On August 2, Europe Starts Requiring a Label on AI Voices. Here Is What Actually Changes.

On 2 August 2026, a set of rules inside the EU AI Act quietly switches on, and one of them is about the sound coming out of your speakers. Under Article 50, AI systems that generate synthetic audio, images, video or text must mark their output so that a machine can detect it as artificially generated. And anyone deploying an AI-generated deepfake — which, in the law's wording, reaches a synthetic voice made to pass as real speech — must disclose that it is artificial. If any part of your work involves an AI voice reaching listeners in Europe, this is no longer a future policy debate. It is a date on the calendar, and it is nearly here.

It helps to see that Article 50 is really two obligations pointed at two different people. The first lands on the provider — the party that builds the AI system and puts it on the market. They have to embed a mark in the output itself: a watermark or metadata that travels with the file and lets a detector confirm it was machine-made, even after it has been copied around the internet. The second lands on the deployer — the party that uses the system to make content and puts that content in front of an audience. They have to disclose, in a way a person can actually perceive, that a deepfake is artificial. Same piece of audio, two separate duties, and — this is the part that trips people up — often two separate companies.

Now put yourself in it. You licensed an AI voice, typed in your script, and generated a narration for a product video that will run in Germany. Are you the provider or the deployer? You did not build the model, so not the provider. But you took its output and published it to an audience, which is close to the textbook definition of a deployer — and the deployer's disclosure duty may be yours. The tool's terms of service will have a view on this, and it may not be the view you would have chosen. The Act even carves out an exemption for work that is clearly artistic, creative or satirical, which sounds reassuring until you try to decide whether your brand's explainer video qualifies. The honest summary is that the line runs straight through the middle of an ordinary AI-voice purchase, and 'the tool handles it' is an assumption, not a fact.

Underneath the legal roles sit two technical layers, and you need both. One is the machine-readable mark on the file — the watermark or provenance metadata the provider is meant to embed, which a person listening will never hear. The other is the human-perceptible disclosure — the on-screen note or the spoken line that tells an actual listener the voice is synthetic. To help everyone hit a moving target, the European Commission finalised a voluntary Code of Practice on the Transparency of AI-Generated Content on 10 June 2026; signing it is one recognised way to show you have done enough. There is a little breathing room, too: for AI systems already on the market before 2 August, the machine-readable-marking part has an extension to 2 December 2026. The expectation that you disclose does not wait that long.

If this is starting to feel like homework you did not sign up for, that is the honest shape of it. You came to AI voice to get an ad or a course out the door quickly and within budget — not to work out whether you count as a 'deployer' under an EU regulation, whether your file carries a compliant watermark, or whether your voiceover needs a spoken disclaimer. But 'fast and cheap' has a way of quietly handing you a compliance question along with the audio, and a raw file from a self-serve tool usually arrives with none of this settled. Whether it is marked, whether you can prove where the voice came from, whether you are the one who has to disclose — all of it lands back on your desk.

This is where working with a studio rather than a raw generator quietly earns its keep. A supplier who takes this seriously hands you audio that already carries a machine-readable mark at the file level, tells you plainly whether a given voice is AI or a real person so you can make the right disclosure call, and keeps the provenance on file — a named professional, an explicit authorization — so that if anyone ever asks where the voice came from, there is a document rather than a shrug. That is close to how Onyx already operates: we came out of running a voice studio in Taiwan since 2008, every AI voice in our library traces to a named actor who recorded under a written licence, and marking and provenance are things we treat as part of delivery, not paperwork we leave on your side of the table.

So the takeaway is not to panic about a date on the near horizon; it is to ask one specific question before you ship anything with an AI voice into Europe: who marked this file, and who has to disclose it — you or your supplier? If you cannot answer that from the tool you are using, that is worth knowing now rather than after 2 August. Tell us what you need voiced, and for which markets, and we will make sure what comes back is labelled, traceable, and clear about which part is yours to disclose — so the deadline becomes one less thing on your list.