On May 12, 2026, China's State Council released its annual Legislative Work Plan, and for the first time it placed AI copyright squarely on the national agenda. The plan flags preparatory revisions to the Implementing Regulations of the Copyright Law — expected to address AI-generated content and training-data questions — and commits to accelerating comprehensive AI legislation spanning data, algorithms, computing power, and intellectual property (MLex; SCMP). It is the clearest signal yet that Beijing intends to convert a patchwork of court rulings into codified rules.
This matters because China has, until now, made its most consequential AI-copyright law from the bench. The question for 2026 is whether codification preserves the pragmatic, creation-friendly approach Chinese courts have built — or freezes it into something more rigid.
What the courts already decided
China's judiciary has not waited for legislators. In Li v. Liu (case (2023) Jing 0491 Min Chu No. 11279), the Beijing Internet Court ruled on November 27, 2023 that an image generated with Stable Diffusion was a copyrightable work, with the human prompter as author. The court found that selecting and arranging prompts, setting parameters, and iterating toward a final image reflected the plaintiff's "aesthetic choices and personal judgment" — enough originality to qualify (China Justice Observer).
That reasoning has hardened into a line of authority. By 2025, Chinese courts had issued roughly five decisions extending copyright to AI-assisted works where a human contribution — typically at the prompt-engineering stage — was meaningful (Authors Alliance). On the output-liability side, courts have policed misuse without strangling the underlying technology: in the Hangzhou "Ultraman" case, the Intermediate People's Court found a generative-AI platform liable for contributory infringement of the right of communication, awarding around 30,000 yuan, because it should have known users were producing infringing images and failed to act (EU IP Helpdesk).
The pattern is telling: block specific infringing outputs through technical measures, but do not negate the training process or order training data deleted. That is proportionality in action.
The case for codifying — taken seriously
The strongest argument for moving from case law to statute is real. Court-by-court development produces inconsistency: a Beijing ruling does not bind a Hangzhou or Guangzhou court, and outcomes can turn on how a particular judge weighs "intellectual investment." Rights-holders, AI labs, and platforms all face uncertainty about what is protected, who is liable, and what counts as a legitimate training-data source. The 2013 Implementing Regulations predate the generative-AI era entirely. A clear national rule on authorship and training-data lawfulness would lower compliance costs and reduce litigation risk — genuine benefits for an industry where DeepSeek alone is reportedly closing a $7.4 billion round at up to a $59 billion valuation (Silicon Republic). Predictability is itself pro-innovation.
Where codification could go wrong
But a statute is only as good as the standard it locks in, and two risks stand out.
First, the training-data question. China's existing Interim Measures for Generative AI Services (in force since August 2023) already require providers to use data from "legitimate sources" and respect intellectual property — but notably softened a draft provision that would have made providers liable for the legitimacy of all pretraining data. That restraint reflected a "moderate leniency" toward upstream training while focusing enforcement on downstream outputs. If the revised regulations swing toward strict, opt-in licensing for every training input — the maximalist position pushed in some Western debates — China would burden exactly the labs it is racing to scale, with no offsetting gain for most individual creators.
Second, the authorship standard. China's courts have been generous: meaningful prompt selection and iteration can create copyright. That stance is criticized abroad — the US Copyright Office takes the categorical view that prompt engineering does not yield foreseeable, human-controlled outputs and so cannot ground authorship (Authors Alliance). Critics warn the Chinese approach risks "copyright saturation." The concern deserves a fair hearing. But the better answer is a calibrated originality threshold — protecting genuine human creative direction while denying protection to trivial one-click prompts — not a US-style blanket exclusion that leaves creators with nothing.
The proportionate path
The Implementing Regulations are still at "reserve-project" status — preparatory, not imminent (MMLC Group). That gives drafters room to get the details right. The goal should be to codify what is working: human-centered, threshold-based copyrightability for AI-assisted works, and output-focused liability that targets infringing uses rather than the act of training.
China's comparative advantage in AI is partly regulatory: it has so far regulated harms without pre-emptively criminalizing the technology. Codification is a chance to make that posture durable and predictable. It is also a chance to squander it, if the new rules import the rigidity Beijing's own courts have wisely avoided. The smart move is to write the case law's pragmatism into statute — and resist the temptation to over-correct.