AI 仿聲催生數碼分身新市場，個人聲音權益面臨挑戰

Here is my new blog about peoples` voices and the way AI can steel them. It was published in Telegraf news site.

Just a few years ago, the phrase “stealing someone's voice” sounded more like a metaphor. Today, all it takes is a few minutes of recorded audio, access to an AI model, and minimal technical skills.

That is why the story of Japanese voice actor Kenjiro Tsuda is interesting far beyond the anime fan community.

Tsuda has one of the most recognizable voices in Japan. He is known for his roles in Jujutsu Kaisen, Yu-Gi-Oh!, and other popular productions. But a few months ago, his voice began to take on a life of its own.

A TikTok account appeared, publishing videos about urban legends, the occult, and conspiracy theories. The videos featured an AI-generated voice that sounded like Tsuda's. The actor himself had never read any of these scripts and had no connection to the account.

That did not stop the page from attracting more than 200,000 followers. According to an estimate cited by AFP, the account could have been generating up to 500,000 yen, approximately $3,200, per month for its owner.

The result was a rather strange, yet distinctly modern, business model. One person spends decades building a career, developing a distinctive voice, and becoming recognizable. Someone else takes that recognition, uploads voice samples into an AI model, and starts making money.

What makes this particularly interesting is that, technically speaking, no conventional theft appears to have taken place. Nobody stole the actor's original studio recordings or distributed a pirated copy of a film. The AI generated entirely new audio recordings containing phrases Tsuda had never spoken.

This is precisely why addressing such cases under existing legislation has proven so difficult.

Stealing a Person Without Stealing Their Content

Over the past few years, discussions about deepfakes have focused primarily on video. A politician appears to make a statement they never made. A celebrity appears in an advertisement for a questionable product. A company executive suddenly joins a video call with employees.

With voice cloning, the situation may be even more complicated.

We still perceive the human voice as a relatively reliable identifier. A message supposedly sent by a company's CEO may raise suspicion. But a phone call in which an employee hears the familiar voice of their boss is perceived very differently on a psychological level.

Meanwhile, the technological barrier to such manipulation is rapidly disappearing.

In the past, convincing voice imitation required a skilled impersonator. Then came sophisticated software solutions. Today, AI models can produce realistic voice replicas using relatively small amounts of source audio.

For public figures, obtaining that audio is hardly a challenge.

Interviews, podcasts, conference presentations, YouTube videos, TikTok clips, television appearances. The more famous a person is, the more material is already publicly available to create their digital replica.

This creates an interesting paradox. Public visibility has always been an asset. Now it is becoming a vulnerability at the same time.

Who Owns a Voice?

Copyright is relatively straightforward. There is a photograph, a musical composition, a film, or a piece of writing. Each has its creators and rights holders.

The situation with the human voice is far more complicated.

A person does not own the copyright to their natural vocal timbre in the same way a musician owns rights to a particular recording.

Modern voice-cloning technologies operate precisely within this legal gray area.

Not a single second of the original recording needs to appear in the final product. An AI model generates an entirely new audio file, making a synthetic voice say words that the real person never uttered.

Yet the commercial value of that file exists precisely because audiences recognize a particular individual.

This was the central issue in Tsuda's dispute with TikTok.

The platform argued that the audio featured a “typical male voice” and that any resemblance to the actor's voice was subjective.

For lawyers, this is a dispute over identification. For the artificial intelligence industry, the question is much broader.

Where is the line beyond which a synthetic imitation becomes a replica of a specific person?

On September 30, the Tokyo District Court took an important step toward answering that question.

Tsuda's request to have the content removed was rejected because the disputed account and its materials had already disappeared from the platform by the time the case was heard. However, the more significant development was that the court recognized that a person's voice could, in principle, be protected under publicity rights.

For the industry, this sends an important signal.

We are gradually approaching the point where a new category of digital property will need to be legally defined. This is no longer about works created by a person, but about the characteristics of the person themselves that technology has learned to reproduce.

Voice, face, facial expressions, speaking patterns, signature phrases, writing style.

Today, we discuss each of these elements separately. Within a few years, AI models may be able to combine them into a fully developed digital double.

At that point, the question of who owns a voice will seem rather narrow.

A much more important question will emerge: who owns a person's digital identity?

This Is Not Just a Problem for Actors

It is easy to see Tsuda's story as a celebrity problem. Actors, singers, and influencers have recognizable voices and faces, so they need to protect them from unauthorized replication.

But the corporate world will face exactly the same challenge.

Imagine the CEO of a large company who regularly gives interviews and speaks at conferences. Several hours of recordings featuring their voice can easily be found online.

That is already enough to create a convincing voice model.

Then the company's CFO receives a phone call from the “CEO,” urgently requesting a financial transfer. And this does not necessarily involve a prerecorded message. Modern systems can synthesize speech almost in real time.

The next stage is obvious. Add a face to the voice.

And at some point, the familiar phrase “I saw and heard them myself” will no longer be sufficient proof that the conversation actually took place with that person.

For businesses, this presents a serious challenge. A significant share of corporate security procedures is built around verifying information. Now companies will also have to verify the identity of the person providing that information.

After Generation Comes the Verification Market

For a long time, the internet operated on a fairly simple principle. If a photograph, audio recording, or video looked authentic, we tended to assume it was genuine until evidence suggested otherwise.

Generative AI is gradually breaking that model.

Soon, the mere existence of a video will prove very little. The same applies to audio recordings and screenshots.

That is why another market will grow alongside the enormous generative AI industry: technologies for verifying authenticity.

Digital signatures, cryptographic proof of content origin, labeling of synthetic materials, corporate identity verification systems, and specialized protocols for authorizing financial transactions.

Businesses will have to implement these tools faster than comprehensive legislation can emerge.

Laws almost always move more slowly than technology. While regulators are still defining the rules for voice cloning, AI models are already learning to reproduce a person's voice, appearance, and behavioral patterns simultaneously.

Authenticity Is Becoming a Commodity

Generative artificial intelligence has an important economic effect that is often overlooked in public discussions.

It dramatically reduces the cost of content production.

Text already costs almost nothing to generate. Images are becoming inexpensive. The cost of voice generation is falling. The same is happening with video.

When producing something becomes cheap and accessible on a massive scale, value shifts elsewhere.

If virtually anyone can create a realistic video of another person, the video itself will no longer be particularly valuable. What will matter is the ability to prove that it is authentic.

This is why Kenjiro Tsuda's story is far more significant than an ordinary dispute between an actor and a technology platform.

It illustrates a problem that will affect not only celebrities in the coming years. Entrepreneurs, journalists, experts, corporate executives, and eventually ordinary users will all have to deal with it.

We are accustomed to protecting passwords, documents, bank cards, and personal data. Soon, we will have to add our own voices, faces, and communication styles to that list.

AI has made copying human identity cheap and scalable.

Now we will have to learn how to prove which one is the original.