Effective Date: October 2026
Whether it’s a multilingual explanatory video, an e-learning module, or a phone announcement, AI voices are increasingly being used for the audio track. This approach is fast, scalable, and cost-effective. As of August 2, 2026, the transparency requirements under Article 50 of the AI Regulation (Regulation (EU) 2024/1689), better known as the EU AI Act, have been largely in effect.
For companies, the most important factors here are the type of AI voice used, who is responsible for its use, and what transparency requirements apply in each specific case.
Key Points at a Glance
- Article 50 of the AI Regulation distinguishes between machine-readable labeling by providers and the disclosure of certain AI content to humans by operators.
- When it comes to voice-overs, it is particularly important to determine whether an audio clip qualifies as a deepfake.
- Whether disclosure is necessary depends on whether the voice is perceived as that of a real or plausibly existing person—not on how it was technically generated.
- In hybrid voice productions, original human recordings are combined with AI-generated versions of the same voice.
- For certain existing AI systems, a transition period for machine-readable labeling extends through December 2, 2026. According to the European Commission’s interpretation, there is no retroactive labeling requirement for content created before August 2, 2026.
This article is not a substitute for individual legal advice; rather, it is intended to provide initial guidance.
Article 50 of the AI Regulation: Which Rules Apply to Voice-Overs
The AI Regulation has been in effect since August 1, 2024, and will be implemented in phases. It follows a risk-based approach. For AI-generated voices and voice-overs, the transparency requirements set forth in Article 50 of the AI Regulation on EUR-Lex are particularly relevant.
Four control areas are particularly important for audio:
- direct interactions with AI systems,
- machine-readable labeling of generative outputs,
- Disclosure of deepfakes,
- Requirements Regarding the Format of a Required Notice.
Important: The machine-readable labeling required under Article 50(2) and the disclosure to the public required under Article 50(4) are two distinct obligations. Therefore, the fact that a voice was generated using AI does not automatically imply an obligation to disclose this to the public.
Providers and Operators:
Who Is Responsible for What?
The AI Regulation distinguishes between providers and operators.
A provider is a natural person, legal entity, or public authority that develops an AI system or has it developed and places it on the market or puts it into service under its own name. This may include, for example, providers of a text-to-speech platform or a voice-cloning tool.
An operator is a natural person, legal entity, or public authority that uses an AI system professionally under its own responsibility. This could be, for example, a company that produces and publishes audio tracks using an AI voice.
The transparency requirements set forth in Article 50 are based on this division of roles.
Transparency Requirements Relevant to Voice-Overs
Transparency Requirements Relevant to Voice-Overs
Art. 50, para. 1: Notice Regarding Direct Interaction with AI
When people interact directly with an AI system, this must generally be recognizable, unless it is already obvious from the context.
This applies, for example, to interactive voicebots, AI voice assistants, or conversational AI avatars, provided that the system in question is intended for direct interaction with natural persons.
A pre-recorded phone message or a pre-recorded voice-over, on the other hand, does not constitute an interaction with the AI system.
Art. 50, para. 2: Machine-readable marking
Providers of generative AI systems must label their output as artificially generated or manipulated in a machine-readable format. Synthetic audio content is explicitly covered.
The methods used should be effective, interoperable, robust, and reliable. Examples include technical metadata or other machine-readable tags.
Art. 50, para. 4: Disclosure of Deepfakes
Operators must disclose deepfakes. In addition, paragraph 4 contains specific requirements for certain AI-generated texts intended to inform the public about matters of public interest.
When it comes to voice-overs, the most important factor is whether the specific audio clip falls under the definition of a deepfake.
Art. 50, para. 5: Form of the Information
Where Article 50 requires that a notice be provided to individuals, it must be clear and unambiguous, provided no later than the first interaction or exposure, and in compliance with accessibility requirements.
The section on labeling practices demonstrates how this can be implemented in practice for audio or video.
When is AI-generated audio considered a deepfake?
The definition of “deepfake” in Article 3(60) of the AI Regulation expressly includes audio content as well.
According to the European Commission’s interpretation in its final guidelines on Article 50 (adopted on July 20, 2026), three criteria must be met:
- The content bears a resemblance to an actual person, object, place, organization, or event.
- What is simulated actually exists or could plausibly exist.
- The content is likely to be mistakenly perceived as authentic or truthful.
Therefore, technical similarity is not the only factor relevant to the legal classification. The specific context of use also plays a role: what matters is how the audience interprets the recording and whether it can be perceived as an authentic depiction or expression.
Actual intent to deceive is not required for classification. Rather, the decisive factor is whether the content, based on its presentation and context, would mistakenly appear to be authentic or truthful.
The criterion of “matter of public interest,” however, is not part of the definition of “deepfake.” It pertains to the separate rules governing certain AI-generated texts.
Which Vote, Which Duty? Three Typical Scenarios
In practice, AI voice-overs can be divided into three typical groups.
1. Voice clone of a real person
In voice cloning, an AI generates a voice that closely resembles and is identifiable as that of a real speaker voice talent .
If the recording can be understood as an authentic expression of the real person based on the criteria described above, it may be classified as an audio deepfake.
Typical examples include:
- a commercial featuring a cloned celebrity voice,
- a corporate video featuring the AI voice of the executive team,
- an e-learning course featuring the cloned voice of a trainer,
- A podcast featuring a computer-generated version of a real person's voice.
In addition to labeling requirements, issues such as consent, rights of use, and the right of publicity regarding the voice must also be reviewed on a regular basis for such productions.
2. Fully synthetic voice
A fully synthetic voice is created without a specific human model.
A deepfake can exist even without a specific real-life model: According to the Commission’s final guidelines, the term “person” also includes realistic AI personas and voices that could plausibly exist. The decisive factor is whether the voice appears to be that of a real person or can be perceived as such. A neutral voice that does not represent a person is typically not a deepfake within the meaning of Article 3(60) of the AI Regulation. Examples include:
- a neutral, off-screen AI voice in a corporate film,
- a synthetic voice in an internal learning module,
- A static, pre-recorded phone message.
For such content, technical labeling by the provider may still be relevant. A deepfake disclosure under Article 50(4), however, is typically not required. The situation may be different for synthetic presenters, testimonials, or interviews that can be perceived as real people.
3. Hybrid Voice: Original recording and AI version of the same voice
In the hybrid voice model, real studio recordings are combined with AI-generated versions of the same voice, for example, through our TTS platform, mainvox. The voice artist voice talent has consented to this use, and the rights are governed by contract.
For example, this makes it possible to update text passages, create additional language versions, produce variations, or expand existing productions.
What makes this unique is that the same vocal sound can exist both as an original human recording and as an AI-generated version. To listeners, the two versions may be barely distinguishable—or impossible to distinguish—acoustically under certain circumstances.
If the AI-generated version is based on the voice of a real speaker, it must be evaluated using the deepfake criteria mentioned above.
We explain how AI voices are created and used to mimic human voices in a separate overview.
Decision Table: Typical Voice-Over Scenarios
The following table summarizes common use cases. It is intended as a general guide; other factors may be relevant in specific cases.
Scenario | Classification | Disclosure Requirements |
E-learning for internal training, fully synthetic voice | Fully Synthetic Voice | This must be determined on a case-by-case basis. Generally, no. |
Public explanatory video produced by a government agency featuring the cloned voice of a real person | Voice Clone | In principle, yes, provided it qualifies as a deepfake |
Audiobook of a fictional novel featuring the cloned voice of a real person | Voice Clone | Disclosure is generally required; for fictional works, the special rule under Art. 50(4) applies. |
TV commercial featuring the cloned voice of a real person | Voice Clone | In principle, yes, provided it qualifies as a deepfake |
Corporate film with a fully synthesized voice-over | Fully Synthetic Voice | This must be determined on a case-by-case basis. Generally, no. |
Podcast Featuring a Cloned Version of the Executive Team's Voices | Voice Clone | In principle, yes, provided it qualifies as a deepfake; consent must also be clarified |
Static phone greeting with a synthetic voice | Fully synthetic, no interaction | Usually not |
Interactive Voicebot in Customer Service | Interactive System | Depends on the specific case; check Art. 50, para. 1 in particular |
E-Learning in the Hybrid Voice Model | Original recording plus AI version | On a case-by-case basis; check each file individually |
In the case of fully synthetic voices, a voluntary transparency notice may be advisable, particularly if the voice is presented as a real person (presenter, testimonial, interview). In this case, there may also be a disclosure requirement under Article 50(4).
What required labeling might look like in practice
In its FAQs on transparency requirements, the European Commission lists visual and audible notices as possible forms.
Audible Alert
For audio-only content, the announcement can be spoken at the beginning, for example:
"The following phone message was generated using an AI voice."or "The following message was generated by a machine."
Even a short voice tag can generally be considered.
Visual cue
For videos or digital players, this information can appear, for example, as “AI-generated audio,” as an on-screen text at the beginning, or as a clearly visible notice within the player.
A notice placed exclusively in the legal notice or in the terms and conditions generally cannot be considered a substitute for the required immediate and conspicuous disclosure. The same may apply to a notice that appears only at the end of the content. According to the Commission’s guidelines, a notice must be conspicuous in practice; for longer videos, a single notice at the beginning may not be sufficient.
EU Icons for AI-Generated Content
The European Commission provides freely usable symbols such as “AI Generated” for content generated entirely by AI and “AI Modified” for content partially modified by AI.
The use of the EU icons is voluntary. However, the icons alone do not automatically satisfy the legal disclosure requirement; the key factor remains that the notice must meet the requirements of Article 50, paragraphs 4 and 5.
Five Steps to Auditing a Single Production Run
- Track AI usage: Which audio files were generated entirely or partially by AI?
- Clarify roles: Who acts as the provider or operator in this specific setup?
- Classify the scenario: Is the voice perceived as a real person, a plausible character, or a neutral off-screen voice? For hybrid voices: Clarify the role of the voice for each file.
- Determining Obligations: Does an interaction, labeling, or disclosure obligation apply?
- Ensure Documentation: How were the production method and labeling status documented?
Exceptions and Special Circumstances
Artistic and fictional works
If a deepfake is part of a work that is clearly artistic, creative, satirical, fictional, or analog, Article 50 provides for a less stringent disclosure requirement.
The presence of such artificial or manipulated content must be indicated in a way that does not impair the presentation or enjoyment of the work.
This may be relevant, for example, for audiobooks, radio plays, games, or other fictional audio formats. It may also apply to podcasts or artistic TV commercials.
This special rule is particularly relevant when an artistic or fictional work itself contains a deepfake—such as an artificially generated voice that realistically mimics a real person.
AI-generated texts
For certain published AI-generated texts on matters of public interest, disclosure may not be required if a human review or editorial oversight has taken place and a person bears editorial responsibility.
Simply correcting spelling or grammar is not enough.
This rule applies to AI-generated text and does not automatically change the transparency requirements for synthetic or manipulated audio content in the same project.
Standard Processing
The obligation under Article 50(2) does not apply to the extent that the AI system merely performs a supporting function in standard processing or does not significantly alter the input data provided by the operator or the semantics of such data.
Whether, for example, AI-powered noise reduction falls into this category depends on the specific application.
Transition Period and Lack of Retroactive Effect
For AI systems placed on the market before August 2, 2026, a transitional provision applies with regard to machine-readable labeling (Regulation (EU) 2026/1744, “Digital Omnibus”).
The requirements set forth in Article 50(2) must be met for these systems only as of December 2, 2026.
In addition, according to the European Commission’s interpretation, there is no retroactive labeling requirement under the transparency rules that will take effect later for content created before August 2, 2026 (for audio, images, and video, the date of creation is decisive).
The Commission recommends voluntary relabeling, to the extent that this is practicable.
Code of Conduct on Transparency for AI-Generated Content
Code of Conduct on Transparency for AI-Generated Content
The EU Code of Conduct on Transparency for AI-Generated Content was officially published on June 10, 2026.
The European Commission and the AI Committee have identified it as a tool to support the practical implementation of transparency requirements. By the end of July 2026, approximately 190 organizations and companies had signed it; this figure reflects the status at that time and may have changed since then.
The code is voluntary and does not replace legal requirements. However, it can provide concrete guidance on their practical implementation.
For audio deepfakes, he recommends, among other things:
- a brief and clear announcement at the beginning,
- if a display is available, an additional visual indicator,
- For longer content, include repeated reminders as needed.
Anyone who does not apply the Code must be able to demonstrate compliance with the legal requirements by other appropriate means.
Agency or company: Who is responsible for the project?
In many voice-over productions, clients, agencies, voice talent , and technical platforms work together. It is therefore crucial how tasks and oversight are distributed within a specific project.
Depending on its setup, an agency can act as the operator itself if it uses AI systems on its own responsibility and integrates their results into production.
Employees generally act on behalf of their company. Freelancers or external service providers may also be considered part of the company, depending on the nature of their involvement.
The contract should specify who:
- documents the use of AI,
- checks the required information,
- implements labeling technically,
- releases the final version.
We address additional questions regarding the legally compliant use of AI in voice-over production in a separate section.
Sanctions and Oversight
Market surveillance is the responsibility of national authorities.
In Germany, this has been governed since July 29, 2026, by the Act on Market Surveillance and the Promotion of Innovation in Artificial Intelligence (AI-MIG). It assigns the Federal Network Agency a central role in coordination, serving as a point of contact, and handling complaints. Other specialized agencies remain responsible for certain areas.
In the event of violations of transparency obligations , Article 99(4) of the AI Regulation provides for fines of up to 15 million euros or 3 percent of global annual revenue, whichever is higher.
For small and medium-sized enterprises, including startups, the lower of the two amounts applies pursuant to Article 99(6).
The actual amount of a penalty depends on the circumstances of each individual case. The amounts listed are maximum limits and are not fines that are automatically imposed.
Other Areas of Law Related to AI Voices
The transparency requirements of the AI Act are only one part of the legal review. Voice cloning and AI-generated voice synthesis may affect other areas of law.
GDPR
To the extent that voice recordings or other information processed in connection with speech production constitute personal data, an appropriate legal basis under the GDPR is required for their processing. Whether the specific requirements for biometric data also apply depends, in particular, on the specific purpose of the processing.
Consent and Rights of Use
When using a real speaker’s voice for voice cloning, AI training, or synthetic speech generation, the necessary rights of use and data protection requirements must be reviewed and clearly defined in a contract. Depending on the specific processing activity, consent may be required or advisable; however, it is not always the only possible legal basis.
In addition, scripts, music, and voice-over performances may be protected by copyright or neighboring rights. We provide an overview under “Usage Rights” and “Buyouts.”
Right to Privacy Regarding One's Voice
Unauthorized voice clones may infringe on personal rights. Therefore, a transparency notice alone does not constitute authorization to clone or use another person's voice.
Competition Law
In the case of advertising and marketing audio, competition law requirements may also apply. For example, if a synthetic testimonial is presented as a genuine statement, this may raise questions about potential misleading practices.
What Companies Should Do to Prepare Organizationally
In addition to reviewing individual productions, companies should establish permanent structures for managing AI voices.
- Track AI workflows. Document the areas where AI voices are used, the systems involved, and the channels through which content is published.
- Define responsibilities. Determine who approves AI systems, coordinates legal matters, and is responsible for publications.
- Develop a labeling standard. Define consistent text modules, symbols, and placement rules for audio, video, and interactive formats.
- Amend contracts. Work with agencies, voice actors, voice talents , and service providers to address labeling, data protection, consents, and usage rights.
- Establish a tracking system. Ensure that the production type, tool used, language version, and labeling status can be traced later. For hybrid voice productions, it should be particularly clear which version is an original human recording and which is AI-generated.
- Train teams. Marketing, L&D, communications, and production teams should be familiar with the requirements and internal processes relevant to their work.
- Involve production partners. Service providers that offer both real voice talent and AI voices can structure production processes accordingly. If you book professional voice talent for specific formats, these recordings themselves are not AI-generated content.
Common Misconceptions
"All AI-generated content must include a visible disclosure."
No. The transparency requirements that apply depend on the specific content and its use.
“The AI provider’s watermark is always sufficient.”
No. A technical label does not automatically equate to a required disclosure to people.
“A notice in the legal notice section is sufficient.”
Not if immediate disclosure is required. In that case, the notice must be visible when the content is accessed.
“AI voices are generally problematic under the AI Act.”
No. AI voices are not generally prohibited. The requirements that apply depend on their specific use.
Conclusion
There is no standard labeling solution for AI-generated voice-overs. The key factors are the production format, the legal role in the project, and the specific use.
Companies should integrate this review into their production and approval processes rather than conducting it only immediately before publication.
We help companies align their branding, corporate voice identity, and production processes. If you’d like to determine which branding approach makes the most sense for your projects, please contact our team.
Frequently Asked Questions About AI Labeling and Voice-Overs
A fully synthetic voice that has no real human model and that does not appear as a real person or is not perceived as such is typically not subject to the relevant disclosure requirement under Article 50(4) of the AI Regulation. Disclosure to the public is therefore generally not required.
Internal professional training modules may also fall within the scope of the AI Regulation. Whether a specific transparency requirement applies depends, in particular, on the format in which the content is produced and the manner in which it is used.
The law does not specify a fixed duration. According to the Commission’s guidelines, the notice must be visible in practice. The voluntary code of conduct recommends that, for longer formats, the notice be repeated as necessary.
According to the European Commission’s interpretation, there is no retroactive labeling requirement for content created before August 2, 2026 (the date of creation is decisive).
The practical implementation may be delegated by contract. However, this does not automatically change the legal division of roles in the specific project.
