracy and the authenticity of facts online.”
As enforcement begins, the coming months will reveal how effectively these transparency measures work in practice—and whether they can keep pace with rapidly evolving AI technologies. Trisha Meyer noted that “we won’t really see results before the end of this year,” given the grace period for labelling systems. For now, the EU has taken a significant step toward making AI-generated content identifiable, even as questions remain about the limits of technical solutions and the willingness of bad actors to comply.