China's Layered Legal Architecture for Avatar and AI Intellectual Property

China has written the most concrete rules anywhere for owning and controlling artificial avatars and the works they generate, and it has done so without passing a single law aimed at either. There is no Chinese statute on artificial intelligence, no dedicated code for digital humans, no legislative act that settles who holds rights over a face conjured by a diffusion model or a voice reconstructed from stolen recordings. What exists instead is an architecture assembled by accretion — civil-law provisions drafted before the generative era, administrative regulations issued in rapid succession, court judgments that read old statutory language against machines their drafters never imagined, and a thickening scaffold of technical standards and enforcement campaigns. The result is not a gap waiting to be filled but a working system, and its logic is best grasped as a set of layers, each answering a different question about what it means to own a synthetic person.

The first question is authorship: when a machine produces an image, who, if anyone, made it. The foundation here is the Third Amendment to the Copyright Law, adopted in November 2020 and effective June 2021, which replaced a closed list of protected works with an open definition — intellectual creations with originality expressible in a certain form — and added a catch-all for other qualifying creations. The amendment never mentioned artificial intelligence, and it imposed no requirement of substantial human creativity; the statutory word is simply originality. The demanding standard that Chinese courts now apply is therefore a judicial construction, not a legislative one. Its landmark expression came from the Beijing Internet Court in November 2023, in Li Yunkai v. Liu Yuanchun. Li had generated a photorealistic portrait with Stable Diffusion, entering more than 150 positive and negative prompts, choosing models, tuning parameters, and iterating before publishing; when another user reposted it stripped of its watermark, Li sued. The court held that this selection and arrangement amounted to genuine intellectual input, that the software could not be an author because authorship belongs to natural persons, and that copyright therefore vested in Li. The damages were trivial — five hundred yuan — but the Supreme People's Court later elevated the ruling to its top nominated cases of 2024, giving it real precedential weight.

What the courts have not settled is where the line falls. Two rulings from Jiangsu province, weeks apart in early 2025, point in opposite directions. In one, a designer who spent hours refining a Midjourney image through Photoshop was granted copyright on the strength of his creative direction. In the other, a plaintiff who admitted she could not reproduce her own AI outputs — because the model's randomness put the result beyond her control — was denied protection on the ground that the expression belonged to the machine, not to her. The divergence exposes an unresolved fault line: whether originality is measured by the human's process of creative direction or by the degree of deterministic control the human holds over what finally emerges. It is the central open problem of AI copyright, and China's courts are working it out case by case rather than by decree.

The second layer treats the avatar not as a work but as a thing, and here Chinese doctrine performs a subtle split. When the Hangzhou Internet Court confronted the virtual human Ada in 2023, it ruled that the avatar could not be a performer and enjoyed no performer's rights of its own — it was, in the court's framing, weak AI wholly driven by human operators. Yet its visual design, the lines and colors and aesthetic choices embodied in it, was a copyrightable work of fine art, and every right in it belonged to the developer who made it. The avatar, in other words, is an object that is owned, not a subject that owns, and the courts have since repeated the point: a 2025 Beijing ruling against a former employee who sold a studio's avatar models without permission again treated the figures as artworks reflecting deliberate aesthetic judgment, with the rights running to their makers. But the same figure casts a second shadow, because an avatar can also resemble a real human being, and that resemblance triggers an entirely different body of law that protects not the maker but the person depicted.

That body is the Civil Code, effective January 2021, whose personality-rights provisions were written just before generative systems made them urgent. The Code protects a natural person's likeness against anyone who falsifies their image using information technology — language that now serves as the front-line weapon against deepfakes and unauthorized AI face-swaps — and, crucially, it dropped the older requirement that such infringement be commercial, so any unauthorized use can qualify. It extends the same protection to voice. That extension produced China's first AI voice-cloning judgment, decided by the Beijing Internet Court in April 2024: a professional dubbing artist discovered that a media company had handed her recordings to a developer who trained a text-to-speech product on them without her consent. The court's reasoning was pivotal. Because the synthetic voice remained identifiable as hers, her voice rights reached the AI reproduction — and those rights were independent of any copyright in the recordings, so the company's ownership of the audio conferred no licence to clone the person. The defendants paid 250,000 yuan and apologized, and the Supreme People's Court later marked the case as typical. The lesson is that owning a recording is not owning the person in it.

Above the courts sits the regulatory layer, and it is here that China's lead is most visible. The Deep Synthesis Provisions, effective January 2023, require that synthetic media carry visible labels and that anyone editing a real face or voice notify and separately obtain consent from the affected person. The Interim Measures for Generative AI Services, effective August 2023, were the first binding rules for the technology anywhere in the world, demanding that outputs uphold core socialist values, that training data meet quality standards, and that services capable of shaping public opinion register their algorithms and pass security review; by late 2025 more than seven hundred services had completed national filing. A 2025 labeling regime then mandated both human-visible marks and embedded machine-readable metadata across every modality. Most pointed of all is the draft Digital Virtual Human measures released for comment in April 2026, which would compel a prominent "digital human" label on such figures, require separate consent before a real person's sensitive data is used to model one, forbid using avatars to defeat facial or voice authentication, and bar the provision of virtual romantic or family relationships to minors. A companion draft on anthropomorphic AI interaction services, circulated at the end of 2025, reaches further into the same social terrain, targeting emotional-companion systems with disclosure duties and usage limits — a sign that the state now means to govern not only how avatars are made but how people are permitted to relate to them. Beneath all of it, the Personal Information Protection Law already classifies faces, voiceprints, and fingerprints as sensitive data demanding separate consent — the doctrinal root from which those consent requirements grow.

Supporting these layers is a scaffold of standards, identity infrastructure, and enforcement. National and industry standards for virtual humans and generative-AI security have proliferated; a blockchain-based real-name identity system lets users act pseudonymously while the state retains the ability to unmask them when legally required; and enforcement runs less through named fines than through the Cyberspace Administration's recurring Qinglang campaigns, one of which removed thousands of AI products and hundreds of thousands of pieces of content in a single 2025 sweep. What is conspicuously absent is the capstone. A comprehensive AI law was drafted, entered the legislative pipeline, and was then quietly dropped from the 2025 schedule; the amended Cybersecurity Law now carries the first AI provisions in national-level legislation, but the unified statute remains a proposal. This is less indecision than method. China's approach favors incremental, sectoral rules — targeted measures, pilots, and standards that manage specific risks while leaving room for an industry still changing shape. For anyone creating, licensing, or contesting rights over digital humans in this jurisdiction, the practical demands follow directly from the architecture: document the human hand in every generated work, secure consent for every real likeness and voice, meet the labeling and filing rules, and watch the courts, because in the continuing absence of a single law it is their accumulating judgments that give the whole structure its concrete meaning.