From 2 August 2026, the EU AI Act requires AI-generated or manipulated deepfakes to be disclosed as artificial. If you create realistic synthetic image, audio, or video that could pass for real, and any part of your audience is in the EU, you have to make clear that it was made or altered with AI. It does not ban deepfakes; it requires that they be labeled. Here is what the rule covers, who it applies to, when, and the exceptions.
What does the EU AI Act require for deepfakes?
The relevant rule is Article 50, the transparency part of the AI Act. Article 50(4) says that anyone deploying an AI system to generate or manipulate image, audio, or video content that is a deepfake has to disclose that the content has been artificially generated or manipulated. The obligation is about honesty with the viewer, not a ban on the technique.
That sits alongside a second duty further up the chain. Under Article 50(2), the providers who build generative AI systems have to mark their outputs in a machine-readable way so the content is detectable as artificially generated. So the law works on two levels: the tool maker marks the output at the source, and the person publishing a deepfake discloses it to the audience.
What counts as a deepfake under the AI Act?
The AI Act defines a deepfake broadly. It is AI-generated or manipulated image, audio, or video content that resembles real people, objects, places, entities, or events and would falsely appear to a person to be authentic or truthful. The test is whether it could mislead someone into thinking it is genuine.
That definition is wider than the popular sense of the word. It is not limited to a face swap of a celebrity. A photorealistic image of a person who does not exist, presented as a real photo, can fall within it. The line is realism and the impression of authenticity, not whether a specific real individual was copied.
Who has to comply, providers or deployers?
Both, in different ways, and it helps to know which one you are.
A provider is whoever develops or places the generative AI system on the market. Providers carry the Article 50(2) duty to mark outputs in a machine-readable format, with technical solutions that are effective, interoperable, and robust as far as is technically feasible.
A deployer is whoever uses such a system to produce and publish content. If you generate deepfake media and post it, you are a deployer, and you carry the Article 50(4) duty to disclose that the content is artificial. As a creator using an AI tool, you are almost always the deployer. The marking at the source is the tool maker's job; the visible disclosure to your audience is yours.
When do the rules apply?
The Article 50 transparency obligations apply from 2 August 2026. That date is the anchor to plan around.
There is some transitional detail still settling, and it matters mainly for tool makers, not creators. A provisional agreement reached on 6 May 2026 (the Digital Omnibus on AI) would give providers whose systems were already on the EU market before 2 August 2026 until 2 December 2026 to meet the machine-readable marking requirement, while systems placed on the market from 2 August 2026 comply from that date. The deployer disclosure duty that applies to you as a creator still starts on 2 August 2026. Treat that date as fixed and the transitional specifics as provisional until the final text is confirmed.
Are there exceptions?
Yes, and they are narrow. Where deepfake content is part of an evidently artistic, creative, satirical, or fictional work, the disclosure obligation is lighter: you still acknowledge that the content is generated or manipulated, but in a way that does not spoil the experience of the work. There is also an exception for use authorized by law to detect, prevent, investigate, or prosecute a criminal offense.
For a normal creator running a synthetic persona, neither exception removes the basic duty. An openly artistic project gets a softer form of disclosure, not a pass on it.
How do you actually disclose and mark AI content?
Two things travel together. The provider-side marking is machine-readable, embedded in the file so detectors and platforms can read it. The deployer-side disclosure is the one a human sees: a visible label that the content was generated or manipulated with AI.
The European Commission has been working out the practical detail. It published draft guidelines on the Article 50 transparency obligations in May 2026, and a Code of Practice on marking and labeling AI-generated content has been in draft with a final version expected around mid-2026. Both were still draft as this was written, so the fine detail of accepted labels may shift. The safe practical move does not change: mark realistic synthetic content visibly and clearly, at the point the audience meets it.
What this means for AI creators and personas
If you run a synthetic persona and any meaningful part of your audience is in the EU, treat the AI Act as another reason to disclose by default rather than a separate burden. The visible label you already use to satisfy platform rules and to keep your audience's trust does most of the work the AI Act asks for. Build the disclosure into how the account operates and you are covering the platform layer, the trust layer, and the AI Act at once.
That is how Cladegrove approaches it: characters that are openly AI, run consistently, with disclosure treated as part of operating the persona. For the wording, see the AI disclosure statement templates; for the platform side, see disclosing AI content on Instagram and YouTube; and for the ownership question, see who owns the copyright of AI images.
See how Cladegrove handles disclosed, consistent personas on the AI disclosure page.
The AI Act reads as heavy when you first open it, but for a creator the practical takeaway is small: if it could pass for real and an EU audience can see it, label it. We were already labeling everything by default for trust reasons, so the AI Act did not change the workflow, it just gave the habit a legal reason to exist. None of this is legal advice; for your own case, talk to a lawyer.
Fabio Ariotti, operator
Common questions
When does the EU AI Act deepfake disclosure rule start?
The Article 50 transparency obligations apply from 2 August 2026, and the deployer disclosure duty that affects creators starts on that date. A provisional agreement reached in May 2026 would give providers whose systems were already on the EU market before then until 2 December 2026 to meet the machine-readable marking requirement, a deadline that applies to tool makers rather than creators. The 2 August date is fixed; some transitional detail for existing systems is still being finalized.
Does the EU AI Act ban deepfakes?
No. It does not ban deepfakes, it requires them to be disclosed. A deployer who creates or manipulates deepfake image, audio, or video has to make clear that the content is artificially generated or altered. The point is transparency, not prohibition, with narrow exceptions for law enforcement and for evidently artistic or satirical work.
Does it apply to me if I am outside the EU?
It can. The AI Act reaches systems and outputs that are placed on or used in the EU market, so a creator outside the EU whose content is shown to an EU audience can fall within scope. If any meaningful part of your audience is in the EU, it is safer to assume the disclosure rule applies to your synthetic content.
What is the difference between the AI Act and platform AI labels?
They are separate layers that stack. Platform rules from YouTube, Instagram, and TikTok are private policies enforced by the platform. The AI Act is law, enforced by EU authorities, and it applies regardless of which platform you post on. Meeting a platform label does not automatically satisfy the AI Act, and vice versa.
