Japan Just Wrote Down the Rule for AI Voices. The Test Is Not How Similar It Sounds.

On 7 August 2026, Japan's Ministry of Justice published a document that almost nobody outside Japan will read, and that anyone commissioning voice work probably should. Its title is the “Report of the Study Group on Civil Liability for the Unauthorised Use of Likeness, Voice and the Like,” and its subtitle says exactly what it is: “Interpretive Guidelines on Publicity Rights Infringement by Generative AI.” The panel behind it was convened in April and met five times, from 24 April to 27 July. What it produced is an official answer to a question that has been argued in comment threads for two years: does a person's voice carry the same legal weight as their face? In Japan, that answer is now on paper, and it is yes.

Start with what it is not. It is not new legislation, and it creates no penalties. The report describes itself as presenting its content as interpretive guidance, written in the hope that people whose business involves likeness and voice, legal professionals, and those who develop, provide or use generative AI can all better predict where civil liability falls. That distinction is worth holding onto, because it cuts both ways. Nothing became illegal on 7 August that was legal on 6 August. What changed is that the reading a Japanese court is likely to apply is now written down, in detail, by the ministry — which makes it considerably harder for anyone to say they did not know.

The test is commercial pull, not resemblance

The anchor is a 2012 Supreme Court judgment, decided on 2 February that year and known as the Pink Lady case — the first time Japan's Supreme Court squarely recognised a publicity right, derived from personality rights. It set out three situations: using the likeness itself as a product to be appreciated on its own, attaching it to goods to differentiate them, and using it as advertising. Infringement arises where the use can be said to be solely for the purpose of exploiting the customer-drawing power the likeness carries. The 2026 report confirms that a voice sits inside the “likeness and the like” of that judgment — a voice is information that identifies one person from another, and a symbol of that person's personality. The panel records no dissent on the point.

Then comes the part worth internalising, because it is the opposite of how most people assume this works. The report notes that scientific comparison of voiceprints can be one factor, but never a decisive one — the question is not how closely two waveforms match, it is whether that person's pull is being used. And a voice that merely resembles someone can still count as theirs: where the degree of similarity, taken together with the surrounding cues that call the person to mind, means the audience identifies the voice as that person's, the report treats it as use of that person's voice. Sounding “close but not identical” is not a safe harbour. Impressions are the instructive counter-example. Mimicry normally does not infringe, precisely because the performer states their own name — the identifying information points at the performer, not the target. Drop the name, let listeners mistake it for the real person, and the report says infringement becomes possible. The line is attribution, not similarity.

Narration and audiobooks are explicitly in the frame

For the first of the three situations — the voice itself as a product to be appreciated on its own — the report's examples start where you would expect, with celebrity voice messages and voice stamps, and then keep going: it records the scholarly view placing narration, audiobooks, car navigation and smart speakers in the same category. That is not an exotic edge case. That is the ordinary catalogue of what AI voices are sold to do. Two further points should give the supply chain pause. First, fame is not the threshold: the report is explicit that it is enough for the voice itself to have the power to promote sales, and that the actor's name or face need not be widely known — a voice recognised only within one genre, or one region, still qualifies if it moves that audience. Second, it reaches upstream. Selling a training dataset built specifically around a particular actor's likeness or voice can itself fall within the framework, potentially as differentiation of the dataset as a product. The exposure is not confined to the finished ad.

It is also not only about famous people. Alongside the publicity right sits a separate right not to have your likeness — and now, expressly, your voice — used at will, which protects mental rather than commercial value, and which belongs to everyone. In one of the report's worked examples, a synthesised voice made to read obscene text, it concludes that there is unlikely to be much difference in the analysis depending on whether the person is a professional voice actor or an ordinary member of the public: the shame, the discomfort, the injury to self-respect do not scale with fame.

Why the guidance arrived now

The scale explains the timing. On 25 June 2026 a Japanese non-profit that monitors likeness and publicity rights published its survey for the year running April 2025 to March 2026. Across TikTok, X and YouTube it counted more than 40,000 posts suspected of infringement, drawing roughly 335 million views, with economic loss to performers estimated on the order of ¥2 to 4.5 billion. Of the 174 companies that responded to its industry survey, 1.1% had a response policy in place — 46.6% had no plans to write one. The group also tested removal one layer up, filing against 20 unauthorised models trained on a single actor's likeness and getting all 20 taken down, after which new models of the same person were posted. A 100% removal rate that does not stay removed is a fair picture of the problem the ministry is answering.

What it changes if you are the one buying the voice

Two obligations now sit on the same audio file, and they are asking different questions. Europe asks whether it is labelled: the EU AI Act's transparency rules for synthetic audio have applied since 2 August 2026. Japan asks whose voice it is, and whether that person agreed. Neither question is answered by how good the output sounds, and neither is answered for you by a self-serve tool. So the practical version, before you generate anything you intend to publish, is short: can you name the human this voice came from, and can you produce the document in which they agreed to this use, in this market, for this length of time? It is also fair to say this is still moving. The panel did not reach consensus on everything — whether an agency, as distinct from the person, can bring a claim was left open — and Japan's IP strategy plan adopted on 12 June 2026 keeps hard law on the table, including whether the Unfair Competition Prevention Act should be amended.

None of this makes AI voice a bad idea. It makes provenance the thing you buy alongside the audio. That is the part we did not change when we moved from running a voice studio in Taiwan since 2008 into AI: every AI voice in the Onyx library traces back to a named professional who recorded it under a written licence that says what it may be used for, and what comes back to you carries that trail rather than leaving it on your side of the table. If you cannot answer “whose voice is this” for a file you are about to publish, that is worth finding out before it runs, not after. Tell us what you need voiced and where it will run, and we will make sure the answer comes with it.