Updated October 2026 · Voice rights, consent and deepfake labels
The EU AI Act requires people who publish AI-generated deepfakes to disclose them, with a visible or audible label at the latest on first exposure. Reports say these transparency duties began on 2 August 2026, with some transitional relief. Check the official Regulation and the final Commission guidance before relying on any detail.
For podcasters and video makers in Europe, one of the most practical parts of the EU AI Act is the transparency rule for deepfakes. It asks people who publish synthetic content that resembles real people, objects, places or events to tell their audience. The rule is simple to state and easy to get wrong in practice, so it is worth reading the details carefully.
This post explains what the rule asks for, what has been reported about its timing, and how a creator can prepare. It draws on the Regulation, on the Commission’s draft guidance and on commentary from several firms. Where those sources disagree, the post says so. It is general guidance, not legal advice.
The Act describes a deepfake as AI-generated or manipulated content that resembles existing people, objects, places, entities or events and could falsely appear authentic. A cloned voice that sounds like a real person saying something they did not say falls within that description. A clearly fantastical scene, such as a dragon or a person flying unaided, is reported to fall outside it, because it cannot be mistaken for real.
The test is whether the content could falsely appear authentic. That makes the question about the audience as well as the content. A voice that listeners would reasonably take as real is the risk the rule addresses.
Commentary on the Act says the duty to disclose sits mainly with the deployer: the person or organisation that publishes the synthetic content. A podcaster who uses a generated voice in an episode, or a video maker who publishes a synthetic presenter, is likely to be the deployer. Providers of generative tools have separate duties, including making outputs detectable where technically feasible.
That split matters. A tool that adds a machine-readable mark does not, on the reports, replace the visible or audible label the creator must add.
According to a summary of the Commission’s FAQ, the label must be visible or audible, and it must appear at the latest on first exposure. A machine-readable mark added by a provider does not satisfy that duty on its own. For a podcast, that suggests a spoken disclosure near the start of the episode, and a written note in the description. For a video, it suggests an on-screen label as well.
The label should be clear, not buried. A disclosure hidden in a long description will not meet the test of appearing at the latest on first exposure, because the listener may never read it.
Reports describe lighter treatment for artistic and satirical works, where the disclosure must fit the work without hampering its display or enjoyment. That is a narrower exception than many creators assume, and it does not remove the need for a disclosure. If you are relying on it, take advice on whether your work qualifies.
Several sources say the transparency obligations began to apply on 2 August 2026. Some of them mention targeted transitional relief for the marking and detection duties on providers, which was contemplated in an omnibus proposal. Whether that relief was adopted, and how it changes the timeline, is not settled in the sources reviewed. The Commission’s draft guidance was published in May 2026, with a final version expected before the August date, but the final text was not confirmed in those sources.
Because these reports do not agree in every detail, check the Regulation itself and the Commission’s final guidance before making any plan that depends on a date.
A summary can tell you the general shape of a rule, but it cannot tell you how it applies to your show, your country or your audience. Read the Regulation itself, and then the Commission’s final guidance for your sector. Note the version and the date you read them, because guidance can be revised. If your work involves a commercial deal or a public body, take advice from a professional who knows EU law.
Keep a short file of the sources you used and when you checked them. If a dispute arises later, a dated record of your reading shows that you made a reasonable effort to follow the rule.
This post relies on the Regulation as summarised in several sources and on the draft guidance. Some sources are blogs and vendor pages, and at least one describes its article numbering inconsistently with the others. Treat the summary as a starting point, and verify it against the official text before you rely on it.
A disclosure protects trust, and the audio decides whether a listener stays. The Vocal Checker shows clipping, noise and level problems in a recording, and the Tunedzy Go page covers cleaning a phone recording. For the wider picture of voice rights, read the overview of voice cloning laws in this series.
Spot clipping, noise, rumble and level problems. Runs in your browser, nothing is uploaded.
Not every one. It applies to synthetic content that could falsely appear authentic. Check the definition in the Regulation.
Reports say no. The label must be visible or audible for the audience.
Several reports say 2 August 2026, with some transitional relief. Check the official text, because sources differ.
It depends on where the content is published and reaches people. Check the scope in the Regulation for your situation.
