EU AI Act: правила прозрачности ИИ вступили в силу — что осталось нерешённым

On August 2, 2026, the European Union officially brought into force the AI content transparency requirements under Article 50 of the EU AI Act — the first mandatory standard for labeling synthetic content for the world's largest digital market. Simultaneously, the Code of Practice (CoP) for generative AI providers came into effect.

What providers are now required to do

Providers of generative AI systems are required, as of August 2, 2026, to fulfill four key requirements under the Code of Practice:

• Multi-level labeling — embedding digital metadata and watermarks into generated content
• Free detection tools — providing users with AI content detectors compliant with EU privacy laws
• Quality standards — ensuring robustness and interoperability of labeling systems
• Compliance processes — regularly testing and confirming fulfillment of requirements

Deployers of AI systems are required to disclose to users that they are interacting with AI — in accordance with Section 2 of the Code. This is the first large-scale mechanism giving audiences the right to know the origin of content.

Why experts call the rules incomplete

The key weakness of the rules is broad exemptions. The EU AI Act does not require labeling for minor modifications (cropping, framing, AI translation), content on matters of public interest, and materials for personal non-commercial purposes. The personal exemption is cited as the most dangerous: it removes privately produced viral deepfakes from regulation — those that cause the greatest harm.

"A label is not protection, and the two should not be confused," note the authors of the

Tech Policy Press report published on the date the rules came into force.

Content provenance information remains optional — tracing the full path of synthetic material across the entire distribution chain is still impossible. Another systemic problem is open models with open weights: voluntary compliance loses force the moment a model is released and fine-tuned by third parties. Public accountability and independent civil society oversight also remain limited.

How reliable is AI content detection?

AI content detectors are not yet capable of serving as a reliable protection tool. According to the TRIED benchmark, specifically developed to evaluate synthetic content detection systems, results remain unreliable and uneven — quality varies significantly depending on language and the subject of the recording. The authors explicitly warn: detectors cannot be regarded as a reliable last line of defense.

When AI content is edited and redistributed, watermarks and metadata are easily lost. Without robust provenance mechanisms, labeling becomes a symbolic gesture.

The trend is not limited to the EU: the states of California and New York are introducing similar measures in 2026, drawing on the principles of the Coalition for Content Provenance and Authenticity (C2PA). However, no jurisdiction has yet addressed the problem of synthetic content systemically.

What this means

The EU AI Act has taken the first mandatory step toward AI content transparency — but critical loopholes remain. Optional provenance, the personal use exemption, and unreliable detectors mean that real protection is still ahead. The key question: will regulators close these gaps, or wait for the industry to sort things out.

Frequently Asked Questions

What is the EU AI Act Code of Practice?

The Code of Practice (CoP) is a mandatory set of rules for generative AI providers in the EU, which entered into force on August 2, 2026, together with the Article 50 requirements. It obliges platforms to implement content labeling, provide free detectors, and regularly undergo compliance checks.

What content is exempt from labeling?

Exempt are minor modifications (cropping, framing, AI translation), content on matters of public interest, materials for personal non-commercial purposes, and content generated in real time where labeling is technically impossible. Experts criticize the personal exemption as the most dangerous from the standpoint of deepfake proliferation.