In February 2024 the Guangzhou Internet Court delivered a first-instance judgment in a copyright dispute over a generative AI service. The case has been described as the first effective judgment holding a generative AI service liable for copyright infringement. Images generated by the defendant’s AI drawing function were substantially similar to the Ultraman character. The court found infringement of the rights of reproduction and adaptation, ordered the defendant to stop the infringement and pay RMB 10,000, and required keyword filtering and similar measures.
The sum is modest, but the judgment puts the duty of care of a generative AI provider on an operational footing.
Facts and holding
The defendant offered an AI drawing service through its website; users could generate images by entering prompts. After comparison, the court held that the output images were substantially similar to the protected work and fell within reproduction and adaptation. The remedy went beyond an injunction and damages: the platform was required to adopt filtering measures. The court did not deal with a single image but with the path by which similar images could be generated repeatedly.
That focus on service design matters. If the same kind of prompt can reproduce similar results, deleting one image does not solve the problem.
Why reproduction and adaptation
The generation process involves learning from and reproducing training material. The court looked at the relationship between the output and the protected work, not at whether the model internally “stores” the original image. The test for substantial similarity remains the comparison an ordinary observer would make. A provider cannot deny infringement risk by saying the model merely generates probabilistically.
The distinction between reproduction and adaptation turns on whether the output retains the basic expression of the work and whether new creative expression is added. For a platform, the practical response is not to litigate the category image by image but to reduce risk at both the input and output ends.
A reasonable and affordable duty of care
The judgment also stressed that a service provider should perform a reasonable and affordable duty of care, and that the burden should not be excessive. That formulation matters. It accepts that a platform cannot manually review every output, but requires technical and managerial measures commensurate with its capability, such as keyword filtering, complaint handling and blocking repeat infringement.
The scope of the duty depends on the service. A service that generates and publicly distributes content differs from one that merely supplies a tool for users’ own use. The charging model also matters: membership fees and per-generation charges mean the provider obtains direct economic benefit from the generated content.
What providers should do
First, manage training data and output filtering separately. Provenance and licensing of training data are an input-side issue; filtering, complaints and takedown are a continuing operations issue. One cannot substitute for the other.
Second, make notice-and-takedown work. A complaint channel, handling deadlines, review records and repeat-infringement blocking must form a closed loop, not just a published email address.
Third, set out usage boundaries in the terms of service. Prohibit clearly infringing uses, retain the ability to restrict abnormal accounts, and keep output records within a reasonable scope.
Fourth, maintain a priority list for well-known characters, personas and styles, and allocate filtering resources to them first.
For copyright and training data handling, see our IP protection service; for output infringement and platform liability assessment, see AI infringement.
This article is based on the February 2024 first-instance judgment of the Guangzhou Internet Court and public reporting. It is for reference only and does not constitute legal advice.