China's Courts Are Quietly Writing the Rules for AI Voices and AI-Made Shows

Most of the loud debate about AI and rights happens in the United States and Europe, where megacases against OpenAI and Stability AI grind through the courts. Meanwhile, in China, a set of smaller rulings is building a practical framework for how likeness and authorship work when a machine does the generating. Two decisions in September are worth reading closely, because they map out where the lines are being drawn.

The Voice Case

On September 29, the Shanghai First Intermediate People's Court closed a case that is being called the city's first dispute over an AI-generated voice. The facts are simple. A woman identified as Wang is a voice actor verified on an app where she posts dubbing work for pay. In January 2025, a friend told her that a company, referred to as Company A, seemed to be using her voice in a customer recruitment campaign. She checked, concluded the audio was probably AI-synthesized from her recordings, notarized the evidence, and sued, asking for 300,000 yuan.

Company A did not deny that the audio was AI-generated. It said it could not confirm the source material because the responsible employee had left, and that it had not known Wang was a voice actor. During the trial, a forensic analysis compared the disputed audio with a fresh recording of Wang's voice. Across 28 acoustic indicators, 24 showed a deviation of less than 10 percent, and 16 of those were under 5.36 percent. The court described the similar and highly similar portions as reaching 90 percent.

The first instance ruled for Wang and awarded 50,000 yuan. Both sides appealed. The intermediate court upheld the finding, and its reasoning is the part that matters. Using a person's voice as training material without consent, and generating a synthetic voice that imitates the person's timbre, tone, and pronunciation style so that it can be identified as that person, infringes on that person's voice rights. The court also placed the burden of proof on the company: as the party controlling the original training data, it failed to show that the source was lawful, so it had to bear the consequence.

The takeaway for anyone building a voice product is uncomfortable for the careless. You do not need to have deliberately targeted a specific person. If the synthesis makes a listener connect the output to a real, identifiable individual, and you cannot prove where your training data came from, you are exposed.

The Short Drama Case

Five days earlier, a court in Wuhan ruled on a different question: does an AI-assisted short drama count as a work protected by copyright?

The case involved a 47-episode micro-drama, roughly an hour of runtime, produced by a company using generative AI tools. It had been filed with the national broadcasting regulator and published on platforms including Hongguo and WeChat Channels. A day after it went live, a second company copied the entire series onto its own channel, changed the title, and ran ads against it. The original producer sued for infringement of its right of information dissemination.

The core dispute was whether AI-assisted content can be a protected work at all. The court held that the test is not whether the final frames were generated by AI, but how deeply a human was involved in the process. In this case, the production went through script planning, prompt design, selection of character images and scene assets, filtering of AI-generated clips, editing, detail correction, and audio and subtitle matching. The court found that the creators had put in substantial, continuous intellectual labor and retained control over the final expression, so the drama was an audiovisual work under copyright law. The copier, which had lifted the whole series and profited from ads, lost.

The most interesting part is the damages calculation. The court said it considered the cost of the compute consumed during creation, including the tokens spent, and the licensing cost of the commercial tools, as factors in setting the award. In the end, it ordered 20,000 yuan. Putting token and tool costs into a copyright damages figure is unusual, and it signals that Chinese courts are starting to treat AI production expenses as a real part of what it costs to make a creative work.

The Pattern Across Both Cases

Read together, the two rulings lean on the same principle. The law still cares about human contribution and about identifiable individuals. A voice can be protected even when it is synthesized, if it points back to a specific person. A drama can be protected even when AI made the frames, if a human made the meaningful decisions.

That is a more pragmatic posture than the muddled fair-use fight in the US, where courts have split on whether training on protected work is transformative. China's Supreme People's Court added to its side of the ledger in September as well, issuing guidance that says AI-generated content accused of copyright infringement cannot escape liability simply because it was machine-generated. Responsibility is assigned according to who had control and what duty of care they owed. The guidance also put the burden on developers to show their training data sources when defending against infringement claims, and it left the biggest questions, including the copyright status of AI outputs and whether unlicensed training is infringement, deliberately unaddressed.

Why These Cases Went Quietly

It is worth asking why rulings like these get so little attention compared with the marquee lawsuits in the United States. Part of the answer is scope. The American cases are about whether an entire training pipeline is legal, which would affect every model provider at once. The Chinese cases are narrower, deciding the liability of one company that used one voice or copied one drama. Narrow decisions make less noise.

There is also a difference in how the systems handle dispute. China's courts have been willing to rule on concrete facts and leave the big theoretical questions to guidance, which is exactly what the Supreme Court did in September when it chose not to address the copyright status of AI outputs or the legality of unlicensed training. The upside is faster, more predictable answers for the parties in front of the court. The downside is that the hard questions stay open, and companies building products have to plan around uncertainty either way.

For foreign companies entering the Chinese market, the two September rulings are a practical heads-up. Voice cloning, likeness reproduction, and AI-assisted production all carry liability that Chinese courts are now willing to enforce, and the burden of proving lawful data sourcing sits with the company doing the generating. That is a higher bar than a simple "we did not mean to" defense, and it applies whether the company is a Chinese platform or an overseas one serving Chinese users.

What It Means for Creators and Builders

For creators, the message is that documentation is now a practical defense. The producer of the AI drama won partly because it could show the full chain of human decisions. A creator who wants to protect AI-assisted work should be able to point to the scripts, the prompts, the selection criteria, and the editing choices, not just the finished file.

For companies building AI voice or video products, the voice ruling is the more immediate warning. Consent and provenance are not optional paperwork. If your model can reproduce a recognizable voice and you cannot say where the training audio came from, the burden of proof sits with you, and it will not be a comfortable place to be.

Neither case settles the whole field, and bigger questions remain open by design in China as well as the US. But while the dramatic lawsuits elsewhere get the headlines, these quieter rulings are doing the work of defining what is allowed. For anyone shipping AI-generated voices, faces, or shows, the practical rules are now being written in ordinary courtrooms, one case at a time, and it is worth reading them as they come.