EU AI Act Deepfake & Synthetic Content Labeling Requirements 2026
Article 50 of the EU AI Act isn't just about political deepfakes. It reaches AI-generated marketing images, synthetic voiceovers, and even AI-written articles — and it requires clear, identifiable labeling before that content reaches anyone in the EU.
What Article 50 Actually Requires
Most EU AI Act coverage focuses on high-risk systems — hiring tools, credit scoring, biometric identification. Article 50 sits alongside those rules but applies far more broadly: it's a transparency obligation that attaches to the output of certain AI systems, regardless of whether the underlying model is classified as high-risk.
If content is AI-generated or AI-manipulated and would appreciably resemble a real person, object, place, entity, or event in a way that could falsely appear authentic, the provider or deployer has to disclose that it's artificial — clearly, and in a way a viewer can actually identify, not buried in a terms-of-service page or metadata nobody sees. The test isn't intent to deceive — it's whether the content could pass as real.
What Content Types Are In Scope
AI-generated or altered images
Product photography retouched or generated with AI that depicts realistic people or scenes, AI-generated headshots, and composited marketing visuals that appear photographic all fall within scope if they could be mistaken for authentic.
Synthetic audio and voice cloning
AI-generated voiceovers, cloned voices used in ads or customer service, and AI-narrated content that mimics a real or realistic-sounding speaker require disclosure under the same standard.
AI-generated or manipulated video
This is the category most people associate with 'deepfakes' — video content depicting people saying or doing things that didn't happen, including AI-generated avatars used in place of real presenters.
AI-generated text on public-interest topics
A narrower rule applies to text: AI-generated content published to inform the public on matters of public interest must be disclosed as AI-generated, unless a human has reviewed it and takes editorial responsibility — a carve-out that puts real weight on documented human review.
Who Carries the Obligation
The AI Act splits responsibility between the entity that builds the AI system and the entity that uses it to generate content:
- Providers of AI systems that generate synthetic audio, image, video, or text content must ensure outputs are marked in a machine-readable format detectable as artificially generated — a technical, embedded signal, not a visible label.
- Deployers — the businesses actually using the tool to produce marketing content, customer-facing media, or public communications — carry the separate obligation to disclose to end users, in a clear and visible way, that the content is AI-generated or manipulated.
- Both obligations can apply at once. A marketing team using a third-party AI image generator needs the tool to embed machine-readable markers, and the marketing team itself needs to add a human-visible disclosure before publishing.
Preparing Before Enforcement Ramps Up
Treat labeling as a publishing-workflow requirement, not a legal afterthought — it has to happen before content ships, not after a complaint.
Undisclosed AI content is a compliance gap you can catch before it ships
RatedWithAI helps teams find and fix the compliance gaps on their web properties before they become complaints. Start with a free scan.
Scan Your Site for Free →Frequently Asked Questions
Does Article 50 apply to internal, non-public AI-generated content?
No. The labeling obligations are triggered by content reaching natural persons — internal drafts, prototypes, or content never published or shown to end users generally fall outside scope. The obligation attaches when the content is disclosed or made available to people.
Is a small watermark or metadata tag enough to satisfy the disclosure requirement?
For the deployer-facing disclosure, the standard is that the label must be clear and identifiable to the person viewing the content — a marker buried in metadata that a typical viewer would never see is unlikely to satisfy that standard, even if it satisfies the separate machine-readable requirement for providers.
How does this interact with existing US or state-level deepfake disclosure laws?
US state deepfake laws typically target narrower categories such as election-related content or non-consensual intimate imagery. The EU AI Act's Article 50 is broader in scope but only applies where content reaches EU audiences — businesses operating in both markets generally need to track both regimes separately.
What is the human editorial review carve-out for AI-generated text?
AI-generated text on matters of public interest doesn't need a disclosure label if a natural person has reviewed the content and holds editorial responsibility for its publication — similar to how a human editor reviewing a wire-service draft can take ownership of the final piece.