Japan Actors Union Releases Generative AI Voice Guidelines: A Voice Actor’s Voice, a Player’s Auditory Memory#

Zhanlan Yingren | October 21, 2026

Hello everyone, Zhanlan Yingren here. The keyword of today’s story is voice — specifically, who owns a voice actor’s voice in the generative AI era, how it can be used, and at what price.

As reported by multiple industry outlets this week, the Japan Actors Union — the organization many reports call Japan’s voice actors’ union — has published an industry rulebook titled Guidelines for the Use of Generative AI Related to Voice, covering roughly 2,600 actors and voice actors and standing among Japan’s first industry guidelines written specifically for AI voice. Per Japanese AI industry outlet Global AI News, the guidelines were formally released on October 15, and coverage rolled in over the following week: games industry outlet The Daily Game reported on October 19, Tencent News’ IT Times channel ran a four-point Chinese-language recap on October 20, and European games industry outlet GI.biz followed on October 21 — a news cycle spanning a full week.

The union’s official statement deserves to be recorded as-is: technological progress will not stop, but living human expression and respect for rights must always sit at the center. That sentence is the bedrock of the whole document, and of the two threads this article pulls on — one is the governance of voice actors’ rights; the other is what the auditory memory built out of voices means to us as players.

A note on sourcing first: every clause, name, date, and figure in this piece stays within what multiple public reports corroborate. Where a detail rests on a single outlet, I flag it individually and extrapolate nothing.

1. First, the Basics: Who Published These Guidelines and How Far They Reach#

Following the roundup by overseas tech outlet Digital Watch Observatory and multiple Chinese-language media, here is the baseline information — pinned down before any clause-by-clause discussion:

The publisher itself deserves an extra sentence: the Japan Actors Union, an industry organization covering roughly 2,600 actors and voice actors. Multiple outlets referred to it as an actors’ union; this piece consistently calls it the union. Its positioning is as one of Japan’s first industry guidelines for AI voice, and its nature is industry guidance rather than law — a point Section 6 takes up in full, because it is where the document’s greatest tension lies.

On terminology, the guidelines define key terms such as performer, performer’s voice, digital replica, and synthetic performer. Do not underestimate this — terms pinned down first give the clauses their boundaries, and that is the precondition for any rulebook to be enforceable.

The scope of application is broader than many readers would expect. As relayed by Digital Watch Observatory, the chain of coverage runs as follows:

1. Data collection — the stage where voice print data is collected from actors is already governed; it is not only regulated at the moment of generation;
2. Model development — training a model on an actor’s voice is itself a standalone stage that requires authorization;
3. Generation and distribution — individual generation sessions and shipped output are auditable case by case;
4. Subsequent reuse — voice material generated now and repurposed later also falls under the guidelines;
5. Extraterritorial activity — even activities carried out outside Japan fall within coverage as long as they target Japanese audiences. This clause is clearly aimed at the playbook of moving recording overseas to dodge oversight.

The attribution requirement is concrete too: AI-generated voice must carry attribution naming the AI provider, the product, and the actor, unless the actor requests anonymity. In other words, under the guidelines’ design, every AI-synthesized line a player hears should, in principle, come with the knowledge of what it is they are hearing.

One sentence on the document’s stance: it does not ban generative AI; it draws procedural boundaries around how a human voice may be used. That is exactly what separates it from a protest statement — protests express an attitude, guidelines set operating rules.

2. The Four Core Clauses, One by One: Every Line Aimed at the Production Pipeline#

The four core clauses are the ones every outlet keeps quoting. I first lay them out in a table — clause content against what it means for the voice production pipeline — then expand each:

Clause 1: Informed Consent, the Heaviest of the Four#

Per the consistent reporting of Digital Watch Observatory and Global AI News, every stage of the generative AI pipeline requires permission — written prior consent from the performer, separately, for training, generation, and output use. More consequential are the exclusions: blanket buyout contracts, verbal consent, implied consent, and retroactive approval are all deemed invalid.

This clause effectively sentences the industry’s standard play — slipping a line into the contract saying the performer agrees to the company using their voice for all purposes, including AI — to death. For actors, it is a move from an all-purpose sell-off back to item-by-item licensing; for companies, legal and process costs rise in earnest, but what they get back is a licensing chain that is clean and auditable.

Clause 2: Compensation — AI Replication Is a New Performance#

AI replication of an actor’s work is deemed a new performance, for which companies must pay the actor royalties or compensation. The elegance of that framing: it skips the argument over whether an AI-synthesized voice counts as the original performance and simply defines it as a new performance — and a new performance naturally demands new consideration.

Live-service games need voice pack updates, anniversary voice lines, and derivative animation; in the past, a one-time recording fee covered none of that downstream use, and now every reuse creates a new payment obligation. For voice actors, this clause changes the shape of income: every reuse of the voice gets priced again.

Clause 3: Translation — the Clause Games Industry Readers Should Watch Closest#

Generative AI translation (automatically converting an actor’s performance into other languages) must preserve the artistic integrity of the original performance, and may not be used in overseas regions without permission.

Consider the hottest direction right now: using AI to convert Japanese voice acting into Chinese, English, and Spanish automatically, at a cost low enough to give every NPC full voiced lines. This clause amounts to saying: if that pipeline wants to run, it first needs the actor’s license for every language version, and the converted performance must not drift out of shape. How to judge artistic integrity is a genuinely hard question, but no overseas use without permission is a clear, hard constraint — localization teams can no longer treat machine-translated voiceover as a free lunch that never touches the actor.

Clause 4: Deceased Performers — the Clause with the Least Room#

Using generative AI to replicate a deceased actor’s voice is prohibited unless the actor explicitly authorized it while alive; even with such authorization, permission from family or from the agency or office the performer belonged to is still required, along with royalties.

Double licensing stacked on royalties — the bar here is deliberately high. For the games industry, this clause touches a real scenario: for titles that have operated for a decade or more, when the original voice actor passes away, does the character’s voice get sealed for good, re-recorded by someone else, or frozen at the last version? The answer the guidelines lean toward is that default replication is off the table.

Read together, the four clauses point in one consistent direction: lift the right to use a voice out of the company’s contract pocket, hand it back to the performer, and price it item by item — by stage, by use, by language.

3. Procedural Baselines: 7 Days, 48 Hours, and Separate Consent#

Beyond the clauses, the guidelines come with procedural requirements — the part that reads most like an operations manual. First, the deadlines in a quick-reference table:

On the multi-source-consistent parts, in detail:

• 7 days in advance: written notice of at least 7 days before providing voice print data to the actor and their agency or starting the development process.
• 48 hours in advance: notice of at least 48 hours before a specific generation session begins.
• Separate consent: consent is obtained separately per project, per usage stage, and per language version — three dimensions, project, stage, and language; change any one of them and you re-license.

The logic of these deadlines is not hard to read: 7 days gives the actor and their agency time to evaluate and negotiate; 48 hours guarantees that every specific generation happens while the actor is in the know; separate consent slams the door on bundling every derivative scenario under a single authorization.

The three single-source items — the one-year validity, the 30-day deletion, the 14-day disclosure — are recorded here as reported, but readers should note their level of corroboration. If more channels confirm them later, those three clauses would turn the right of withdrawal from a paper right into an executable process: an actor could not only say no before signing, but also exit mid-collaboration with a 30-day data deletion deadline — a fairly strong check on anyone hoping to accumulate voice data long term.

Put the four core clauses next to this procedural layer and the document’s real shape comes into focus: this is not a declaration of principles but a process rule that tries to plug directly into voice production schedules. When voice print data changes hands, when a generation session opens, how long each authorization lasts, how it gets renewed — every one of these gets a numeric floor.

4. These Guidelines Did Not Fall from the Sky: the 2024 Protests, the Tsuda Lawsuit, and Masako Nozawa’s Remarks#

To understand the guidelines, you have to place them back into the arc of Japan’s voice acting industry over the past two years. Laid out in order, there are three milestones:

Milestone 1: 2024, Organized Protest#

Per the timeline from industry tracker AIBase, in 2024 a number of Japanese voice actors formed an organization to protest AI voice abuse, arguing directly that existing copyright law leaves voices under-protected. In other words, the industry’s anxiety did not begin with these guidelines; the organized protest has been running for at least two years — and the guidelines are the first written fruit of that two-year arc.

Milestone 2: Kenjiro Tsuda Sued TikTok — a Lost Case That Set a Precedent#

Voice actor Kenjiro Tsuda sued short-video platform TikTok after AI videos using his voice without authorization spread widely across social platforms. As relayed by AIBase, the case was dismissed after the videos in question were taken down — but the lawsuit left behind something more important than the verdict: it established the precedent that an individual’s voice is protected under publicity rights.

Tsuda’s representative works are usually listed as Attack on Titan, Chainsaw Man, and Yu-Gi-Oh!, with The Daily Game adding Jujutsu Kaisen. One look at that list makes it obvious his voice is bound to a long string of hugely popular characters — the demonstration effect of the case inside the voice acting community speaks for itself.

In a sense, these guidelines were published standing on that precedent: the precedent settled whether a voice is protected, and the guidelines try to answer how it may be used with respect for that protection. From case-by-case defense to written rules is a standard, solid step in labor-rights action.

Milestone 3: Masako Nozawa at the Press Conference — No Different from Theft#

Per AIBase’s Chinese-language relay, Masako Nozawa — the voice actor known to generations of audiences as Son Goku in Dragon Ball — said at the press conference that using her voice without permission is no different from theft, that voices and performances deserve respect, and that even small-scale use requires prior consent; she also pointed out that AI-replicated voices lack emotional expression.

One clarification: the full text of her remarks currently exists only in this one outlet’s Chinese relay, and details should defer to a complete on-site record. But even in relay form, the two layers are clear: the first half is a rights claim — prior consent even for small uses, aimed squarely at the what-harm-is-one-short-clip rationalization; the second half is an artistic judgment — AI replicas lack emotional expression, which we take up in the next section.

The phrase no different from theft speaks in the language of legal rights, but it lands on emotional ground: recognition. When a character has walked alongside a player for ten or twenty years, what the player remembers is not just the lines, but the breath, the pauses, and the rasp of that voice — things the union’s vocabulary calls performance, and players’ memory calls this character is simply this voice. Connect the three milestones and the line is straight: the 2024 organized protest supplied the attitude, Tsuda’s lawsuit supplied the legal footing, Nozawa’s remarks supplied the emotional footnote, and these guidelines gather all of it into clauses that can be written into a contract.

5. A Voice Actor’s Voice Is the Bedrock of Players’ Auditory Memory#

At this point, a few more words from the editor’s chair — this story hits close to home for an old Chinese-MMO player like me.

I have been playing since the Stone Age era — the classic MMORPG. Games back then had no voice acting; a character’s emotions lived entirely in text boxes and a few system sound effects, and our image of what a character sounded like was something we imagined ourselves. Later, Chinese MMOs laid down voice acting piece by piece — someone was reading the main story, NPCs started talking, login screens got theme vocals — and not long after that, when a character changed voice actor, players would argue on forums for days. What is that argument really about? Players discovering that the auditory memory inside their heads has been tampered with.

Japanese RPGs pushed this to another extreme. A character’s voice is part of how the character is built; players can close their eyes and identify who is speaking the second a line lands, and that ear itself is an asset banked over years of play. Masako Nozawa says AI replicas lack emotional expression, and players’ gut feeling comes from the same source: a synthesized line can approach the timbre, but the rhythm of the breathing, the hesitation inside a line, the restraint of a laugh cut short — that living human expression still cannot be replaced.

And hearing is a brutally picky form of memory — when the art style changes, players say the visuals got an upgrade; when the voice changes, they only say this is not my character. Sound has no middle ground, because it is the way a character speaks beside your ear, and close enough does not cut it. I have lived this moment myself: after a game I had played for years updated, a long-standing character’s voice changed, and from the very first line the world felt wrong — same character, same art, same skills, but the moment that familiar line came out, a nameless strangeness pressed in. Only later did I understand: that sense of familiarity never came from the character itself; it was banked day after day by a voice that kept the character alive.

So what moves me most in these guidelines is not any enforcement clause but that line in the union’s statement — living human expression must sit at the core. On the surface, voice actor protection is about employment and pay; underneath, it is about players’ auditory memory — and seeing that thread written formally into an industry rulebook is, in my experience, rare. A voice is not a resource file in an asset library; it is a performance, and behind every performance is a living person.

6. Guidelines Are Not Law: the Real Teeth Await a Law Change, and Players Will Not See Changes Soon#

Time to pour the cold water and manage expectations: these guidelines are not legally binding. Per The Daily Game and Tencent News, the union says it will push for the guidelines to become an industry-wide standard and push the government to legislate protection for voice rights. In other words, for now this is industry self-regulation whose force runs on consensus, public opinion, and leverage at the negotiating table — not on court subpoenas.

The legislative movement, however, is real: per Digital Watch Observatory, the Japanese government is considering revising the Unfair Competition Prevention Act to provide legal protection against unauthorized use of AI-generated voices, and the guidelines landed right alongside that discussion.

That draws a clear timeline: industry rules first set the operating standard, government law revision follows with enforcement force, and two years of organized protest plus one precedent-setting lawsuit laid the ground for this timeline. The guidelines’ real teeth will most likely only grow in once the Unfair Competition Prevention Act revision lands.

One more stark fact: as of now, not a single Japanese AI company or platform has publicly committed to adopting these guidelines — The Daily Game states plainly that no company has publicly agreed to adopt them. The rules are out and the regulated are silent; that is the probation period every self-regulation document goes through, and whether negotiating practice turns it into convention or legislation turns it into obligation should become clear within the next year or two.

Potential Effects on Voice Production Pipelines and Cost Structures in Games#

The three points below are pipeline reasoning built from the clauses — analysis, not established fact:

1. Voice contracts will be rewritten. Three-stage written consent plus separate licensing means blanket voice authorization clauses are no longer acceptable; companies will need to sign separately for training, generation, and output use — a predictable increase in legal workload.
2. Localization pipeline costs will change shape. AI translation must preserve artistic integrity and be licensed per language version, landing directly on multilingual voice pipelines — precisely the scenario AI replication most wants to cheapen. Every company will re-run the math between money saved and royalties owed.
3. Voice planning for long-running characters moves up to signing time. The deceased-actor clause plus the consent validity cap (one year at most, per the single source) means titles with operating horizons measured in decades now have to think through renewal and succession of voice rights at contract time.

The Direct Takeaway for Players#

This is an industry rule plus a legislative signal, not a game patch note — in the short term, you will not see any instant change in voice production inside any game you play.

The two follow-ups worth watching: progress on the Unfair Competition Prevention Act revision, and whether a first publisher publicly commits to adopting the guidelines. The day either lands, this story stops being industry news and becomes player news.

7. This Story Is Not the Same as the AI Topics We Have Covered Before#

Regular readers may remember we have covered AI topics often: how AI game production tools are rewriting the development pipeline, AI-rebuilt classic games hitting rights holders’ red lines, generative AI usage among Japanese game developers soaring to 85.8%, and the copyright dispute over an OpenAI employee’s AI-made imitation of Legend of Runes being halted and pulled.

This time, though, I want to draw a clear line between this story and those. In our earlier pieces, the subject was the tool — how much faster AI can make development, whether AI-generated content infringes, whether the tools feel good to use. The subject here is the person — one person’s voice, one actor’s right to their performance, one person’s ownership of an auditory memory. A tool’s efficiency can be quantified; the boundary of a right can only be negotiated. The former asks what AI can do; the latter asks what AI must not touch. The guidelines’ answer is plain and hard: not touching starts with written consent at every stage.

One more note: some readers may have expected a clause-by-clause comparison with labor agreements among Western peers. That comparison cannot be done right now — the cross-checkable public material available this time lacks content directly comparable to the terms of Western agreements, and writing one anyway would mean leaning on imagination, which we will not do. When comparable public material exists, we will pick the topic back up.

Finally, a checklist of this article’s watch points, for follow-up verification:

Final Thoughts#

From the pure text boxes of Stone Age to a full industry rulebook over whether a single voice line may be AI-replicated, twenty-odd years on, the relationship between players and voices has been formally put on the table. As an old player who once had to imagine what characters even looked like, I know exactly how much weight the phrase auditory memory carries — it is not written in config files; it is written in thousands of hours of play.

It is too early to say what form these guidelines will ultimately take in the games we play. But the direction is clear: as technology’s cost curve keeps falling, someone has to hold a line for human expression — not to block the technology, but to make it knock before touching a person’s voice.

Technology moving forward is the tide, and the union’s statement did not dodge it: technological progress will not stop. But the second half of that sentence matters more — living human expression, and respect for rights, must always sit at the core. Here is hoping that the new voices we hear in games over the next decade still come from living throats and carry the warmth of breath; and that every voice line AI helps synthesize tells us plainly whose it is, where it came from, and who said yes.

Voices deserve that much care.

— Zhanlan Yingren, NTGame Editor