Two recent decisions regarding copyright and generative artificial intelligence (AI) handed down by Chinese courts are notable. In one, a court found that output created using Stable Diffusion could be subject to copyright. In the other, a court found that output generated by a generative AI system could infringe the reproduction and adaptation right under Chinese copyright law. These decisions raise interesting issues but leave many questions that go to the core of our copyright frameworks unanswered.
Copyright in computer generated output
A decision of the Beijing Internet Court ruled that a picture generated by an artificial intelligence (AI) model called Stable Diffusion
was a work of fine art and that the plaintiff, who used the model to create the picture, was the author and copyright owner of the picture.[i] The court also found that the defendant, a blogger who used the picture without permission had infringed the plaintiff’s rights in the work.
The plaintiff used Stable Diffusion to generate a picture of a beautiful woman (depicted here) under the dusk light. The court examined whether the picture generated by the plaintiff using Stable Diffusion constituted a work and what type of work it was.
The court found that the picture was a work of fine art, as it was a graphic art work composed of lines and colors and was of aesthetic significance.
The court also found that the picture reflected the plaintiff’s intellectual investment and personalized expression, as he designed the presentation of the character, selected prompt words, arranged the order of prompt words, set parameters, and selected the seed picture that he wanted to start from.
The decision explains in detail the specific and detailed prompts, parameters, weights, and seeds used by the plaintiff including prompts to successively modify outputs to produce the ultimate image.
The court accepted the plaintiff’s claim that image created reflected an original work of art. It found as a fact that the plaintiff was able to generate a desired picture by changing the prompt words, parameters and other input commands.
The Chinese court did a nuanced analysis that examined the control the plaintiff had over the creation of the image and found that it met the standard of originality under Chinese law stating:
In layman’s terms, the Stable Diffusion model works in a way that a human does: it acquires some abilities and skills through learning and accumulation, and it can generate a picture based on the text descriptions input by humans – drawing the lines and doing the colors, and presenting man’s creative ideas in a tangible way. In this case, the plaintiff wanted a close-up of a beautiful woman under dusk light, so he entered the following prompt words into the Stable Diffusion model: “ultra photorealistic” and “color photo” for the art type; “Japan idol” for the subject, along with detailed description of the character such as skin, eyes, and braid color; “in locations”, “golden hour”, and “dynamic lighting” for the environment; “cool pose” and “viewing at camera” for the way the character is presented; and “film texture” and “film simulation” for the style. The parameters were also set. Based on the initially generated picture, the plaintiff added some prompt words, modified the parameters, and finally got the picture he wanted. From the time the plaintiff had an idea about the picture to his final selection of the picture involved, the plaintiff did some intellectual investment, such as designing the presentation of the character, selecting prompt words, arranging the order of prompt words, setting parameters, and selecting the picture that he wanted. The picture involved reflects the plaintiff’s intellectual investment, so it meets the element of “intellectual achievement”…
Generally speaking, “originality” requires that the work be completed independently by the author and reflect the author’s personalized expression. “Mechanical intellectual achievements” are excluded. For example, if a work is completed based on a certain order, formula, or structure, different people will get the same result; as the expression is singular, the work does not have originality. And one has to decide according to the specific situation whether an AI-generated picture reflects the author’s personalized expression. Generally speaking, when people use the Stable Diffusion model to generate pictures, the more different their needs are and the more specific the description of picture elements, layout, and composition is, the more personalized the picture will become. In this case, there are identifiable differences between the picture involved and the prior works. In terms of the generation process of the picture involved, the plaintiff did not draw the lines himself, or instruct the Stable Diffusion model everything on how to draw the lines and do the colors; the lines and colors that constitute the picture involved are basically done by the Stable Diffusion model, which is very different from the conventional way of people using brushes or software to draw pictures. However, the plaintiff used prompt words to work on the picture elements such as the character and how to present it, and set parameters to work on the picture layout and composition, which reflects the plaintiff’s choice and arrangement. The plaintiff input prompt words and set parameters and got the first picture; then he added some prompt words, modified the parameters, and finally got the picture involved. Such adjustment and modification also reflect the plaintiff’s aesthetic choice and personal judgment. During the trial, the plaintiff generated different pictures by changing the prompt words or the parameters. One can infer that with this model, different people can generate different pictures by entering different prompt words and setting different parameters. Therefore, the picture involved is not a “mechanical intellectual achievement”. Unless there is contrary evidence, it can be found that the picture involved is independently completed by the plaintiff and reflects the plaintiff’s personalized expression. In summary, the picture involved meets the element of “originality”.
Comments on finding that copyright subsisted in the generative AI image
While the court concluded that copyright could subsist in the image it did not analyze the scope of that copyright. As Prof. Mark Lemley pointed out in an article on the topic, it is possible that the copyright in generative AI content would be narrow as it would not extend to expression generated by the AI system, but likely only to the expression over which the user could demonstrate it causally produced. Thus , while the decision appears to recognize the possibility that a copyright may subsist in the use of generative AI as a computer assisted tool, it does not examine the even more complex question as to the scope of protection in such works.
The decision of the Chinese court appears to be a departure from the approach being taken by the U.S. Copyright Office guidance which states that a work created using generative could be protectable by copyright if the “AI contributions are the result of …an author’s ‘own original mental conception”. However, based on the office’s understand of the technology “users do not exercise ultimate creative control over how such systems interpret prompts and generate material” and thus are not eligible for copyright protection over this content. Accordingly, the Copyright Office has rejected AI-generated content even when presented with evidence that the content was created using hundreds of prompts, including most recently in the Suryast decision.
The U.S. Copyright Office approach to the copyrightability of content produced using generative AI tools has been criticized as being inconsistent with the copyright laws including the law in the U.S. and the European Union.[ii] Generally, the test for originality recognizes that copyright can subsist if a person exercises a sufficient level of control over a tool that rises to the minimal level of originality required for copyright subsistence, whether it a minimal amount of creativity under U.S. law, a minimal amount of skill and judgement under Canadian law, or the minimal requirements under other copyright laws such as those in the European Union which requires a sufficient degree of human intellectual effort.
The approach taken by the U.S. Copyright Office also appears to be at odds with the draft Inventorship Guidance for AI-Assisted Inventions issued by the US Patent and Trademark Office. There the USPTO expressed the view that although a human being must be an inventor for an invention to be patentable, a human being can be an inventor if the individual makes a significant contribution to the invention. This can “be shown by the way the person constructs the prompt in view of a specific problem to elicit a particular solution from the AI system.” The position of the USPTO may reflect a different view on the question of whether there is room for a significant human contribution when an invention is conceived and reduced to practice using AI tools.
The court apparently also did not address whether the defendant would have infringed any moral rights of the plaintiff or any violation of rights in copyright management information, if any.
Infringement by computer generated output
The Guangzhou Internet Court issued a landmark judgment in China’s first case concerning the infringement of artificial intelligence-generated content (AIGC). The case involved a text-to-image AIGC service provider that generated images based on the text descriptions of its users. As depicted below, some of the images were seen as being substantially similar to the character Ultraman, a famous Japanese cartoon character. From summaries of the case, the court ruled that the AIGC provider infringed the exclusive licensee’s rights of reproduction and adaptation and ordered it to stop the infringement and pay compensation.[iii]
The court found that the AIGC provider had access to the original works of the Ultraman series through its online database and that the AIGC output shared the same or similar expressions of the original works, such as the appearance, posture, and background of the characters. The court also found that the AIGC provider adapted the original works by generating new images based on the users’ text descriptions, which altered the original expressions and created adaptations (derivative works). Therefore, the court concluded that the AIGC provider reproduced and adapted the original works without authorization, and thus infringed the exclusive licensee’s rights.
Comments on findings that copyright was infringed by the AI generated output image
The decision appears to accept that because a generative AI system is trained using a particular work that output that is similar to such a work is infringing. This may be the case, but this assumption may not be true as generative AI systems are capable of producing random images or output that resemble an input work without necessarily having copied it from the input work.
To infringe copyright in Commonwealth countries such as the U.K, Canada, India, and Australia, the plaintiff must establish that the allegedly infringing work was copied from the original work and that there is a reproduction of a substantial part of the original elements in the allegedly infringed work. These concepts have analogues under U.S. law which also requires both probative and illicit copying for there to be an infringement. Substantial similarities can play a role in both prongs of the analysis. Where substantial similarities are present and access has been proved, or if the similarities are striking, courts can presume copying, which can be rebutted by the defendant. But, with generative AI that presumption may be questioned as substantially similar content may be created by or without copying from an input works. The analysis is even more complicated if the allegedly infringed work was itself generated using generative AI tools and such content, if protected at all, would be subject to only thin protection. Thus, the courts would have to assess the implications of how generative AI operates in assessing both probative and illicit copying.
The recent decision from the Chinese court did not appear to wrestle with whether the traditional copyright framework for assessing infringement can still be relied on or needs adapting.[iv]
The protected work and the alleged infringing works in the Chinese case are depicted below.
Further comments on the decisions involving copyright and generative AI
As highlighted above, the two Chinese decisions raise more questions than they answer. The decisions got me thinking about whether the images I created for this blog’s feature image would be protected by copyright. To generate the synthetic images, I used OpenAI’s Dalle-E3. My intent was to generate an image that depicted a girl and a cartoon character that were somewhat inspired by the images in the two cases in a Chinese court setting. I went through many iterations, some of which were close and some not, based on somewhat similar prompts and attempts to modify some of the images. Examples are shown below.
As you may infer somewhat similar prompts generated quite different images suggesting that the lion’s share of the expression, if not all of it, was attributable to Dalle-E3. My prompts could be equated to uncopyrightable ideas transmitted to AI the system. But, even assuming that there some super slim copyright protection, perhaps limited to selection or arrangement of the artifacts in the image (which I had some limited control over), that would perhaps provide some protection only against virtual copying. But, it would also need to be established in an infringement case that even any striking similarities were due to copying and not due to coincidental generation of the substantially similar synthetic content.
[i] Li v. Liu, (2023 Jing 0491 Min Chu No., unofficial translation online.
[ii] Edward Lee, Prompting Progress: Authorship in the Age of AI Florida Law Review, Vol. 76, 2024 Forthcoming; AI Originality Revisited: Can We Prompt Copyright over AI-Generated Pictures? Tianxiang He, GRUR International, ikae024, https://doi.org/10.1093/grurint/ikae024, 07 March 2024; P Bernt Hugenholtz & João Pedro Quintais, “Copyright and Artificial Creation: Does EU Copyright Law Protect AI-Assisted Output?” (2021) 52:9 IIC – Int Rev Intellect Prop Compet Law 1190–1216.
[iii] Guangzhou Internet Court (2024) Yue 0192 Min Chu 113. (2024粤0192初113. This post is based on summaries of the decsions here, See, China – Court Decides Artificial Intelligence Generated Content Infringes Copyright https://www.hg.org/legal-articles/china-court-decides-artificial-intelligence-generated-content-infringes-copyright-66549; Landmark Ruling In China: AI Service Found Guilty Of Copyright Infringement – Latest News In English | Aan World https://aanworld.com/ai-18/; EN Ultraman defeats AI generated copies (hfgip.com) https://www.hfgip.com/news/ultraman-defeats-ai-generated-copies; China’s First Case on AIGC Output Infringement–Ultraman – KWM – Seagull Song https://www.kwm.com/cn/en/insights/latest-thinking/china-s-first-case-on-aigc-output-infringement-ultraman.html.[iv]
[iv] On the issues of how generative AI may affect the potential scope of copyright protection for generative AI content and the substantial similarity tests used in copyright infringement analyses see, Lemley, Mark A., How Generative AI Turns Copyright Upside Down (July 21, 2023). Available at SSRN: https://ssrn.com/abstract=4517702 or http://dx.doi.org/10.2139/ssrn.4517702.