AI-generated Image Found Infringed for the First Time in China

Mr. Li operated Stable Diffusion to produce a picture and then posted it onto his own social media page. Ms. Liu was a blogger who posted an article in which she had inserted Li’s picture—titled The Spring Breeze brought Tenderness (“the image”)—as an illustration. The picture appeared in Liu’s blog without Li’s consent and with the watermark removed. Li sued Liu for copyright infringement for CNY 5,000 and injunctive relieves at the Beijing Internet Court.[1]

A work subject to copyright protection refers to an intellectual product that is original in the fields of literature, art or science and is expressed in a certain form. This definition of a copyrighted work was instrumental in how the court decided to approach this case. Several main questions were raised: whether the image constitutes a work under the definition of copyright law; whether the Plaintiff Li holds the copyright for the image; and whether the defendant’s use constitutes infringement of copyright with liabilities. The majority of the court’s reasoning focused on question number one. Accordingly, the court proceeded to analyze the respective elements.

Intellectual Production

Intellectual production refers to any product resulting from a human’s intellectual activities, the court pinpointed. A work shall faithfully demonstrate the intellectual investment of human beings. Stable Diffusion is a model trained by a myriad of pictures and corresponding descriptive text which is able to yield an image with corresponding traits that the prompt instructions offer by adopting the homology of the syntax of the prompt and the pixels in the pictures. It does not merely take available pictures straight from online search engines or perform combinations of several factors predetermined by the programmer. Specifically, Stable Diffusion uses accumulated capability to delineate, contour and draw the lines in place of human manual tasks.

In the present case, the Plaintiff prompted Stable Diffusion, set up relevant parameters, added other prompt terms and adjusted further parameters in a first draft, before finally choosing an image to his satisfaction. The court found that throughout the process, from the conception to the decision, the Plaintiff had devoted his intellectual activities to, for instance, the style of the presenting figure, the choice and orders of prompts, the arrangement of parameters and finally the choice of an ideal output image. Therefore, the element of intellectual production was met.

Originality

A work must be independently completed by a producer with the producer’s personalized expression, the court reasoned. On the contrary, a mechanical display of an intellectual product is deemed to be lacking originality, for example, a product completed in a particular order or displaying a distinct formula or structure.

In the use of Stable Diffusion, the more distinctive the input prompt procedure is and the more clearly defined the description of element layouts is, the more an output demonstrates personalized expression. It is to some degree correct to deduce that the image was drawn by Stable Diffusion; however, the Plaintiff by himself designed the female figure and its presentation via his choice of prompts and rearranged the image’s layout through parameters, notwithstanding his further modification of the image with further prompts and other parameters. All of these activities were a reflection of the Plaintiff’s aesthetic judgments and personal choices in creating the final image. In the court’s hearing, the Plaintiff operated the same generative AI model on-site to produce different images by inputting different prompts and parameters. That is, clearly the image in dispute was not a mere mechanical display of an intellectual product; it was one displaying originality.

Work in artistic field

A generative AI model which has no free will is not a subject at law, the court emphasized. When a person uses a model to produce an image, this is essentially no different from using a tool to make creations. In other words, it was the human who dedicated intellectual investments rather than the model. Encouragement of creation is the core policy purpose of copyright law. So long as an image—despite being generated by a machine—can present a person’s original intellectual investments, the image shall be considered a work protectable by copyright law.

Besides, the image in dispute was a piece of planar creation made up of lines and color and with aesthetic implications. Hence, it is considered an artistic work.

Ownership

In principle, the copyright of a work belongs to the author. Stable Diffusion was merely a tool that the Plaintiff used to create an image, as forementioned. The designers of Stable Diffusion did not pre-determine the output of the model, nor were they involved in the steps of image generation. Furthermore, the designers of Stable Diffusion disclaimed any rights in association with the outputs. In view of the above, neither the Stable Diffusion model nor the designers were the authors of the image in dispute.

On the contrary, the Plaintiff was the person who gave direct prompts, set parameters and chose the image. The image was produced as a result of the Plaintiff’s intellectual input and thus demonstrated the Plaintiff’s personalized expression. The Plaintiff authored the image and therefore was the owner of the image.

The image in dispute was a protectable work as per copyright law. The court continued to the infringement determination.

The court found the Defendant to have infringed the Plaintiff’s right of paternity and the right of information network dissemination through their use of the image and removal of the watermark on the same. As a result, the court awarded light damages amounting to CNY 500 and also ruled the defendant to issue a public apology in her blog.

From a comparative perspective, notably, this trial decision turned out to be in marked contrast to the stance of the US Copyright Office (“USCO”) in February 2023. In the Zarya of the Dawn Letter of the USCO, it is quoted that “[A] person who provides text prompts to Midjourney does not ‘actually form’ the generated images and is not the ‘master mind’ behind them.” “The information in the prompt may ‘influence’ the generated image, but prompt text does not dictate a specific result.” The USCO emphasized that there is a huge difference between a user’s prompts and the images that an AI model actually produces. Thus, the user, lacking meaningful control of the nature of the outputs, is not the output’s creator.

Returning to the present case, it was apparent that the scientific theories of an AI model’s generation of outputs were not sufficiently reasoned so as to offer a convincing analytical basis. This could form the basis of one of several arguments in a further challenge from the defendant. However, up until the deadline of appeal no parties filed for an appeal. The case is therefore finalized and turned effective.

[1] (2023) Jing-0491-MinChu-No. 11279 (2023.12.27)