EU AI Content Labeling: Transparency Code, Deepfakes, August 2026 Rules

What changed

The European Commission published a Code of Practice on marking and labelling AI-generated content on June 10, 2026. It is a voluntary tool, drafted by independent experts, that offers practical guidance to help providers comply with the AI Act. The underlying transparency rules become effective in August 2026.

The AI Act sets two layers of obligation. First, providers must ensure that AI-generated content is identifiable — a broad requirement to mark synthetic outputs so they can be detected. Second, certain content must be clearly and visibly labelled for people, specifically deepfakes and AI-generated text published to inform the public on matters of public interest.

Responsibility is split. Providers of generative AI systems bear the primary duty to make outputs identifiable and properly marked. Deployers — the businesses and people who use those systems to publish — share obligations to apply visible transparency measures. In plain terms: the tool maker must mark the output, and the publisher must disclose it.

Why it matters and for whom

The rules draw a useful distinction between machine-readable marking and human-visible labels. Marking is the technical signal embedded in a file — a watermark or metadata that detectors and platforms can read. Labelling is the disclosure a person actually sees: a caption, badge, or note saying the content is AI-generated. The AI Act wants both, for different audiences.

For media, this hardens existing editorial practice into law. AI-written or AI-assisted pieces on matters of public interest, and any deepfake imagery, need visible disclosure. For marketing and advertising agencies, synthetic visuals, voiceovers, and AI avatars in campaigns fall squarely in scope. For influencers and creators, the era of quietly posting AI-generated images or cloned voices without a label is ending in the EU.

The practical risk is uneven readiness. Many providers already embed marking — Google's SynthID is one example — but visible labelling at the point of publication is patchier, and that duty often lands on the deployer. A brand that uses a compliant tool can still fall short if it strips or omits the consumer-facing label.

There are limits and open questions. The deepfake rule includes proportionality for artistic and satirical works, and the public-interest text rule is narrower than "all AI text." Pure marketing copy is treated differently from news. Getting these boundaries right is exactly what the Code of Practice is meant to help with.

The voluntary nature of the code is easy to misread. The code itself is not binding, but the AI Act obligations it explains are. Signing up to the code is a way to demonstrate good-faith compliance and reduce regulatory risk; ignoring it does not exempt anyone from the August 2026 rules. For organizations, the calculus is simple — the code is the clearest available map of what "compliant" looks like before enforcement practice exists.

This also slots into a wider pattern. Major providers are already shipping machine-readable marks by default — Google's SynthID watermark on Gemini output is one example — which suggests the technical marking side will move faster than the human-facing labelling side. The gap to close is at the point of publication, where a deployer decides whether the audience actually sees a disclosure. That is the part regulators will be watching, and the part most likely to trip up brands and creators who assume a compliant tool covers them.

What to do and timing

The key date is August 2026, when the transparency obligations apply. The Code of Practice published on June 10 is the bridge — voluntary, but the clearest signal of what regulators expect.

Providers should ship robust machine-readable marking by default and document it. Deployers — newsrooms, agencies, brands — should build a labelling step into publishing workflows now: decide where the visible disclosure appears, who is accountable for it, and how to avoid stripping marks during editing or re-encoding. Because the duty is shared, contracts between tool vendors and publishers should make clear who guarantees what.

Do not wait for enforcement clarity to start. The cheapest compliance is a habit built before the deadline, not a scramble after it.

Bottom line

• The EU's new Code of Practice (June 10, 2026) guides compliance with AI Act transparency rules that take effect in August 2026.
• AI content must be machine-readable marked, and deepfakes plus public-interest AI text must be visibly labelled, with duties split between providers and deployers.
• Newsrooms, agencies, brands, and creators operating in the EU should treat AI disclosure as a built-in workflow step, not an afterthought.

Expect labelling to become a baseline expectation across the wider web, as EU rules pull platforms and global vendors toward disclosure by default.