Japan Actors Union sets generative AI rules for voice actors

The Japan Actors Union (JAU) has published its first guidelines for the use of generative AI in vocal audio work, putting consent, disclosure and performers' control at the centre of a debate that has become impossible for the entertainment industry to ignore.¹

The 21-page document, titled “Guidelines for the Use of Generative AI on Vocal Audio Work”, contains six sections and 22 articles.² It is aimed at protecting actors and voice actors whose recorded performances can be copied, modelled or repurposed by AI systems without their knowledge. The JAU presented the guidelines at a press conference on Thursday, where Dragon Ball voice actors Masako Nozawa and Bin Shimada discussed the personal impact of unauthorised voice cloning.¹

Why the guidelines matter

The union's intervention arrives as synthetic voices become easier to create and distribute. The guidelines define vocal audio broadly, covering recordings of actors and voice actors, including private recordings. They also describe AI use widely: the definition includes voice synthesis, deep learning, machine learning, modelling through prompts, metadata configuration and the operation of software interfaces.¹²

That breadth matters because the risk is not limited to a recognisable digital performance released as a finished song. A voice can be collected as training material, shaped into a model, or used to generate lines that a performer never recorded. The guidelines treat those stages as part of the same rights conversation rather than waiting until a fake performance has already reached an audience.

Nozawa, best known internationally as the voice of Son Goku, said actors' voices were being used without consent. Shimada described hearing an AI version of Ado's song “Usseewa” made with a voice based on his Broly performance, saying he could not accept a voice being published online to gain followers or profit when the actor had not agreed to it.¹

What the rules cover

The document's first section sets out its purpose, definitions and criteria for unauthorised use. Later sections establish standards that prioritise human performers and set out rules for notification, written consent and disclosure.²

The proposed safeguards include the right to request crediting and removal, protection against unfavourable treatment for refusing or withdrawing consent, and limits on how long consent remains valid. The guidelines also address compensation, oversight, audits, data-log requests and requests to stop or remove unauthorised AI-generated vocal work.¹²

There are dedicated provisions for Japanese-language vocal audio used in animation and dubbed content, alongside separate rules for multilingual localisation and overseas distribution. That is particularly relevant to anime, where a performance may travel across territories and languages while remaining closely associated with a specific actor.¹

The JAU's approach also connects with a wider push by Japanese performers to establish practical protections. Voice actor Yuki Kaji launched the Soyogi Fractal voice-synthesis programme in 2024 partly in response to concerns around copyright and unauthorised voice use. Other performers formed the No More Unauthorized Generative AI group in late 2024, while Kenjiro Tsuda pursued legal action over alleged unauthorised imitation of his voice on TikTok.¹

The legal position is still developing. Anime News Network reports that both a Ministry of Justice expert panel and a Tokyo District Court recognised voice actors' voices as legally protected under the right of publicity, although Tsuda's lawsuit was dismissed after the relevant videos had been removed.¹ The JAU guidelines therefore function as both a practical framework for industry behaviour and a statement of what performers expect from companies building or deploying synthetic-voice systems.

Our opinion

The strongest part of the JAU's proposal is its insistence that consent must be specific, written and reversible. “AI use” is too broad a label for a single yes-or-no permission: training, internal testing, commercial release, localisation and future model reuse can all carry different consequences. Treating them as separate uses gives performers a meaningful choice instead of asking them to sign away control of an unknowable future product.

The guidelines will not, by themselves, stop every unauthorised clone from appearing online. Enforcement, transparent data records and workable removal processes will decide whether the principles have teeth. Even so, the JAU is addressing the real issue: a voice is not merely a file that can be copied. It is part of a performer's identity, livelihood and relationship with an audience. Any company using synthetic vocal audio should have to explain where the recording came from, what the model will do and how the performer can say no.

Hero image: Photo by Odin Reyna on Pexels.