AI-cloned voices: consent comes before convenience
An AI can rebuild any voice. But cloning a person’s voice needs consent, labelling and a clean purpose. My assessment as an auditor and managing director.
In short
An AI-cloned voice sounds like a real person but was never spoken by them. Cloning a person’s voice without consent violates their personality rights and data protection law. If the clone goes into circulation, the labelling obligation of the EU AI Act applies as well. Convenient is one thing, permitted is another.
Auf Deutsch lesen: deutsche Fassung
At a trade fair recently I was shown an AI tool that rebuilds any voice in a few minutes. A few minutes of audio material are enough, then the cloned voice says everything you type in. The salesman raved about the possibilities: the boss’s voice for the training video without him having to go into the studio. The voice of the well-known employee for the telephone announcement. A podcast in your own voice, even when you are ill in bed. Sounds practical. And this is exactly where it gets delicate. Cloning a voice has become technically trivial. Legally and humanly it is anything but harmless. This is my assessment as an auditor, not legal advice: convenient is one thing, permitted is another.
Your voice is like your signature
I come from the workshop and think in such images. A person’s voice is like their signature or their face. It belongs to them, it is an expression of their personality, and it vouches for them. If someone speaks with my voice, they speak in my name, whether I said it or not. German law protects this. The general right of personality, derived from Article 2(1) in conjunction with Article 1(1) of the Basic Law (Grundgesetz), the German constitution, also covers a person’s own voice. Nobody has to tolerate their voice being rebuilt without their agreement and having words put into their mouth that they never said.
This applies to the managing director just as much as to the employee on the hotline or the external voice actor. A voice is not ownerless property that you collect because a few recordings are lying around. It belongs to the person who has it, and nobody else may dispose of it without further ado.
And the harm is not theoretical. With a cloned voice you can make a person say things they would never say: promote a product, make a political statement, give a commitment that binds them. Whoever hears it believes the familiar voice. In the end the person affected answers for words that never passed their lips. That is an interference with the core of what makes a person who they are.
Without consent the clone is a violation
The crux is consent. Anyone who wants to clone and use the voice of a real person needs that person’s permission. And an informed one: the person must know what their voice is used for, to what extent, for how long, and whether they can withdraw it. A blanket yes, hidden somewhere in the employment contract, is as a rule not enough for this. A consent that permits everything and for ever is not real consent.
Anyone who thinks this is no problem within their own company is especially wide of the mark. Clone the boss’s voice because he is the boss, after all? Or the other way round: may the company clone an employee’s voice because she is employed? No. The employment relationship does not suspend the right of personality. Anyone who wants to clone a voice in the company needs the voluntary, specific agreement of the person concerned, not a ruling from above.
The voice is also personal data
Second level: data protection. A voice recording is personal data. Processing it in order to build a voice model from it is data processing under the GDPR. It needs a legal basis, a clearly defined purpose and transparency towards the data subject. Anyone who secretly trains a voice model from old recordings already violates these basic rules. And unlike a single photo, a voice once trained can hardly be recalled once the model is out in the world.
There is one more level. If the voice is processed in such a way that it uniquely identifies a person, for example as a voice profile for authentication on the telephone, that is biometric data under Article 9 GDPR. That is the special category with the highest protection. Stricter conditions apply to it, and explicit consent is usually the only viable route. So anyone who clones a voice quickly finds themselves in the most sensitive area of data protection, not at its edge.
When the clone goes out: the labelling obligation
Let us assume consent has been properly obtained and data protection has been clarified. Then a third obligation is still open. As soon as artificially generated or manipulated audio content goes into circulation, the transparency obligations under Article 50 of the EU AI Act apply, which are in force from 2 August 2026. Such content must be recognisable as artificially generated or altered. Listeners have a right to know that they are not hearing a human, but a machine that sounds like a human.
The order matters, and it is often turned around. Labelling does not replace consent. A label under a voice clone does not cure missing permission. I have described in detail elsewhere how you label AI content in concrete terms. Here the principle counts: labelling is mandatory, but it is the last stage, not the first. First the permission, then data protection, then the labelling.
A label under a voice clone does not turn missing consent into existing consent. Labelling is mandatory, but it is the last stage, not the first.
The misuse scenario: the boss’s voice on the telephone
There is a tangible reason why I am so blunt on this topic. Cloned voices have long been a tool for fraud. The pattern is well known: an employee in accounts receives a call, on the line the familiar voice of the managing director, urgent, a transfer has to go out immediately. Only it is his cloned sound. Anyone who pays on the strength of such a call falls for a fraud, punishable under Section 263 of the German Criminal Code (Strafgesetzbuch, StGB).
The sober consequence for every company: a voice is no longer an identity document. It no longer serves as proof that the person at the other end really is the one you believe you are hearing. Anyone who ties approvals for payments or data to a familiar voice alone has an open gate. Critical instructions need a second, independent channel for confirmation, not the gut feeling that the voice does sound real.
The same honesty applies in the other direction, towards the customer. A voice system for customer contact, for example on the hotline or in a voice assistant, must be built in such a way that people learn that a machine is speaking and not a human. Under Article 50(1) of the EU AI Act this obligation falls on the provider. Whoever deploys the system does not switch the notice off. People must not be left believing that they are speaking to a real person when a system sits at the other end.
The simplest way out: a voice that belongs to nobody
With all of this, a simple question that many skip is worth asking: does it have to be the voice of a real person at all? There are two very different things that are often lumped together. One is the cloned voice, the replica of a specific, real person. The other is a synthetic voice, an artificial voice that is not modelled on any specific person. It belongs to nobody, so it does not touch any right of personality and needs no consent from a person concerned.
For many purposes in a company exactly that is the clean route. The announcement in the telephone queue, the narrator’s voice in e-learning, the read-aloud function on the website: this does not need a cloned colleague, a neutral synthetic voice does just as well. It spares you the whole question of consent and personality rights. All that remains is the labelling that it is a machine, and that is easy to fulfil. Anyone who clones a real voice should do so deliberately, because exactly this person is wanted, and not out of embarrassment because the tool happens to offer it.
I therefore advise companies to adopt a clear rule: the synthetic voice is the default. The clone of a real person is the exception, which needs documented consent and a good reason. That way you turn the burden of proof in your own house the right way round, away from convenience and towards a conscious decision.
What ISO 42001 requires of a voice cloning project
Anyone who wants to use voice clones responsibly finds the right framework in ISO/IEC 42001, the standard for AI management systems. It requires a clearly defined, legitimate purpose, a documented impact assessment for systems with effects on people, information for users and affected persons, and objectives for responsible use, among which the standard counts human oversight. Translated to a voice clone, that means:
- Purpose and legal basis are documented: which voice is cloned for what, and on which permission does that rest?
- Consent has been obtained, is demonstrable and can be withdrawn, specific rather than blanket.
- Outputs are labelled as artificially generated wherever they go into circulation.
- There is a safeguard against misuse: no approvals on the basis of voice alone, a second channel for critical actions.
- A human retains oversight of what is published with the cloned voice.
An auditor does not need ten minutes to see whether this is in place. He asks about the consent, the purpose, the labelling and the safeguard against misuse. If one of these is missing, the finding is clear. And in practice several are usually missing at once, because the tool is bought first and the thinking only starts afterwards.
My clear position: cloning a voice just because it is possible and convenient is not progress, but a violation you can see coming. With consent, a clean purpose, labelling and a safeguard against misuse, a voice clone can be a legitimate tool, for accessibility, for localisation, for efficiency. Without that it is a breach of trust that quickly becomes a criminal offence. A person’s voice belongs to them, not to whoever happens to have a few recordings and an AI tool.
Are you planning to use AI voices or other generative AI in your company? Check in a few minutes whether consent, labelling and controls are in place before the first clone goes out.
To the AI readiness checkPrimary sources
- EU AI Act, Regulation (EU) 2024/1689, Article 50 (transparency obligations)
- Article 9 GDPR (special categories, biometric data)
- General right of personality, Article 2(1) in conjunction with Article 1(1) of the German Basic Law
- Section 263 of the German Criminal Code (fraud)
- ISO/IEC 42001:2023 AI management systems
Frequently asked questions
May I clone the voice of an employee or the boss if it benefits the company?+
Only with that person’s voluntary, informed consent. The employment relationship alone is not enough. The right of personality in one’s own voice stays with the person. A blanket clause in the employment contract is as a rule not sufficient.
Is it enough simply to label the AI voice clone?+
No. Labelling under Article 50 of the EU AI Act is mandatory as soon as the content goes into circulation, but it replaces neither consent nor the legal basis under data protection law. Labelling is the last stage, not the first.
Is a cloned voice biometric data?+
The voice recording is initially ordinary personal data. If the voice is processed in such a way that it uniquely identifies a person, for example for authentication, it is biometric data under Article 9 GDPR, the special category with the highest protection.
How do I protect my company against fraud with cloned voices?+
Stop treating a voice as proof of identity. Critical approvals for payments or data need a second, independent confirmation channel. Train the team on the pattern of the urgent call with the familiar voice.
When is a voice clone permissible at all?+
When consent, a clear purpose and legal basis, labelling of the outputs and a safeguard against misuse come together. Then it can be useful for accessibility, localisation or efficiency.
Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)
Last updated: 10 November 2026. Researched and reviewed to the best of our knowledge; not a substitute for individual legal advice.
Questions about your own case?
In a free 15-minute intro call we assess where you stand on ISO 42001, ISO 27001 and the EU AI Act, honestly and without a sales pitch.