Skip to content
Entertainment

Is an AI voice clone a copyright infringement? What courts decided in 2026

AI voice cloning and copyright: a voice is not a work, and copyright protects the recording, not the timbre. New York, Tokyo, Denmark and Munich decided in 2025 and 2026 where the artist sues and what a company must get authorised.

By Cesar Augusto Taborda Lima

No. A person’s voice is not a work, so cloning a voice, on its own, is not a copyright infringement. That is the position in Brazil and it is what the courts that have faced the question have said. Copyright protects the recording and the fixed performance. The timbre, the thing that lets a listener recognise who is singing, is protected by a different route, that of personality rights. The distinction looks subtle, but it decides everything: whom you sue, with what evidence, for how long and under which law.

There is one important exception, and it lies in training. Whoever feeds the model with protected recordings is back in copyright territory, and the Munich court has just said so without hedging.

§ 2. The protection of performers extends to the reproduction of their voice and image when associated with their performances.

Law 9,610/1998 (Brazil’s Copyright Act), art. 90, § 2 (free translation).

The voice is not a work. The recording is

Brazil’s Copyright Act protects “creations of the mind” (art. 7) and leaves out ideas, methods and concepts as such (art. 8, I). A voice is not a creation. It is an attribute of the person, like a face or a signature.

Around the voice, the Act protects other things. The composition, with or without lyrics (art. 7, V). The phonogram, which only the producer may reproduce (art. 93, I). And the performance: the performer has the exclusive right to authorise or forbid its fixation, reproduction or making available to the public (art. 90, I to V). That protection reaches the voice, says § 2, “when associated with their performances”.

The proviso in § 2 is the whole point. If someone lifts a stretch of the original recording, the phonogram and the performance are reproduced, and copyright answers. If a model learns the timbre and sings a song the artist never recorded, no fixed performance is being copied. What is being used is the vocal identity. Protection then comes from the Constitution (art. 5, X and XXVIII, “a”) and from the Civil Code, which allows a person to forbid unauthorised commercial use of their image and words (art. 20) and to demand that the harm stop, with damages (art. 12). Where the use is advertising, the STJ, Brazil’s Superior Court of Justice, does not require proof of loss (Precedent 403).

New York: imitating is not copying

Two American voice actors, Paul Lehrman and Linnea Sage, found their voices cloned in a synthetic voice-over service and sued the company, Lovo. In a decision of 10 July 2025, Judge J. Paul Oetken of the federal court in Manhattan rejected the copyright theory. Protection, he wrote, “does not extend to this kind of imperfect mimicry, even when accomplished using advanced technology rather than more traditional techniques like musical covers or impersonations”. Copyright deals with expression, not with something “as abstract and intangible as a ‘voice’”. The statute had already said as much in its own way: rights in a sound recording do not reach another recording made of independently fixed sounds, “even though such sounds imitate or simulate” the original (17 U.S.C. § 114(b)).

The actors did not leave empty-handed. New York’s right of publicity (Civil Rights Law §§ 50 and 51) survived, together with the contract and consumer protection claims. The shape is the same as Brazil’s: copying the recording is a copyright matter; imitating the timbre is a personality matter.

Tokyo: the voice as a symbol of the person

Kenjiro Tsuda voices Kento Nanami in Jujutsu Kaisen. An anonymous TikTok account posted more than a hundred and eighty videos using a synthetic version of his voice. On 30 September 2026 the Tokyo District Court found an infringement of publicity rights, the use having been made with an obvious intention to exploit the commercial appeal of the voice, and left a line that will be quoted for a long time: “the human voice is symbolic of individual personality, just like one’s portrait”. The request to remove the videos was refused only because the account had already been deleted. It was the first time a Japanese court recognised a right to one’s own voice.

Denmark decided to do it differently. Bill L 49, tabled in the Folketing on 8 October 2026, writes into the Copyright Act a right of every person over their appearance, voice and movements. Realistic digital imitations may not be published without consent, which can be withdrawn at any time, and it is the publisher who has to prove the consent existed. Protection lasts fifty years after death. A separate provision covers imitations of a performer’s rendition, with collective management allowed. Caricature, satire, parody and social criticism are left out, except for dangerous disinformation. The intended start date is 1 January 2027, if the bill passes.

The Danish model settles proof and duration, two things that under personality rights are left to the judge. The price is putting into copyright law something that is not a work. It is a choice, and Brazil has not made it: art. 66 of the AI bill, PL 2,338/2023, merely requires that the use of voice and image by AI systems respect the personality rights of the Civil Code.

Training is a separate account

Everything so far has been about the output. Training the model is a different story, and there copyright is back in charge.

On 31 July 2026 the Munich Regional Court I decided the claim of GEMA, Germany’s music collecting society, against Suno, the AI music generator (the court’s official press release, case 42 O 763/25). Six well-known works had been used in training. With simple prompts, the system gave them back recognisably. The court held that memorisation of the works inside the model infringes the reproduction right, that Germany’s text and data mining exception (§ 44b UrhG) does not cover it, that the outputs communicate the works to the public, and that the company is liable, not the user who typed the prompt. For the part of the training done in the United States, it applied fair use and found that every factor weighed against Suno. An appeal is still open.

Brazil does not need a new statute to get close to that. Use of a work requires prior and express authorisation, and the Act lists among the forms of use “inclusion in a database, storage in a computer” and “any other forms of use now existing or yet to be invented” (art. 29, IX and X). Art. 46 has no data mining exception. PL 2,338/2023, passed by the Senate and stalled in a special committee of the Chamber of Deputies since 2025, would only confirm the logic: a public summary of the content used in development (art. 62), an exception only for non-commercial research (art. 63), a right to prohibit use (art. 64) and remuneration for training (art. 65).

The market is already moving that way without waiting for the law. Since June 2025, Ecad, Brazil’s music royalties collecting body, requires registrations of works and phonograms to say whether generative AI was used, in whole or in part, with the tool and the prompt. Since 25 September 2025, Spotify accepts vocal impersonation only with the authorisation of the artist imitated.

What to do with this

For the artist, the first question is simple: did the synthetic voice use a stretch of your recording, or did it only learn your timbre? The answer chooses between a copyright claim and a personality claim, and nothing prevents both where both things happened. Before notifying the platform, preserve the evidence. The account disappears, as it did in Tokyo.

For the company commissioning a synthetic voice, the AI vendor’s licence settles the copyright side of the model. It does not buy the consent of the person whose timbre the model reproduces. That consent is personal, specific, revocable, and it has to be in writing.

For anyone licensing a catalogue to developers, the money is in the training. That is where the contract should charge.