On 23 April 2026 the media disclosed details of China’s first AI model infringement case. The plaintiff operated a “transformation comic” special-effect product; the defendant launched a “shojo comic” special-effect. The court held that the two effects were similar, that the defendant’s product constituted a substantial substitute for the plaintiff’s and caused substantial harm to the plaintiff’s competitive interests, amounting to unfair competition, and awarded damages plus reasonable costs of RMB 1.6 million, upheld on appeal.
What is protected is not code but the competitive interest from investment
The reasoning deserves close reading. The court did not focus on copying of model code or weight files but on the chain “investment — effect — substitution”.
According to the disclosure, the court noted that an operator forms a particular model effect through data training and tuning, and that the other party replicates that effect at low cost and brings it to market, substituting the first product and harming its competitive interests. This evaluation path bypasses copyright’s strict requirement of “expression” and instead applies the general clause of the Anti-Unfair Competition Law.
For AI companies this opens a new enforcement channel and draws a new risk line. A model’s effect can, under certain conditions, be protected.
Two kinds of risk for model companies
First, the risk of being infringed. Effects, parameter configurations and data-processing pipelines formed through substantial investment are protectable competitive interests. Companies should retain training processes, iteration records and effect comparisons to prove the link between investment and effect. See our IP protection service.
Second, the risk of being accused. Model development aimed at “reproducing a competitor’s effect” carries markedly higher risk. Even without directly obtaining the other’s model files, approximating a competitor’s effect through reverse engineering, bulk testing or data feeding may be found to be unfair competition.
Boundary judgments in practice
First, correspondence of investment. Whether the protected effect derives from provable substantial investment is the starting point.
Second, substitution. Whether the defendant’s product substantially substitutes the plaintiff’s is key to finding harm. Overlapping functions, converging effects and overlapping target users all strengthen this.
Third, means of acquisition. Whether the other’s model capability, training data or tuning experience was obtained by improper means directly affects the finding of unfairness.
Fourth, market impact. Loss of customers, disruption of pricing or loss of market share is the basis for calculating harm.
Compliance suggestions
First, put “model effects” into the IP asset register. Alongside patents, trademarks and copyright, as assets to be managed and protected.
Second, keep a process archive of training and tuning. Data sources, training logs, version iterations and effect-evaluation records are the evidence for future claims.
Third, treat “benchmarking a competitor” cautiously. Development aimed at explicitly reproducing another’s product effect should proceed only after legal assessment, with evidence of independent R&D retained.
Fourth, watch the lawfulness of data sources. A model’s competitiveness increasingly depends on data, and lawful sourcing is both a competition and a data-compliance issue — see our data-compliance service.
IP protection for AI models is forming a new combination of rules: copyright governs expression, patents govern technical solutions, trade secrets govern undisclosed processes and data, and anti-unfair competition governs competitive interests formed through investment. A company’s asset register and evidence archive need to cover all four lines — see our IP protection service.
This article is based on public information concerning China’s first AI model infringement case disclosed on 23 April 2026. It is for general reference only and does not constitute legal advice.