EU AI Labelling Rules Turn Deepfake Disclosure Into A Platform Workflow Test

AI Disclosure Moves Into The User Interface

The European Union has turned AI-content transparency into a detailed design and operations problem for platforms, publishers and app operators.

Under Article 50 of the EU AI Act, the new Code of Practice sets out how AI-generated or AI-manipulated content should be identified when people first encounter it.

The framework separates providers, which build AI systems, from deployers, which use those systems to create and publish content.

That distinction is important because the visible disclosure burden falls heavily on the party placing the content in front of users, not only on the company that built the model.

For deployers, the clearest target is deepfake content: AI-generated or manipulated images, audio or video that can resemble real people, objects, places, entities or events and appear authentic.

The rules also cover AI-generated text on matters of public interest when it is published without human review, editorial control or an accountable publisher.

Labels Must Appear Before Ambiguity Spreads

The Code of Practice is specific about timing.

AI disclosures must appear no later than the first interaction or first exposure, so a label buried after a video, placed below a text item or added after the viewer has already engaged with the content would not meet the stated approach.

The EU provides icons for fully AI-generated material, partially modified material and a minimal AI marker that can support a custom interactive layer.

Custom labels are still allowed, but they must preserve the central AI identifier and remain clear at a context-appropriate size.

Placement rules vary by format.

Visual labels should sit where overlays do not obscure them, such as a corner of an image or video.

Video labels need to appear at the start and then recur at regular intervals, including after ad breaks or other interruptions.

For published text, the disclosure should be near the headline, at the top of the item or in an early colophon-style position.

Accessibility Becomes A Compliance Requirement

The rules do not treat disclosure as a purely visual badge.

Audio-only deepfakes need a spoken notice in plain and simple language at the beginning, with repeated reminders for long-form or live audio.

If a screen is available, such as a smartphone or car display, a visual icon is required alongside the audio notice.

Accessibility requirements add another operational layer.

The Code calls for audio descriptions of visual labels, tactile or haptic cues for audio content, high-contrast icons, screen-reader compatibility and detectability by assistive technologies.

That makes the obligation relevant to product design, content-management systems and moderation workflows, not just legal policy teams.

Artistic and satirical uses are not exempt.

They receive more flexible placement options, but the content still needs a disclosure that does not interfere with the work’s display or enjoyment.

For websites, apps and smart glasses, the label can sit beside the content or inside an interface overlay if users can perceive it without taking a dedicated action.

Publishers Need Accountability, Not Just Icons

The most sensitive area is public-interest text.

Licensed media organisations can rely on editorial processes if those processes satisfy the Code’s requirements.

Other deployers must establish internal editorial-control policies that identify the person or legal entity responsible for publication and explain the organisational measures used for human review or editorial control.

The Code also pushes deployers toward documentation, staff training and correction channels.

It encourages mechanisms for trusted flaggers, independent researchers, academics and fact-checkers to report missing or incorrect disclosures, with mislabelling reviewed and remedied without undue delay.

The practical watchpoint is whether platforms can make these disclosures reliable across video, audio, text and mixed interfaces without turning labels into decorative compliance marks.

The EU framework shows what transparency should look like at the content layer; enforcement will depend on whether deployers can prove that labels, accessibility features, editorial responsibility and correction processes work in normal publishing flows.