In short

Anyone cloning AI voices needs more than an AI label: consent, GDPR and the EU AI Act converge. This guide shows creators how to document consent, clarify the chain of rights, and embed governance in everyday production.

Table of contents

• Why consent is mandatory in voice cloning

• The voice as data: what the GDPR requires

• EU AI Act: labelling obligations from August 2026

• Clarifying the chain of rights: who must give consent

• International framework: ELVIS Act and NO FAKES Act

• Documenting consent: contracts, revocation, deletion

• Governance in practice: a checklist for creator teams

• Frequently asked questions

• Sources

Key takeaways

Berlin Regional Court, ruling of 20 August 2025: an AI voice clone without consent obliges the user to pay a fictitious licence fee per video.

From 2 August 2026, the transparency obligations of the EU AI Act also apply to synthetic audio content.

Biometric voice data falls under Art. 9 GDPR: processing generally only with explicit consent.

Consent and labelling are separate obligations: an AI label does not replace the speaker's consent.

The ELVIS Act (Tennessee) and the NO FAKES Act (S.1367) tighten voice protection, including for creators with a US audience.

Voice Cloning: Consent and Governance for Creators

In short

Voice cloning without consent is expensive: the Berlin Regional Court awarded a voice actor EUR 4,000, and from 2 August 2026 the EU AI Act will require AI-generated voices to be labelled. This guide shows creators the applicable rules and a clean consent workflow.

Why voice cloning is now a governance issue

Voice cloning has long been part of everyday creator production: with just a few minutes of speech material, it is now possible to create a deceptively real voice imitation, and with every advance in the models it becomes harder to distinguish the original from the replica.[1] It is precisely this quality that makes the technology so valuable for podcast teams, course providers and studios, because a recorded voice can be reused across many languages and formats. But it also makes it a target for attack: the same technology is driving a wave of fraud that hits consumers and brands alike.

A look at the USA shows the scale: in 2024 alone, US consumers lost around 2.95 billion US dollars to imposter scams – fraud schemes in which perpetrators pose as familiar people, authorities or companies.[2] AI-generated voice clones are a key accelerant. Consumer Reports therefore submitted a petition to the FTC with more than 75,000 signatures calling for providers of voice-cloning products without effective safeguards to be held accountable.[3]

Why consent is mandatory in voice cloning

On the other side stand the economic interests of those whose voice it is: voice actors, musicians, actors and podcasters live off their recognizability, and unauthorized voice clones can significantly undermine this livelihood.[1] Legally, a voice is not free material but personal data and an element of personality at the same time.

Anyone who uses an AI voice that resembles a real person needs that person's permission. This was clarified by the Berlin Regional Court in its ruling of 20 August 2025 (case no. 2 O 202/24): a YouTuber had the unmistakable voice of voice actor Manfred Lehmann imitated by AI and used it to dub two videos on his channel without asking the actor. The court saw this as an infringement of the right to one's own voice, which is part of the general right of personality, and ordered him to pay a fictitious license fee.[4]

The thesis of this article is therefore: consent and documentation are not bureaucratic overhead, but the prerequisite for multilingual production to remain scalable at all. The legal situation has narrowed dramatically in a short time, a German court has for the first time imposed a fictitious license fee for an AI voice clone, and from August 2026 the EU AI Act will make the labeling of synthetic voices mandatory. Those who translate these obligations into a fixed workflow can produce multilingual formats with a clear conscience. Those who do not risk cease-and-desist letters, fines and reputational damage that wipe out any efficiency gains.

Recognizability is enough, no original needed

The crucial point for creators: the court did not base its decision on the original recording, but on recognizability. The voice in the videos did not come from Lehmann but was an AI imitation, and that is precisely what did not protect the YouTuber. In the court's view, it made no difference to the assessment that it was not "the" voice of Lehmann but an AI-generated imitation. Because, given the similarity, the audience could assume that Lehmann had spoken the text himself, this constituted an infringement of his general right of personality.[5]

Legal situation in Germany: the right to one's own voice

Three factors shaped the ruling and give creator teams a practical line of review:

In Germany, the voice is protected by the general right of personality, which derives from Art. 1(1) and Art. 2(1) of the Basic Law in conjunction with § 823(1) of the German Civil Code (BGB). The Federal Court of Justice expressly recognized this right to one's own voice as early as the so-called Marlene Dietrich ruling (case no. I ZR 49/97): the voice is a distinguishing feature of a person, just like the image. The general right of personality protects in both directions, both against interference with personal honor, for example through an imitation that contradicts one's own values, and against unauthorized economic exploitation.[1][1]

Assignability: what matters is whether a not insignificant part of the audience can attribute the voice to a specific person, not whether the original recording was used.[5]

Commercial context: the YouTuber used the voice to increase the reach of his channel and a linked online shop. This economic connection significantly sharpened the balancing of interests.

Freedom of art and expression does not protect everything: the videos had satirical content, but a use of another person's identity features that is predominantly economically motivated does not fall under these freedoms.[4]

In addition, the GDPR applies: voices are not freely available data but are subject to both personality rights and data protection law. For creator teams this means concretely: before any use of a voice, whether for training an AI model or for publishing voice imitations, consent is required that is specific to the respective purpose and remains revocable at any time. Without this consent, severe fines loom, and affected individuals can demand cessation, deletion and information about the processing.[1]

For podcasters and course providers this means: even the use of an AI voice that merely sounds similar can be unlawful if the person is recognizable and has not consented. Consent is therefore not a formality but the fundamental prerequisite of every voice cloning project.

How sharply this right is now being enforced is shown by the amount of damages: the court set the fictitious licence fee at 2,000 euros per video, 4,000 euros in total, and additionally ordered the channel operator to reimburse pre-litigation legal costs.[4]

The voice as data: what the GDPR requires

In addition to personality rights, data protection law applies – and it kicks in earlier than many think. A voice recording is first and foremost personal data, because it can be attributed to a natural person through sound, tone and manner of speaking. What matters is the technology: if the process used extracts unique voice characteristics, for example to create a voiceprint and synthesise the original voice, the data generally becomes biometric data within the meaning of Art. 4 No. 14, Art. 9 GDPR.[6] Systems designed to imitate a specific voice practically always work with exactly such feature extractions.

Legal basis	What it protects	Consequence for creators
General personality right (Art. 1, 2 GG in conjunction with § 823 BGB)	The voice as a recognisable personality trait, recognised since the Marlene Dietrich ruling of the BGH	No use of a recognisable voice without consent, not even as an AI reproduction
GDPR	The voice as personal data	Purpose-bound, revocable at any time, explicit consent before processing
Berlin Regional Court, judgment of 20.08.2025 (case no. 2 O 202/24)	Protection also against AI-generated imitations where the audience is confused about attribution	Fictitious licence fee of €2,000 per video, here €4,000 in total

Biometric data are subject to a general prohibition on processing, which can only be overcome under the narrow exceptions of Art. 9(2) GDPR. For voice cloning projects, in practice only one route is viable:

What the Berlin Regional Court ruling means for creators in concrete terms

Explicit consent: tacit approval or a legitimate interest is not sufficient for biometric voice data; consent must be given freely, informed and explicitly.

Data protection impact assessment: in most voice cloning cases, an impact assessment under Art. 35 GDPR is required, because processing biometric data poses a high risk to the individuals concerned.

Information obligations: data subjects must know for what purpose their voice is processed, how long voiceprints are stored and who receives the data.

When in doubt, treat it as biometric: if it cannot be reliably determined whether the technical process extracts unique identification features, controllers should take the legally safe route and assume biometric status.

The ruling is more than an isolated case; it works as a practical guide because the court worked through every line of defence of the YouTuber one by one. The first point concerns satire and artistic freedom: while the court did recognise satirical content about the German federal government in the videos, which in principle falls under the protection of freedom of expression and art, the defendant still lost the balancing exercise because the well-known voice primarily served as bait: it was meant to make the videos more attractive, boost clicks and revenue, and at the end of the videos the operator pointed viewers to his online shop. A satirical engagement with the speaker's person or voice itself simply did not take place.

Anyone working with external service providers should examine the processes used in detail and coordinate with their own data protection officer. The higher protection status is not a bureaucratic obstacle: voices are immutable characteristics, and misuse, for example through deepfakes, can have serious consequences for the person affected.

The second point is particularly important for teams working with AI tools: the licence of the AI platform does not convey any rights to the voice. The defendant had argued that he had paid for the software and had therefore acquired a right of use to the generated voice. The court rejected this: since the speaker had never consented to the creation and distribution of his voice on the AI platform, the buyer could not acquire a derived right of use. Only the person themselves can release rights to their voice.

EU AI Act: labelling obligations from August 2026

The third point concerns commercial use: as soon as an imitated voice serves, even indirectly, to increase reach, clicks or sales – for example in product videos or on channels with shop links – the use is commercial, and courts then tend to rule strictly in favour of personality rights. The court additionally considered the interference particularly serious because viewers could have gained the impression that the speaker identified with the defendant's content and products.

Consent and labelling are two separate obligations. Even those who have the speaker's valid consent must identify synthetic content as such. The transparency obligations of the EU AI Act (Article 50) apply from 2 August 2026 and are addressed to two different addressees:[8][9]

The practical consequence can be summed up in a single test question: before every use of a realistic AI voice, the team should clarify whether the audience will attribute it to a real person. If the answer is yes, or even just maybe, that person's consent is required, or the use is off the table. When in doubt, it is advisable to make it clearly recognisable that the voice is artificial, or to refrain entirely from imitating real people.

Role	Obligation under Art. 50 AI Act	Relevance to everyday creator work
Providers	Must label synthetic audio, image, video and text content in a machine-readable format so that it is recognisable as artificially generated or manipulated (Art. 50(2))	If a team runs its own AI voice solution, it is the provider and must mark the outputs in a machine-readable way
Deployers	Must clearly and recognisably disclose deepfakes when publishing the content; the provider's machine-readable marking is not sufficient for this (Art. 50(4))	Creators, networks and studios that publish AI voices must declare the synthesis visibly to the audience

Deepfake is defined narrowly here: it is an AI-generated or AI-manipulated image, audio or video content that resembles real persons, objects, places, facilities or events and would falsely appear to a person as authentic or truthful.[8] Three elements must therefore come together: the resemblance, the real-world reference and the false impression of authenticity. A cloned speaker voice in a podcast generally meets these criteria.[9]

Regardless of the deadline, the platforms' rules already apply today. YouTube requires disclosure of realistic-looking, altered or synthetic content in the upload process, TikTok has its own AI toggle, and Meta requires a declaration for realistic-sounding AI audio.[10][11] Anyone who fails to label risks platform sanctions up to content removal, quite independently of the statutory obligation.

Clarify the chain of rights: who all needs to consent

Providers must mark the outputs of their AI systems in machine-readable form and make them recognisable as artificially generated or altered (Art. 50(2))[12]

Deployers, i.e. the brands, agencies and creators using the technology, must disclose deepfake audio as artificially generated, audibly or visibly, at the audience's first perception (Art.[12] 50(4))

The disclosure must be clear and distinguishable, at the latest at the moment of first exposure, not in a linked description[13]

Artistic, satirical or fictional content is subject to a milder form of labelling that does not impair the use of the work[12]

The speaker's consent is necessary but often not sufficient. Several protective rights converge on a voice recording, and overlooking even one of them makes the entire project vulnerable. For professional productions such as audiobooks, dubbing, advertising spots or multilingual podcast releases, this means: the chain of rights must be fully clarified before the first training run starts.

The scope of the disclosure obligation is important here: it expressly applies even when the creator has consented to the voice cloning.[13] A hidden, machine-readable watermark in the audio file does not replace the audible disclosure either, because that marking is aimed at systems, not at listeners.

Performer's related rights: Professional speakers, actors and voice-over artists hold rights in their performance and decide whether and how the recording is used.

Producer's phonogram rights: Recording studios and producers own their own protective rights in the recordings. Even if the speaker consents, use may be impermissible without the producer's consent; all rights holders involved must be included.

AI voice replication as an adaptation: When an imitated or altered voice is generated from the recording using AI, this often constitutes an adaptation in legal terms. The alteration does not release you from the obligation to obtain the rights.

Text and data mining under Section 44b of the German Copyright Act (UrhG): Reproductions of lawfully accessible works are permitted; a rights holder's reservation of use is effective for works available online if made in machine-readable form, and the reproductions must be deleted when they are no longer needed for text and data mining.[14]

Post-mortem use: Personality protection does not end abruptly with death. In commercial use of a deceased person's voice, the heirs generally have a say.

For implementation, the European Commission has published a voluntary Code of Practice on transparency for AI-generated content, which around 190 companies and organisations had joined by the end of July 2026. Anyone following the Code can rely on it for compliance with labelling obligations and gains legal certainty in all member states.[15] The obligation itself exists independently of membership. Violations can be penalised with fines of up to 15 million euros or 3 percent of worldwide annual turnover, whichever amount is higher.[13]

Publicly accessible, by the way, does not mean freely available: The fact that a recording can be accessed as a podcast or on YouTube does not mean it may be used as training material. In practice, a simple rule pays off: no voice cloning project starts before it is documented in writing which rights holders have consented.

International framework: ELVIS Act and NO FAKES Act

Obtaining consent correctly: What a valid consent contains

Anyone localising content for a US audience, such as a podcast with an English version or a course for the American market, encounters additional regulations that go beyond the GDPR. The USA have no uniform federal law, but a rapidly growing body of state legislation and a nationwide draft bill, both of which expressly protect the voice.[16]

A consent is only as good as its wording. That consent must be given before use, be purpose-bound and remain revocable is described above; what matters here is the level of detail. A blanket consent is not enough: a general usage clause covering the recorded material does not cover voice cloning, because training an AI model on the voice and generating new, synthetic speech are a distinct use that deserves its own clause.[13]

Tennessee led the way: On 21 March 2024, Governor Bill Lee signed the Ensuring Likeness Voice and Image Security Act (ELVIS Act), the first US state law to expressly include the voice in the protection of the right of publicity. The law covers both the actual voice and its simulation, and thus targets the unauthorised AI replication of the voices of songwriters, performers and artists.[17]

Explicit permission: Voice cloning is expressly named, not derived from a general usage rights clause[13]

Purpose limitation: The specific languages, markets and platforms on which the clone may be used; consent to a German version does not automatically cover a Spanish advertising version[13]

Term: A time limit parallel to the usual usage rights periods rather than unlimited use[13]

Separate remuneration: A separate fee for the voice cloning, distinct from the payment for the original recording[13]

Revocation process: A named route for how consent can be revoked, and what then happens to existing content[13]

At the federal level, there is the NO FAKES Act of 2025 (S.1367, 119th Congress), introduced on 9 April 2025. The draft creates a nationwide intellectual property right in digital replicas of voice and likeness, which can be licensed by any person or rights holder and survives death. The right is complemented by a notice-and-takedown procedure against unauthorised digital replicas, similar to what is known in copyright law.[16]

The most important timing effect lies elsewhere: rights clearance should be settled before the recording, not only when the dubbing is due. Anyone collecting the voice for a multilingual production anyway is best off negotiating the cloning clause in the same breath as the recording itself, while both sides are still in conversation and compensation can be considered together.[13] Consents renegotiated after the fact are more expensive, slower, and considerably more legally vulnerable in a dispute. The more specific the wording, the less room there is for later conflict.

Build consent practice once: A documented consent process covering purpose, languages, channels, and retention period works across countries, instead of inventing new contracts for every market.

Think documentation along: Anyone who can prove when and for what a consent was given is prepared both against German and US claims.

License instead of gray zone: Both the ELVIS Act and the NO FAKES Act explicitly reward licensed use; rights holders can grant the voice contractually, which increases demand for clean consents.

Documenting consent: contracts, revocation, deletion

The obligation becomes a process when three building blocks come together: a precise consent text, clean contracts with service providers, and a deletion concept. The consent text should be concrete rather than blanket, because the more precisely purpose and scope are described, the more robust the consent is in a dispute.

Word the consent text concretely: Purpose, target languages, distribution channels, retention period, and right of revocation belong in the text and are documented, not agreed verbally.

Regulate data processing: Anyone using voice cloning service providers needs written data processing agreements that ensure GDPR-compliant handling of voice data.

Check international transfers: If the provider is based outside the EU, appropriate safeguards for the data transfer must be reviewed, in particular standard contractual clauses.

Define a deletion concept: Voice recordings, training material, and voiceprints must not be stored indefinitely, but deleted as soon as they are no longer necessary for the agreed purpose.

Revocation possible at any time: A consent can be revoked, so it must remain traceable which consents were given at what time and which content they cover.

For creator teams this means: the legal framework depends on the distribution market, and it is moving fast. Anyone planning a multilingual production today that will be available in the USA should keep an eye not only on German consent practice but also on the regulations of the target markets. The good news is that the requirements are moving in one direction: everywhere toward consent, labeling, and a documented chain of rights. A governance workflow covering these three elements is therefore future-proof not only in Germany but internationally.

Revocation in particular is underestimated in practice: a speaker can withdraw their consent, and then the team must be able to reconstruct which projects rest on which legal basis. Anyone who archives consents centrally and links them to project files can respond in an emergency within minutes instead of weeks. Such an approach pairs well with a structured localization playbook that anchors the approvals as a fixed step in the production process.[18]

Governance workflow for creator teams in practice

Governance in everyday work: a checklist for creator teams

From the previous sections, a process can be derived that any creator team, podcast publisher, and agency can map in four building blocks. The goal is not to slow down production but to make it plannable: every new language version then inherits the already documented chain of rights instead of renegotiating it each time.

Rules only help if they arrive in the production process. For podcasters and podcast networks, but equally for course providers and studios, this means: consent and labeling become fixed steps in the workflow, not a decision made per individual edit. Three anchors have proven effective:

Build a consent register: for every voice, document who spoke it, which purposes, languages, markets and platforms the consent covers, how long it runs and how it can be revoked[13]

Embed a labeling workflow into production: every AI-generated dubbing receives an audible notice at first exposure, such as a short spoken note at the beginning, plus a visible label where the format allows it[13]

Carry out vendor due diligence: before selecting a tool, check whether the vendor supports machine-readable marking and whether its license terms clarify that rights to the voice remain with the voice owner, because the software license does not replace consent

Define a decision framework: determine when an AI voice is used and when a human voice actor is booked, for example when a clear, audible AI label would break the format[13]

Risk analysis before every use: Can the voice be attributed to a real person? Is there a commercial connection? Is the chain of rights fully clarified? Is labeling planned? These four questions follow directly the guideline of the LG Berlin II ruling.

Define roles: the team defines who obtains consents, who reviews them legally and who archives them. Without clear responsibility, the documentation stalls.

Labeling as a project step: the consent status and the planned labeling are firmly anchored per project, for example in the localization workflow of a tool like ContentHub Studio, instead of leaving it to individual cases.

Building block	Core question	Concrete implementation
Consent register	What may this voice do, where and for how long?	One entry per voice with purpose, languages, platforms, duration and revocation path
Labeling	Does the audience know it is hearing AI?	Audible notice at first exposure, visible label in addition
Vendor due diligence	Who is liable, and whose rights apply?	Review of marking support and license terms before tool deployment
Decision framework	When is an AI voice the wrong path?	Book a human voice actor when labeling would break the format or consent is missing

The combination of risk analysis, roles and a fixed workflow step turns a legal obligation into a routine that scales as more languages and channels are added. Anyone rolling out voiceover in multiple languages already has many moving parts per project; that is exactly where the governance check belongs, not in a separate spreadsheet nobody opens. Our overview of AI voiceover in multiple languages shows how multilingual productions can be set up technically.

This is where the circle closes back to technology: anyone scaling multilingualism with voices, for example in a localization workspace like ContentHub Studio, needs this governance as a foundation so that consent and labeling grow with every language output. Such a tool can accelerate production in over 100 languages, but it replaces neither the documented consent of the voice owner nor the audible labeling that the EU AI Act will require from August 2026.[13] Teams that anchor both from the start can later produce every new language version from a clean basis instead of reconstructing rights afterwards. Those who skip documentation today will pay for it tomorrow as liability risk, as delay or as lost audience trust.

In the end, the effort pays off twice: consent protects against cease-and-desist letters and fictitious license fees, labeling protects against platform sanctions and fines, and together they protect the audience's trust. Creators who treat consent and labeling as two separate, documented obligations can use voice cloning without taking on a legal risk with every release.

Frequently asked questions

Is voice cloning without the speaker's consent legal?

As a rule, no. The Berlin Regional Court II ruled on 20 August 2025 (case no. 2 O 202/24) that even the AI-based reproduction of a recognizable voice infringes the right to one's own voice, even without the use of original recordings. The channel operator was ordered to pay 2,000 euros in fictitious license fees per video. In a commercial context, documented consent is practically mandatory.

From when must I label AI-generated voices under the EU AI Act?

The transparency obligations of the EU AI Regulation apply from 2 August 2026. Providers must then mark synthetic audio content in machine-readable form; deployers who publish such content must clearly and perceptibly disclose deepfakes upon first publication. A machine-readable marking alone does not fulfil the deployer's disclosure obligation. However, platform rules from YouTube, TikTok and Meta already require labels today.

Is the human voice a biometric datum under the GDPR?

Under certain conditions, yes. The voice is always personal data; if the technology used extracts unique characteristics (such as a voiceprint), it is considered biometric data under Art. 9 GDPR. In that case, processing is only permissible with explicit consent, and in most cases a data protection impact assessment is required. If in doubt, the biometric status should be assumed for legal certainty.

Is a note such as 'voice was generated with AI' in the credits sufficient?

No, because labelling and consent are two separate obligations. The Berlin Regional Court II made clear that a mere label does not release the use of another person's identity features. The note fulfils transparency requirements at best, but replaces neither the speaker's consent nor the clarification of the chain of rights from related rights and sound recording rights.

Who must consent in the chain of rights for a voice cloning project?

At minimum the speaker, via their related rights, and the producer or phonogram producer, via the rights to the recording. Since the AI reproduction is often legally considered an adaptation, even a modification does not release from the licence. For deceased persons, the heirs usually have a say in the commercial use of the voice.

What do the ELVIS Act and the NO FAKES Act regulate in the USA?

Tennessee signed the ELVIS Act on 21 March 2024, the first law to expressly include the voice in the protection of personality rights. The NO FAKES Act (S.1367) creates a nationwide licensable right to digital replicas of voice and likeness, including a notice-and-takedown procedure. Creators with a US audience should therefore build their consent practice with documentation from the outset.

What happens if a person withdraws their consent?

Consent can be revoked at any time. Upon revocation, the legal basis ceases to exist, so further processing of the voice recordings and the voiceprints derived from them must be stopped immediately, regardless of whether the agreed purpose is still ongoing. Purpose limitation is only the regular deletion trigger in ongoing operations. Therefore, every consent given should be documented with purpose, languages, channels and date in such a way that a revocation can be traced in the process and implemented immediately.

Sources

[1] oppenhoff.eu – https://www.oppenhoff.eu/de/news/detail/der-schutz-der-eigenen-stimme-in-zeiten-von-ki-voice-cloning/

[2] ftc.gov – https://www.ftc.gov/system/files/ftc_gov/pdf/csn-annual-data-book-2024.pdf

[3] advocacy.consumerreports.org – https://advocacy.consumerreports.org/press_release/more-than-75000-consumers-urge-ftc-to-crack-down-on-ai-voice-cloning-fraud/

[4] dr-bahr.com – https://www.dr-bahr.com/news/bekannter-synchronsprecher-bekommt-wegen-geklonter-ki-stimme-schadensersatz.html

[5] urheberrecht.org – https://www.urheberrecht.org/news/7818/

[6] dsn-group.de – https://www.dsn-group.de/datenschutz-notizen/die-stimme-als-biometrisches-datum-0659597

[7] oppenhoff.eu – https://www.oppenhoff.eu/en/news/detail/protecting-your-own-voice-in-times-of-ai-voice-cloning

[8] digital-strategy.ec.europa.eu – https://digital-strategy.ec.europa.eu/de/policies/code-practice-ai-generated-content

[9] artificialintelligenceact.eu – https://artificialintelligenceact.eu/de/article/50

[10] support.google.com – https://support.google.com/youtube/answer/14328491?hl=de

[11] humave.de – https://www.humave.de/blog/ki-kennzeichnung-auf-social-media-was-instagram-tiktok-youtube-und-linkedin-wirklich-verlangen

[12] artificialintelligenceact.eu – https://artificialintelligenceact.eu/article/50

[13] makeinfluence.com – https://www.makeinfluence.com/en/academy/ai-voice-cloning-and-dubbing-of-creator-content-rights-and-disclosure-under-the-eu-ai-act

[14] gesetze-im-internet.de – https://www.gesetze-im-internet.de/urhg/__44b.html

[15] digital-strategy.ec.europa.eu – https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content

[16] govtrack.us – https://www.govtrack.us/congress/bills/119/s1367/text

[17] hklaw.com – https://www.hklaw.com/en/insights/publications/2024/04/first-of-its-kind-ai-law-addresses-deep-fakes-and-voice-clones

[18] ratgeberrecht.eu – https://www.ratgeberrecht.eu/aktuell/voice-cloning-und-recht/