ShadowGuardEU AI Act Labeling Rules for Creators: AI Video and Deepfakes (2026)
Since August 2, 2026, Article 50 of the EU AI Act requires two things: AI tools must mark what they generate, and whoever publishes a realistic AI deepfake must tell viewers. If your audience includes people in the EU, the second part can apply to you. This guide explains who is covered, what counts as a deepfake and what a valid label looks like. It is a plain-language summary, not legal advice.
What Changed on August 2, 2026
The transparency obligations in Article 50 started applying on August 2, 2026. On July 20, 2026 the European Commission published its final guidelines on them, and it has confirmed a voluntary Code of Practice on Transparency of AI-Generated Content as adequate. AI generators already on the market before August 2 have until December 2, 2026 to add machine-readable marking. Content generated before August 2, 2026 does not need to be labelled retroactively.
Two Duties, Two Different People
AI tools: mark every output
Providers of AI systems that generate synthetic audio, images, video or text must mark the output in a machine-readable format so it can be detected as AI-generated or manipulated. In practice that means Content Credentials (C2PA) metadata and watermarks.
Whoever publishes a deepfake: tell the viewer
Deployers of an AI system that generates or manipulates a deepfake must disclose that the content is artificially generated or manipulated. The marking inside the file does not count as that disclosure.
Does This Apply to Creators?
The AI Act defines a deployer as anyone using an AI system under their authority, "except where the AI system is used in the course of a personal non-professional activity." A clip made for friends falls outside that. A channel run for income, sponsored content and work for clients are professional activity, so creators publishing realistic AI content for an EU audience should expect the disclosure duty to apply. The guidelines don't name influencers specifically, and borderline cases, such as a hobby account that earns a little, are not settled.
What Counts as a Deepfake
The test is objective: it doesn't require an intention to deceive, and photorealism is an indication rather than the deciding factor. What matters is whether the content resembles a real person, object, place or event and would falsely appear authentic.
| Example | Visible label? | Why |
|---|---|---|
| A realistic AI video of a real person saying something they never said | Yes | It resembles an existing person and would falsely appear authentic |
| Realistic AI footage of an event that never happened | Yes | Events and places count, not only people |
| A face swap or a cloned voice in otherwise real footage | Yes | The guidelines list face replacement, voice cloning and composites as changes that need marking |
| Noise reduction, minor cropping or translated subtitles | No | Standard editing that does not substantially alter the content is excluded |
| A talking cartoon animal or an obviously impossible scene | No | No audience would take it as genuine |
Sources: Regulation (EU) 2024/1689 (AI Act), Articles 3 and 50; European Commission, Guidelines on the transparency obligations under Article 50 (July 20, 2026); European Commission, Code of Practice on Transparency of AI-Generated Content.
How the Label Has to Look
- It has to be clear and noticeable at first exposure — a note that is easy to overlook, such as one buried in terms of use, does not count.
- Machine-readable marking alone does not discharge the obligation to tell viewers.
- If people may not see the start of the content, the disclosure has to be repeated later; for live content it has to stay on.
- The Commission finalised three EU icons: one for fully AI-generated content, one for AI-modified content and a basic icon with an interactive second layer. Audio-only content needs a spoken disclaimer instead.
Artistic and Satirical Work Gets a Lighter Rule, Not an Exemption
For deepfakes that are part of an evidently artistic, creative, satirical or fictional work, the duty is limited to disclosing the AI generation in a way that doesn't hamper the display or enjoyment of the work. The guidelines add that this character must be evident, that where a work is partly informative the informative character prevails, and most advertising does not qualify.
Where C2PA Fits In
The provider duty is why more AI-generated files now arrive with Content Credentials already inside them. Platforms read that metadata: TikTok says it has labelled more than 3 billion videos as AI-generated using Content Credentials, creator labels and invisible watermarking. So the metadata in your export can trigger a platform label whether or not you add one yourself — and removing it doesn't remove your own duty to disclose.
Fines
Breaches of Article 50 can be fined up to €15 million or 3% of worldwide annual turnover, whichever is higher, with lower caps for small companies. Enforcement sits with national market surveillance authorities.
What To Do About It
- Decide whether your account is professional. Personal, non-professional use is outside the deployer duty. A channel run for income, sponsored posts and client work are not.
- Label realistic AI people, places, events and voices. Turn on the platform's AI label and add a visible on-screen note where viewers in the EU may not see the caption.
- Keep the label visible where viewers enter. Clips, stitches and long videos are often watched from the middle; a label only in the first second may not be enough.
- Don't strip Content Credentials. Removing the file's metadata doesn't remove your duty to disclose, and platforms increasingly read it anyway.
- Make satire obviously satire. The lighter rule for artistic and satirical work only applies when that character is evident to the viewer.
- Check the file before you post. Know whether your export carries AI Content Credentials, so a platform label never comes as a surprise.
Key Terms
Frequently Asked Questions
I'm not based in the EU. Does this apply to me?
Article 50 is tied to AI output used in the EU, not to where you are based. How national authorities will pursue individual creators outside the EU is not settled, but if EU viewers are part of your audience, labelling realistic AI content is the safe default.
Is the platform's AI label enough?
The guidelines don't name platform labels. A label shown on the video when it starts playing is the closest fit to "clear at first exposure"; a hashtag or a line deep in the caption is the kind of disclosure that can be overlooked.
Do I need a label if I only used AI to edit?
Not for standard editing that doesn't substantially alter the content, such as noise reduction, minor cropping or translation. Face replacement, voice cloning and composites are treated differently.
Can ShadowGuard tell me whether my video is a deepfake?
No. ShadowGuard reads the Content Credentials in MP4 and MOV files and shows whether they declare AI generation and which tool wrote them. Whether a video resembles a real person and needs a label is a judgment it doesn't make — and nothing here is legal advice.
Know What Your File Says Before It Goes Live
ShadowGuard reads the Content Credentials in MP4 and MOV files and shows whether they declare AI generation, alongside the frame-by-frame visual check.