Blog
Human voice-over vs AI: legal boundaries, voice cloning and fines of up to €15 million
The EU AI Act, deepfake audio, performers' rights and fines of up to €15 million — what a company should know before publishing a synthetic voice.
Only a few years ago the question was whether artificial intelligence would ever sound convincingly human. Today the more relevant question is: when is AI allowed to sound like a particular human being?
Modern speech synthesis can produce natural, fluent audio within seconds. Voice cloning goes further. A model can reproduce the timbre, rhythm and speaking characteristics associated with a real person.
From a production perspective, the attraction is obvious. No recording session. No waiting for the talent. A line changes in the script and a replacement file can be generated immediately. Then somebody asks a rather uncomfortable question: whose voice is this, and who authorised us to use it?
Since 2 August 2026 that question has become even more important in the European Union. Article 50 of the EU AI Act now applies to transparency obligations concerning certain AI-generated and manipulated content.
An AI voice is not automatically illegal
There is nothing inherently unlawful about synthetic speech. Companies can use text-to-speech systems, generic synthetic voices and legitimately licensed voice models.
The situation changes when the generated audio begins to resemble an identifiable real person. Under the AI Act, a deepfake can include AI-generated or manipulated audio that resembles an existing person and could falsely appear authentic or truthful to someone hearing it. Deployers using AI to create such deepfake audio are required to disclose that the content has been artificially generated or manipulated.
That distinction matters enormously in advertising. A commercial does not have to mention a famous voice artist by name to create a problem. If the entire idea is to make listeners think, “That sounds exactly like him”, the absence of a name does not automatically make the use legitimate.
“We didn't copy his recording. We only created a similar voice”
This is where discussions about generative audio often become misleading. The legal question is broader than whether somebody copied a particular WAV file.
A recognisable voice can be connected with a person's identity and personality rights. In Poland, for example, civil-law protection of personal interests is based on an open catalogue, and legal scholarship has long discussed the protection of a recognisable voice as an acoustic equivalent of a person's image.
A separate layer can arise from performers' rights. A professional voice artist's creative delivery may qualify as a protected artistic performance, depending on the circumstances.
This is why the useful question is not merely: “Did we steal the recording?” It is: “Are we commercially exploiting a recognisable vocal identity without permission?”
A licence to use a recording is not necessarily a licence to clone a voice
Suppose a voice artist records a radio commercial. The client receives the agreed right to broadcast it for one year. Does that automatically mean the client may upload the recording to an AI service, build a digital replica of the talent and make the replica read entirely new scripts? That is a very different use.
- A licence to publish an existing recording online does not automatically mean permission to train a voice model.
- Permission to edit a recording does not necessarily mean permission to generate hundreds of new statements which the performer never actually made.
Whenever AI cloning is planned, the agreement should deal specifically with it rather than assuming that ordinary recording rights somehow cover everything.
The €15 million number is real – but it needs context
The AI Act's transparency requirements under Article 50 apply from 2 August 2026. For deepfake audio, deployers must disclose its artificial origin. The Act's penalty framework places breaches of Article 50 transparency obligations in a category carrying administrative fines of up to €15 million or 3% of total worldwide annual turnover for the preceding financial year. SMEs, including start-ups, are subject to the more favourable cap mechanism provided for in Article 99.
That does not mean that every badly labelled voice-over will result in a €15 million fine. It is a statutory maximum, and enforcement has to take the circumstances and proportionality into account. But it changes the commercial conversation. Labelling the origin of synthetic media is no longer something that should be discussed five minutes before a campaign goes live.
What about GDPR and biometric data?
Another misconception is that every voice recording automatically constitutes biometric data. It does not. Voice data becomes particularly sensitive from a biometric perspective when it undergoes specific technical processing for the purpose of uniquely identifying a person. The European Data Protection Board gives voice-model recognition and voiceprints used to identify users as an example of such biometric processing.
A voice-cloning project may therefore involve several legal layers at the same time:
- personality and identity rights,
- performers' rights,
- contractual licensing,
- the AI Act,
- data-protection law,
- and potentially liability connected with what the synthetic person is made to “say”.
That is why the cheapest voice file can sometimes become the most expensive part of a campaign.
Provenance is becoming as important as audio quality
Advertising producers already understand music licensing. Where did the track come from? Who owns it? Can it be used in paid advertising? For how long? In which countries? Synthetic voices should be treated with the same discipline.
Before using an AI-generated commercial voice, it makes sense to establish who licensed the model, how it was created, whether it resembles an identifiable person, whether commercial advertising is permitted, whether there are territorial or sector restrictions, and whether the output must be labelled as AI-generated.
A sentence such as “our supplier says the rights are fine” may not be much help once a campaign is already running.
A real voice artist still needs a licence. The difference is provenance
Professional voice-over recordings have never been legally frictionless. Usage rights for an internal corporate film differ from those for national radio advertising or a paid online campaign.
The practical advantage is that a traditional production has an identifiable performer, a documented recording session and an agreed licence.
A professional RPM Studio voice bank therefore provides more than a selection of attractive voices. It creates a traceable production chain between the client, the studio and a real performer. RPM's voice bank currently includes hundreds of Polish and international voice talents working across dozens of languages.
RPM Studio operates recording facilities in Warsaw, Wrocław and Tarnów and handles voice-over recording, editing, sound post-production, mixing and mastering.
The technology behind AI may be more futuristic. But if someone asks twelve months later where the voice in a campaign came from, “we have the documented rights to this recording” remains a remarkably modern answer.
Can a cloned voice be detected?
A suspicious recording can be compared with verified samples of a person's real voice. Such analysis can measure the strength of the similarity between voice characteristics. A strong match, however, does not by itself prove that AI was used. A genuine recording of the same person should also show a high level of similarity.
Voice analysis is therefore useful evidence, but it is not a magical “AI / not AI” switch. When provenance is uncertain, companies should preserve original voice tracks, project files and licensing documentation instead of relying purely on what the audio seems to sound like.
RPM describes this process in more detail on the voice verification page and in the article on checking whether a recording uses a cloned voice.
AI is here to stay. Audio procurement needs to catch up
Synthetic speech has legitimate and useful applications. It can work extremely well in prototypes, accessibility tools, internal applications, automated announcements, localisation workflows and productions based on properly licensed synthetic voices.
The problem begins when technological speed replaces rights clearance. The true cost of voice production is therefore no longer just the price of generating or recording a file. It also includes certainty. Are we actually entitled to publish the voice we are about to use?
That question gives professional voice-over recording and a documented voice bank a new role. Not because AI is inherently bad. Because professional audio should have more than good sound. It should also have a known source and clearly defined rights.