Explaining the EU Act on AI, Deepfakes and Disclosure. A Practical Guide and Breakdown of What it means for Creators and Film

Reading Time: 7 minutes

Published: January 12, 2026 | Last Updated: January 13, 2026

Add FilmDaft as a preferred source on Google

The EU AI Act deepfake disclosure is a transparency rule that requires a clear notice when AI is used to generate or change an image, audio, or video in a way that could falsely seem real. It does not permit you to use someone’s face, voice, or likeness. It also does not replace other rules around consent, privacy, copyright, or defamation.

This guide focuses on how disclosure works in practical terms for film and video, meaning what you show the viewer, what you add to the file, and what you document in case someone questions it later. This isn’t legal guidance. When in doubt, contact an entertainment lawyer.

Why deepfake disclosure matters in film and video

Film has always been built on illusion. But with AI, it’s now possible to make something look like it really happened, even when it didn’t.

A deepfake is an AI-generated or AI-altered audio, image, or video that convincingly imitates a real person, place, object, or event, so it can seem authentic. Deepfake is a portmanteau of deep learning (a type of machine learning) and fake.

Disclosure helps people understand what they’re looking at, especially when fake media looks like real evidence.

Here are Robert Downey Jr and Tom Holland inserted in Back to the Future using deepfake technology.

Deepfakes copy the cues people trust

Most people judge a scene based on what looks real. A believable close-up, a familiar voice, or a natural reaction can all suggest something actually happened. AI can copy those same signs. If your scene includes a real person’s voice or face, or if it shows something that looks like real footage, it can mislead people even if your project is fictional.

Disclosure builds trust, especially once clips are shared on platforms outside of your control, like trailers, social media edits, memes, and news commentary.

AI works like VFX, but faster and easier to misuse

Body doubles, stunt work, matte paintings, and compositing are common in film. AI tools now do similar work with faces and voices, but at a faster speed and lower cost. That speed makes it easier to publish realistic scenes without normal checks like legal review, broadcaster approval, or production notes.

If you need help on consent issues around AI-generated faces or voices, see Consent and Digital Replicas: What Creators Should Know.

Fiction doesn’t protect you from risk

What you make might be fictional, but what matters is how it’s seen. A synthetic clip inside a movie is different from that same clip shared as a teaser. The EU AI Act focuses on the person watching the content. That means you need to think about how and where your work will show up.

For more on ethics and responsibility, explore Ethics, Law, and Provenance in AI Filmmaking and the overview Artificial Intelligence in Filmmaking: A Practical Guide and Overview.

Where the EU AI Act puts the deepfake rules

The EU AI Act covers many kinds of AI tools. The rule about deepfake transparency comes from its transparency obligations, especially Article 50. The basic idea is simple: if AI-generated content might confuse someone about what’s real, you have to label it clearly.

Article 50 is the key rule for disclosure

Article 50 covers several types of AI transparency. One of those duties is for deepfakes. If your AI tool generates or changes media that could be mistaken for real people, events, or places, you must give a clear notice to viewers.

“Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated.”
Source: EU AI Act (Regulation (EU) 2024/1689), Article 50(4)

Important dates to plan for

The law came into force on 1 August 2024. The deepfake disclosure rule becomes fully enforceable on 2 August 2026. This matters for planning credits, delivery files, and marketing material.

The EU is also preparing support materials, like a Code of Practice on AI labeling. These may help clarify what counts as “clear and distinguishable” in real releases.

Provider vs deployer—what it means for you

The law uses role-based terms that map to real workflows:

Provider: the company that creates or supplies the AI system. They’re responsible for adding machine-readable labels to outputs.
Deployer: the person or team who uses the AI tool in a professional setting. This includes you, if you use AI to generate or change content that could pass as real.

If your AI tool doesn’t add a label, you still have to disclose. That means using on-screen notices, your own metadata, or file-level signals that don’t get stripped during upload.

What the EU means by “deep fake”

Many people think a deepfake only means a face swap. The EU AI Act uses a wider definition. It includes anything that looks like it could be real—people, objects, places, or events—if it was generated or changed by AI.

The legal meaning in simple terms

A deep fake is AI-made or AI-edited content that:

Resembles existing reality (a real person, place, or event), and
Could look authentic to a viewer, even if your project is fictional

Examples that usually count

Here are situations where disclosure is often required:

A synthetic voice imitating a real person
A recreated video of a real place that looks like news footage
A fictional scene that includes a synthetic clip people might mistake as real evidence

The Roadrunner documentary (2021) used an AI-generated voice of Anthony Bourdain. It caused backlash because people expected real audio in that format. Viewers felt tricked.

Cases that may not count—but require judgment

The law makes space for editorial judgment. Some uses are more like normal post-production:

AI tools that only clean up or retouch footage
Minor visual changes that don’t change the meaning

There’s also a rule for fiction, satire, or creative works. If it’s clear the work is fictional, disclosure must still happen—but it can happen “in an appropriate manner.” That usually means an opening or end credit is enough, unless the scene itself could still cause confusion.

What disclosure means in real-world production

Disclosure isn’t just a label. It’s a practical part of your post-production workflow. It’s about making sure the person watching your work understands what’s real and what’s not.

Make it visible and understandable

Viewer-facing notices matter more than metadata. Platforms often strip technical data, and people don’t inspect file properties. If your project might mislead someone, disclosure must be clear and readable on the version they actually see.

Places where disclosure can appear

On screen: opening card, lower third, or end credits
In context: social captions, trailer notes, broadcaster metadata, festival programs
In the file: metadata, watermark, or content credentials (if supported)

A repeatable workflow to follow

This step-by-step process helps you plan ahead and avoid confusion:

Inventory: list all AI-generated or edited media
Classify: ask “Could this be mistaken for real?”
Decide: is it a deepfake under the Act? Write your reasoning
Place: pick the best spot to add disclosure for that use
Mark: add metadata or machine-readable tags if possible
Write: use clear wording like “Some images were artificially generated or manipulated”
Test: check if the label is readable and shows up correctly on devices and platforms
Archive: save your notes and versions in a “proof pack”

Disclosure choices depend on the context

The law says disclosure must be “appropriate.” What’s appropriate depends on the type of work, the viewer’s expectations, and the release format.

Fictional features and series

In fictional work, the audience expects some visual trickery. You can often disclose in the end credits or with an opening title. A good example is Rogue One: A Star Wars Story (2016), which digitally recreated legacy actors. It didn’t mislead viewers because the work was clearly fictional. Still, trailers and behind-the-scenes clips need separate labels.

Documentary, factual, or news-style formats

In nonfiction, viewers assume they’re seeing reality. Even small changes (like synthetic voices or archival-looking inserts) can break that trust. In these cases, disclosure usually needs to appear near the altered material and again in the credits. This helps protect clips from being reposted out of context.

Trailers, social posts, and ads

Short formats travel quickly and often without your full framing. Add a visible on-screen notice plus a matching caption. This survives reposts better than metadata alone.

Provenance: showing what you did after the fact

Provenance means the records you keep to explain your choices. It helps when someone asks how your project used AI. It also helps your team stay on the same page when multiple vendors or versions are involved.

How machine-readable marking works

AI providers are supposed to add metadata that marks content as synthetic. This can include watermarks or content credentials. But not all platforms preserve this data. Re-encoding, cropping, and screen recording can remove it. That’s why visual disclosure and a paper trail are both important.

What to save in your “proof pack”

Tool info: tool name, version, vendor, and settings
Asset history: source footage, renders, exports with timestamps
Reasoning: why the tool was used and how you picked the placement
Consent forms: signed approval for faces, voices, or likenesses
Release tracking: which versions went where, with screenshots

Consent and disclosure solve different problems

Consent says whether you are allowed to use a likeness. Disclosure tells the viewer what they are seeing. You need both. Even with consent, you can mislead people without disclosure. Even with disclosure, you can break the law without consent.

For help with consent language, see Consent and Digital Replicas: What Creators Should Know.

Common mistakes that cause bad disclosure

Most teams don’t break this rule on purpose. But a few common ideas lead to bad decisions. Here’s what to avoid:

“It’s only a face swap, so it’s not a deepfake”

The law includes cities, events, and voices—anything that looks real when it’s not.

“I added metadata, so I’m covered”

Metadata is useful but unreliable. Most viewers never see it. Platforms can strip it.

“Fiction doesn’t need disclosure”

Fiction gets some flexibility, but not a full pass. You still need to add a visible notice.

“I’ll just put a line in the end credits”

That can work for fiction, but not for clips or factual content. Placement matters.

“Disclosure is part of our marketing”

Disclosure is not promo. It’s a plain-language label that helps people understand what they’re watching.

How to check your work before release

You can catch confusion before it happens. Run tests on both the viewer and the upload process.

Test what people see

Pick your riskiest scene. Show it to someone who hasn’t worked on the project. Ask them what they think is real, what seems fake, and whether the disclosure helped them understand it.

Test the platform output

Upload a test version. Check whether captions, metadata, and disclosure show up correctly. Pay extra attention to mobile playback, since that’s where many viewers will see it.

Keep your language consistent

Use the same sentence across all formats: credits, captions, press notes, and delivery forms. That avoids confusion. If you need simple language for non-technical stakeholders, try What Is AI? A Plain-English Guide for Creators.

Summing Up

EU AI Act deepfake disclosure is a transparency rule that applies to synthetic or AI-manipulated content that could be mistaken for real. Treat disclosure as part of your post workflow. Identify AI content, decide how visible it needs to be, place the notice clearly, and keep your records. Fiction may allow a softer placement. Factual formats and standalone clips often need earlier and clearer labels. Disclosure helps build trust, but it doesn’t replace consent or legal responsibility.

Read Next: Wondering where ethics meet AI tools?

Start with our full AI in Filmmaking overview to see how generative tools are changing writing, production, editing, and design.

Then head into our AI Ethics, Law & Consent section for real-world guidance on consent, disclosure, documentation, and accountability. These articles focus on practical risks and workflow choices—not just legal theory.

Whether you’re using voice models, AI clean-up, or generative images, this section helps you plan responsibly and protect trust in every phase of production.

Also, check out our full guide on AI Tools for Filmmaking to compare models, task types, and how different tools handle writing, editing, color, audio, and animation.

Official sources for the legal text and guidance

If you want to read the exact legal wording, start with the Official Journal version of the AI Act on EUR-Lex. The links below are the safest “source of truth” when you need to confirm definitions, timelines, and Article 50 transparency duties.