EU AI Act Article 50: What Voice Work Must Disclose

Since 2 August 2026, if you publish audio that a machine generated or manipulated, you have to say so. Article 50 of the EU AI Act is now enforceable, and it reaches further into voice work than most people in the industry realise. It applies whether the voice is fully synthetic, cloned from a real performer, or a human recording that an AI tool has altered.

This guide covers what the obligation actually says, who carries it, and what counts as compliance. It is written for people who commission, record or publish voice — not for lawyers.

What changed on 2 August 2026

The AI Act phases in over several years. Article 50 — the transparency chapter — is the part that landed this August. It creates two separate duties that are easy to confuse.

The first is machine-readable marking. Systems that generate synthetic audio must mark their output so that it can be detected as artificially generated. This is invisible to the listener. It lives in the file.

The second is disclosure to people. Where the content is a deepfake — audio that resembles a real person — whoever publishes it has to tell the audience that it is artificial. This one is visible, and it has to appear at first exposure. Buried in a footer after the listener has already heard the clip does not count.

Both duties can apply to the same piece of audio at the same time. They are not alternatives.

Provider or deployer — which are you?

The Act splits obligations between providers and deployers, and the split matters because most voice studios turn out to be both.

A provider builds or supplies the AI system. If you have trained a voice model, fine-tuned one, or ship a text-to-speech product under your own name, the marking obligation is yours.

A deployer uses an AI system in the course of their work. If you licence a commercial TTS engine and use it to voice a client's explainer video, you are a deployer, and the disclosure obligation is yours.

Here is where studios get caught. A production house that fine-tunes a cloned voice for a client and then publishes the result is acting as both. It owes the marking and the disclosure. Assuming your vendor has handled the first one is a common and expensive mistake — check the contract, and check the actual output files.

What machine-readable marking means for audio

The Act does not name a technology. It states an outcome: the output must be marked in a machine-readable format and detectable as artificially generated or manipulated. The Code of Practice finalised on 10 June 2026 fills in the gap, and it points toward a layered approach rather than a single technique.

In practice that means two things working together:

• Cryptographically signed metadata — C2PA manifests that travel with the file and record how it was made
• Imperceptible watermarking — a signal embedded in the audio itself, designed to survive compression and re-encoding

Metadata alone is fragile. Strip it, re-encode the file, or run it through a platform that rewrites headers, and the provenance is gone. Watermarking survives that, which is why the layered approach exists. Available options include AudioSeal, Google's SynthID for audio, and C2PA manifests for the distribution chain.

Be honest about the state of this technology. Audio watermarking is less mature than the image equivalent, and robustness against determined removal is still an open problem. The obligation is legal, not technical — you have to make a genuine effort with the tools that exist, not achieve perfection.

The artistic exemption, and its limits

Article 50 carves out content that is evidently artistic, creative, satirical or fictional. For that material, disclosure can be minimal and non-intrusive — it must not spoil the work.

This is narrower than people assume. The exemption softens how you disclose. It does not remove the duty. A satirical podcast using a synthetic politician's voice still discloses; it just does not have to stamp a warning over every line.

And the exemption does not touch the marking obligation at all. Artistic intent has no bearing on whether the file carries a machine-readable signal. That duty stands regardless.

One more thing worth stating plainly: deepfakes must be labelled even when the content is entirely lawful. Having permission from the person whose voice you cloned does not exempt you. Consent and disclosure are separate questions.

What this means if you hire or record voice

For buyers, the practical change is contractual. Anyone commissioning voice work should now be asking suppliers a direct question: is any part of this synthetic or AI-processed, and if so, what marking is applied? Get the answer in writing. The disclosure obligation follows the published content, which usually means it follows you.

For performers, the Act adds weight to something the industry was already circling. Voice cloning consent is now entangled with a public labelling regime, which makes the scope of a licence much harder to leave vague.

The German market has already moved on this. The VDS Gagenkompass — the fee survey published by the German voice association — states explicitly that even unlimited usage rights do not include use for machine learning or artificial intelligence. That has to be licensed separately, if at all. It is a useful model: treat AI training as its own right, priced and negotiated on its own terms, never bundled into a buyout.

That principle is being tested right now. Much of the German dubbing industry has refused Netflix work since February 2026 over a contract clause covering exactly this — AI training rights bundled into a general rights assignment. We track that dispute in our page on the Netflix dubbing conflict.

Penalties, and how seriously to take them

Infringements of Article 50 carry fines of up to €15 million or 3% of global annual turnover, whichever is higher. That is the tier below the prohibited-practices ceiling, but it is not a rounding error for anyone.

Enforcement in the first year is likely to focus on large platforms and obvious bad actors rather than mid-sized production companies. That is not a reason to ignore it. It is a reason to get the paperwork right now, while the cost of doing so is a contract clause and a metadata check rather than a remediation project.

A practical starting point

Three steps cover most of the exposure:

1. Audit what you publish. Identify every piece of audio in your pipeline that was generated or altered by an AI tool. Include the ones you think of as minor — noise reduction is fine, but generative fill and voice conversion are not.
2. Ask your vendors for their marking documentation. If a TTS supplier cannot tell you what marking their output carries, that is your answer about how prepared they are.
3. Write AI use into your talent contracts explicitly. Separate the recording licence from any AI training or voice-cloning right. Price them separately. Do not let a total buyout silently swallow both.

There is also a simpler route, and it is worth saying out loud: for a great deal of commercial work, a real performer remains the faster and cleaner option. No marking obligation, no disclosure question, no provenance chain to maintain. If you need a native-language voice for a campaign, briefing a human sidesteps the entire compliance question — and usually sounds better in the languages that matter to you.

Where synthetic voice genuinely is the right tool — scale, iteration speed, languages you cannot staff — then treat Article 50 as a design requirement rather than an afterthought. Build the marking into the pipeline, not onto the end of it.