Japan's Voice Actors Just Wrote the AI Consent Rules

Masako Nozawa has voiced Son Goku since 1986. She is 89. In 2025 she became the first voice actor to receive one of Japan's highest cultural honors.

Speaking to NHK this month about videos that recreate her voice with AI, she said this: "This is not my voice. Using it without permission is unforgivable. It's the same as theft, and it makes me angry."

She also said the copies can be technically flawless and still fail. "It might even be a perfect copy. But it has no heart, does it?"

Those quotes would be a nice human-interest story on their own. They landed alongside something more consequential: on October 15, the Japan Actors Union published guidelines on AI voice use, covering roughly 2,600 actors and voice performers.

What the guidelines actually require

The core provision is consent before the fact. Performers must agree before their voices are used to train AI models, to generate new recordings, or to appear in published content. The guidelines also set out compensation and protections against unauthorized use.

That ordering is the important part. It's not a takedown mechanism, it's a permission mechanism. The default flips from "use it and remove on request" to "ask first," which is a much bigger change than it sounds, because it puts the burden of proof on the person doing the cloning rather than the person being cloned.

It also gives performers something individual court cases rarely deliver: a checklist. Before anyone signs a synthesis deal, the union has defined what a fair one contains.

A union is a strange and effective place to write a standard

Most countries are approaching this through legislation or litigation. Japan is doing it through collective organization, and there's a reason that might work faster.

Legislation is slow and general. Courts are slower and specific to whoever filed. A union guideline is neither. It's a market standard that 2,600 working professionals can refuse to work without, which means producers have to price it in before the first project rather than after the first lawsuit.

It also solves a coordination problem that individual performers can't. A young voice actor asked to sign a broad synthesis clause has very little leverage. A union saying "this clause is not acceptable" gives them room they didn't have to negotiate alone.

The obvious limitation: a guideline isn't enforceable law. If a studio ignores it, the remedy is industrial rather than legal. But in an industry where the same four agencies book most of the work, industrial pressure is not nothing.

Nozawa isn't against the technology

The detail most coverage skipped: she also said she finds AI exciting.

Her objection is specifically about authorization, not about synthesis existing. "I can't forgive using it without asking," she said. "I want people to follow the rules."

That position is more useful than a blanket rejection, because it points at a workable line. The technology is fine; the missing piece is a permission layer and a payment. Which is exactly what the guidelines try to construct.

It's also the same position taken by Megumi Ogata, who voiced Shinji in Evangelion and called AI voice cloning "heartbreaking" in August while asking the government to intervene. Two very different performers, the same ask: not a ban, a rulebook.

Why voice is a harder case than image

There's a structural reason performers are winning this argument faster than visual artists.

A voice is unambiguously one person's. Nobody argues that a voice is a genre or a technique. When someone clones Nozawa's Goku, there is no plausible claim that they were merely inspired by a style — they reproduced an identifiable individual's instrument.

Style is the ambiguity that has kept image litigation tangled for three years. Voice doesn't have that escape hatch. Which is why the voice cases have moved faster, and why the standards being written now will probably become the template that visual likeness rules get modeled on.

For anyone building products that synthesize faces or voices, that's the signal to read: the consent architecture is being defined by performers first, and it will be exported.

The dubbing market is where this gets expensive

Anime voices are the visible case, but dubbing is the commercial one.

Every foreign film and series released in Japan gets re-voiced, and that work is a large, steady, unionized slice of the industry. It's also the most obviously automatable part of the job: match the lip flaps, keep the character's register, deliver on schedule. A model that can do passable dubbing at a tenth of the cost is an existential offer to the people who do it for a living.

Which is exactly why consent rules matter more here than in any other corner of voice work. If dubbing becomes synthetic by default, the profession shrinks to a supervision role. If it requires consent and payment, it becomes a licensed input — the same structural position a songwriter is in when a cover version gets made.

The guidelines don't resolve that fight. They just make sure performers are at the table when the terms get set, rather than finding out afterwards from a trailer.

What this means in practice

Three things, for three different people.

If you're a performer, the useful artifact is the checklist. Ask for the scope of training use, the duration, the territory, whether sublicensing is allowed, and what happens to derivatives when the contract ends. Get the answer in writing. "We won't clone you" is not a clause.

If you're building a product, consent is becoming a data problem before it's a legal one. You need a record of who agreed to what, tied to the specific voice model, with an expiry. Studios that can produce that record on request will win contracts; studios that can't will lose them regardless of output quality.

If you're an audience, the interesting shift is that "is this a real person" is becoming the wrong question. The right one is "did this person agree." The first is a detection problem that nobody is winning. The second is a paperwork problem that can actually be solved.

The line that will get quoted

"Without heart, it's no good, whatever the role. Even a voice that's only crying is different when you cry with real feeling."

You can read that as an old performer defending her craft against a machine, and plenty of people will. But it's a more precise claim than that. She's saying the thing being copied isn't the sound — it's the decision behind the sound, the choice to cry a particular way on a particular take.

You can't license that, and you can't synthesize it. You can only hire the person who does it. The guidelines exist to make sure that hiring still happens, and that when it doesn't, someone gets paid.