AI voice clones under Article 50: consent settles the rights, not the transparency

Voice cloning has moved from experiment to production tool in a short time. Dutch agencies produce radio and TV commercials with AI voices of well-known actors, with explicit consent and a licence agreement drafted by an IP lawyer. That contract settles the rights properly. It does not settle the transparency duty.

That distinction is the core of this piece. Consent and personality rights sit in a different regime from Article 50 of the AI Act. The first question is whether you may use the voice. The second is whether the audience knows it is hearing an AI voice. Since 2 August 2026 that second question is binding law, and the answer does not depend on how good your contract with the voice actor is.

A consented voice clone is still a deepfake

The AI Act defines a deepfake as AI-generated or manipulated audio, image or video content resembling existing persons, objects, places or events, that would falsely appear to a person to be authentic. That definition looks at what the material does to the listener, not at whether it was created lawfully.

This is where the intuition of many agencies goes wrong. The reasoning runs: we have consent, so it is not a deepfake. But consent affects lawfulness, not perception. A listener who hears a familiar voice in a commercial assumes that person recorded it. If they did not, the very impression Article 50(4) addresses has been created. Whether the voice was paid for it changes nothing about what the listener believes they hear.

The Commission is explicit about this in its guidelines of 20 July 2026. Point 124 states that relying on the lighter regime for creative work, discussed below, cannot justify failing to respect the fundamental rights of individuals or the rights of rightsholders under Union law on intellectual property or data protection. The two regimes sit alongside each other. They do not replace one another.

Why commercials fall outside the artistic exemption

Article 50(4) provides a lighter regime for deepfakes forming part of evidently artistic, creative, satirical or fictional work. The duty does not disappear there, but it may be met in an appropriate manner that does not hamper the enjoyment of the work. Many producers hope a commercial qualifies, since advertising is creative work after all.

That hope does not hold. Point 122 of the guidelines excludes content whose nature is exclusively informative or commercial and recognisable as such. More importantly: where a deepfake combines several characters, for example informative and creative, the informative character always prevails and the ordinary labelling requirements apply. A commercial is commercial in nature by definition, however creative the execution.

Among its examples the Commission explicitly lists an AI-manipulated video with a realistic synthetic influencer showing only the functionalities of a sponsored product as a case falling outside the lighter regime. A radio commercial with the cloned voice of a well-known figure promoting a product sits in the same category.

The picture may differ for a production clearly recognisable as art or satire. Point 122 then requires that it be evident to the exposed person that the content falls in that category, and prescribes that those categories be interpreted strictly. Content whose nature may be unclear or ambiguous to the public falls outside it. When in doubt, the lighter regime is therefore not the safe choice but the risky one.

Who carries which duty in the chain

A commercial quickly involves four parties: the platform generating the voice, the agency creating the clone and delivering the production, the advertiser commissioning it, and the media agency buying and airing the spot. Article 50 allocates duties not by who pays the bill, but by role per obligation.

For a voice agency this usually means two things at once. If you use an external voice platform, the machine-readable marking under paragraph 2 sits with that platform and becomes a supplier question: does it mark the output, and how do you evidence that. If you publish yourself, or deliver to an advertiser who publishes, the visible or audible disclosure under paragraph 4 sits with the party actually bringing the spot to the public.

A common mistake is assuming the supplier's watermark suffices. Point 117 of the guidelines rules this out explicitly: deployers cannot rely on the machine-readable marking the provider embedded under paragraph 2, because those markings are not immediately clear and distinguishable to the people exposed to the content. A watermark in the file is not a disclosure to the listener.

Point 12 adds that in complex production and distribution chains you must take proportionate measures to ensure the label is genuinely displayed to the intended audience at the moment of first exposure. Concretely: contractual arrangements with distribution partners are part of compliance, because a label lost in transit is no label.

Point 14 closes an escape route that is tempting in advertising. A legal person remains the deployer even when it involves freelancers or contractors under its authority in operating the system. So you cannot push the duty onto the hired producer.

What this means in practice for a voice agency

The code of practice as a route, not an obligation

On 10 June 2026 the Commission published a code of practice on transparency of AI-generated content. Signing is voluntary and not signing does not constitute non-compliance. The code has a section for providers on machine-readable marking and one for deployers on deepfakes.

For those who sign and comply, demonstrating compliance is easier. Point 148 of the guidelines states that non-signatories are expected to demonstrate by other appropriate means how they meet paragraphs 2, 4 and 5, for example through a gap analysis against a code assessed as adequate, and are likely to receive more requests for information. Point 149 adds that commitments implemented in line with such a code may be taken into account as a mitigating factor when setting the level of a fine.

For an agency working structurally with synthetic voices that is a real trade-off: sign and follow the framework, or build a file of your own that withstands comparison with it.

Where this goes wrong

The three mistakes we see most often with synthetic voice are these. The first is assuming consent removes the transparency duty, which it does not. The second is assuming the commercial falls under the creative exemption, while the informative and commercial character prevails. The third is relying on the supplier's watermark, which is precisely what does not suffice towards the audience.

All three are straightforward to fix, and all three are cheaper to fix before a campaign runs than after.

The full breakdown per paragraph sits in our practical guide to Article 50, the provider and deployer split is worked out here, and the article text itself is in the AI Act Explorer. Organisations wanting to record this per production, from role determination to disclosure and registration, start at Embed AI. For teams that must recognise during production when a disclosure is needed, LearnWize provides role-based training.