When generative AI is used to create a voice that sounds like a real person, what should AI creators and companies be watching for?
On September 30, 2026, the Tokyo District Court (46th Civil Division) addressed that question in a case brought by Kenjiro Tsuda, a voice and screen actor known to anime fans as the voice of Kento Nanami in Jujutsu Kaisen, against TikTok Pte. Ltd., the Singapore company that operates TikTok. The court stated that a person’s voice, like their portrait, is a symbol of the individual’s personality, and may in certain circumstances be protected by the right of publicity. Kei Hirano, one of the plaintiff’s lawyers, said it was the first ruling to expressly recognize voice as falling within the right of publicity.
The court did not, however, find that the narration at issue actually infringed Tsuda’s rights. Because the posts and their data were deleted while the case was pending, it dismissed his claim for removal. The ruling does not draw a simple line that “making an AI voice that sounds like someone is illegal.”
Even so, it matters to anyone using AI-generated voices in content, advertising, social media, film, or games. Even if a person’s voice was never used as training data, legal issues can arise when the output evokes that specific individual and is used to exploit their customer appeal.
Key takeaway
The question is not only “Whose voice was used for training?” It is also: Who does the final output sound like? Does it trade on that person’s customer appeal? How is the voice presented, and how is it monetized?
The ruling is best read as opening the door to a broader debate over how the personal and economic value of a voice should be treated in the age of AI.
Table of Contents
What Was at Issue in the Tsuda Case?
The case (Tokyo District Court, Case No. 2025 (Wa) 70635) was an action seeking the removal of posted content. At issue were 188 videos uploaded between July 2024 and September 2025 by an anonymous user’s TikTok account. Each paired narration with visuals and subtitles on what the judgment describes as “trivia on urban legends, paranormal phenomena, and the like.”
Tsuda argued that the narration resembled his voice and sought removal under either Japan’s Unfair Competition Prevention Act or the right of publicity.
A website run by the uploader, however, carried a statement to the effect that the narration was not made by training an AI on Tsuda’s voice without permission, but by training it on the voice of a friend who is good at voice impressions. The court recognized only that this statement appeared on the website; it made no finding that a friend’s voice was in fact used. The judgment also notes that comments on the videos included viewers saying the voice sounded like Tsuda, and others doubting the uploader’s explanation.
Primary source (in Japanese): Tokyo District Court, Judgment of September 30, 2026, Case No. 2025 (Wa) 70635
The Court’s General Principle: Voice Is a Symbol of Personality
The most closely watched part of the judgment was the court’s statement that a person’s voice, like their portrait, is a symbol of the individual’s personality.
Japan has no statute expressly codifying the right of publicity; the doctrine has developed through case law. The Supreme Court first addressed it squarely in its 2012 Pink Lady decision, describing it as the right to exclusively exploit the customer appeal of a person’s name, portrait, and similar attributes. The Tsuda ruling expressly places voice within that framework.
The court then set out a general test: the unauthorized use of, for example, a performer’s voice infringes the right of publicity where the voice has customer appeal that promotes the sale of goods or services, and the use is made for the sole purpose of exploiting that customer appeal.
The court did not, however, decide whether the narration in the 188 videos met that test. The significance of the ruling lies not in drawing an infringement line around specific audio, but in giving voice a legal foothold within the right of publicity.
Why Was the Removal Claim Dismissed?
The uploader deleted the account on May 29, 2026, while the case was pending. Because no request to restore it was made within 30 days, the court found that the account data, including the videos, had been deleted from TikTok’s servers as of the end of June 28, 2026. Tsuda argued that published data-retention periods do not always match actual practice, so it was unclear whether the data was truly gone, but the court rejected that argument.
TikTok had also argued that, because the videos no longer existed, Tsuda lacked a legal interest in bringing the suit and the case should be thrown out. The court rejected that argument too. Whether the content still existed, it held, was not a question of Tsuda’s interest in bringing the suit but a substantive question of whether a right to removal actually existed.
It then concluded that, even assuming TikTok could be subject to a removal obligation, there was nothing left to remove, and dismissed the claim on the merits.
The judgment states expressly that the court would not decide the remaining issues. It did not rule on Tsuda’s claims under the Unfair Competition Prevention Act, or on whether TikTok was under any obligation to remove the posts.
The Ministry of Justice Had Already Issued Guidance on Voice
On August 7, 2026, Japan’s Ministry of Justice published a report by its study group on civil liability for the unauthorized use of portraits, voices, and similar attributes, subtitled “Interpretive Guidelines on Infringement of the Right of Publicity by Generative AI” (English title ours). Prompted by concerns that the unauthorized use of portraits and voices through generative AI has become serious, the report sets out how existing law and case law apply to the right of publicity, the right not to have one’s portrait and similar attributes used without justification, and the Unfair Competition Prevention Act.
The report says that a voice both identifies a person and symbolizes their personality, and that there was no dissent within the study group that voice falls within the scope of the right of publicity. It also says that nationwide fame is not required for customer appeal: a voice can have it within a particular field or region.
Where the court’s ruling stayed at the level of general principle, this report offers more concrete guidance for people actually making and deploying AI voices.
Reference (in Japanese): Ministry of Justice, Japan: Publication of the study group report on civil liability for unauthorized use of portraits, voices, etc.
If the AI Never Trained on the Person, Can It Still Be a Problem?
For AI creators, the most striking detail in the case is the uploader’s claim that the AI was trained on a friend who is good at voice impressions. Even if that were true, would it matter if the output still sounded strikingly like Tsuda? The court did not decide that question, but the Ministry of Justice report offers a framework.
What matters is who listeners think is speaking
Where a voice resembles a real person, the report says the analysis should weigh the degree of similarity together with any additional cues that bring that person to mind, and ask whether customers would identify the voice as theirs (pp. 26–27). Scientific tools such as voiceprint analysis can be one factor, but they are not decisive (footnote 56).
The test, in other words, is not acoustic similarity alone. It is who the listener believes they are hearing.
Undisclosed impersonation can cross the line
The report notes that impersonators ordinarily perform under their own names, so impressions are generally not treated as use of the original person’s “voice” and do not infringe the right of publicity.
However, where the impersonator’s name is not disclosed, a voice resembling someone with customer appeal is used, and audiences may take it to be that person, there is room for infringement (p. 27). The same logic may apply to AI voices: how a voice was technically made may matter less than whose voice it is presented as.
Keeping the person out of the training data doesn’t eliminate the risk
On portraits, the report states that even where neither the training data nor the prompt includes the person’s actual likeness, a similar image can still be produced by adjusting prompts, and that using such an image for the sole purpose of exploiting the person’s customer appeal may infringe the right of publicity (footnote 194).
That passage concerns portraits, and no court has ruled on the point for voice. Still, the practical lesson is clear: “we didn’t train on that person” is not, on its own, an answer to every legal question about commercially using their identity.
Does Labeling It “AI Voice” Make It Safe?
Disclosures such as “This is not the real person” or “This voice is AI-generated” matter, but their legal effect depends on the right involved. In this case, TikTok argued that because the uploader’s website made clear the voice was not Tsuda’s, there could be no confusion. As noted above, the court did not decide that point.
The Ministry of Justice report explains that under Article 2(1)(i) of the Unfair Competition Prevention Act, which covers causing confusion with a well-known indication, a disclaimer, such as a note on photo merchandise or calendars stating that the product has no connection to the person pictured, may mean the confusion requirement is not met (p. 74). By contrast, Article 2(1)(ii), which covers misappropriating a famous indication, does not require confusion at all. Tsuda relied on both provisions.
The right of publicity raises a further question. It is not limited to whether viewers mistook the content for the real person; it also asks whether the use is made for the sole purpose of exploiting that person’s customer appeal.
Our editorial view, then, is that disclaimers can reduce the risk of confusion, but they should not be treated as a complete answer under either unfair competition law or the right of publicity.
Five Practical Checks for Anyone Creating AI Voices
1. Check not just the training source, but who the output sounds like
Even if the person’s own recordings were never used, the result may strongly evoke a specific individual. For commercial use, review where the data came from, how similar the final output is, and how it is presented.
2. Ask whether “sounding like them” is the selling point
If a voice draws views because it sounds like a particular celebrity, or lifts ad performance because audiences associate it with that person, the use may be trading on their customer appeal.
3. Look at how the voice is monetized
Ad revenue, platform payouts, product sales, customer acquisition, and promotions all raise the commercial stakes and call for closer scrutiny.
4. Don’t rely on disclosure alone
Labeling content as AI-generated is an important transparency measure, but it does not by itself resolve every rights issue.
5. If you are deliberately recreating a real person, build on permission
When a commercial AI voice is designed to evoke a real actor, voice actor, narrator, singer, influencer, or other identifiable person, the safer approach is to structure the project around permission from that person or the appropriate rights holder.
VTubers need a more nuanced look. The report notes that the voice of a fictional character, where it is not a real person’s voice, does not carry customer appeal protected by the right of publicity. But where a performer’s voice is identifiable as a real person’s and their customer appeal is being exploited, further analysis is needed.
AI Creators’ View: Voice Is Becoming an Identity Asset
AI voice has mostly been discussed as a technical question: how to generate convincing speech from data. Read alongside the Ministry of Justice report, however, the Tsuda ruling makes clear that a voice is more than sound. It carries personal and economic value, identifying an individual and conveying their reputation, popularity, and persona.
As AI makes it easier to produce a voice that sounds like someone without copying their recordings, the central question moves beyond training data. Who does the audience hear? Whose value is being used?
Editor’s Note
Rights management in the AI era can no longer stop at asking what went in. We also need to ask whose personal and economic value the final output invokes.
For AI creators, building that judgment into the production process is becoming part of the work itself.
Conclusion: It’s Not Only About Whose Data Was Used
The Tokyo District Court stated, as a general principle, that a person’s voice, like their portrait, is a symbol of the individual’s personality and may be protected by the right of publicity in certain circumstances. It did not find that the narration in this case infringed Tsuda’s rights, and it did not decide the Unfair Competition Prevention Act claims or whether TikTok had any obligation to remove the posts.
The practical issues for AI voice users are nonetheless coming into focus: training source, final output, similarity, presentation, purpose, monetization, and permission. Anyone using AI voices in commercial content should treat these as a single design question, not a checklist of separate ones.
This article provides general information based on the judgment and public materials and is not legal advice on any specific matter.
Key References
- Tokyo District Court – Judgment of September 30, 2026, Case No. 2025 (Wa) 70635 (in Japanese)
- Ministry of Justice, Japan – Publication of the study group report on civil liability for unauthorized use of portraits, voices, etc. (in Japanese)
- Ministry of Justice, Japan – Full text of the study group report (in Japanese)
- Supreme Court of Japan – Pink Lady judgment of February 2, 2012 (English translation, WIPO Lex)
- Bengo4.com News – Report on the Tsuda ruling (in Japanese)
