In November 2023, Beijing Internet Court rendered a verdict in China’s first case concerning the copyrightability of AI-generated pictures LI v. LIU, which ruling triggered mixed reactions among the AI industry and the public. The plaintiff, Mr. LI (the “Plaintiff”), filed a copyright lawsuit alleging that the defendant, Ms. LIU, a blogger on Baijiahao platform (the “Defendant”), violated his copyrights in the AI-generated picture, including his right of authorship and the right of dissemination via information networks. In its decision, Beijing Internet Court found that the AI-generated picture was a copyrightable work involving human authorship and that Defendant was liable for copyright infringement.
On February 24, 2023, Plaintiff generated a few pictures using Stable Diffusion, a U.S.-based text-to-picture artificial intelligence (“AI”) service. He labelled one of such pictures as “Spring Breeze Brings Tenderness—AI generated picture” (春风送来了温柔) and posted it on a popular Chinese lifestyle social media platform “Little Red Book” (Xiaohongshu). Defendant, a Chinese blogger, then published an article titled “Love in March, Among Peach Blossoms” (三月的爱情,在桃花里), using Plaintiff’s AI-generated picture “Spring Breeze Brings Tenderness” as an illustration in her article, but removing Plaintiff’s user ID and the watermark of Little Red Book from the picture. Plaintiff soon sued Defendant for copyright infringement, including violating his right of authorship and the right of dissemination via the internet.
Plaintiff’s AI-generated picture “Spring Breeze Brings Tenderness”
In this case, the court focused on the following issues: (1) whether the subject AI-generated picture constitutes a copyrightable work and therefore subject to Chinese copyright protection; (2) if yes, whether Plaintiff is the copyright owner of the subject AI-generated picture; and finally (3) whether Defendant should be held liable for copyright infringement.
1. Whether the Subject AI-Generated Picture Constitutes Copyrightable Work.
On the first issue, Beijing Internet Court ruled that the subject AI-generated picture “Spring Breeze Brings Tenderness” constitutes a copyrightable work—a work of fine art, and therefore, subject to Chinese copyright protection.
The Beijing court started its analysis by listing out the criteria for a work to be protected under Chinese copyright law, including: (1) whether the work falls within the fields of literature, art and science; (2) whether it possesses originality; (3) whether it has a specific form of expression; and finally (4) whether it is intellectual creations (by humans).
With regard to the first and third criteria, the Beijing court held that because the subject picture is akin to commonly seen photographs and paintings, it has satisfied these two criteria.
With regard to the criteria of “intellectual creations”, the court held that a (copyrightable) work should reflect the intellectual input/contributions of human beings. In this case, the court found that, Plaintiff has provided intellectual inputs throughout the subject picture-generation process, including: (1) choosing the preferred AI service provider (i.e., Stable Diffusion) among many other alternative AI picture service providers to render the picture style that Plaintiff prefers; (2) inputting around 30 “Prompts” and over 120 “negative Prompts” to determine the output of the AI-generated picture; and (3) setting and re-setting various technical parameters to produce, choose and re-arrange the pictures that Plaintiff favors. As such, the court determined that the subject picture has reflected Plaintiff’s intellectual input, thus meeting the criteria of “intellectual creations”.
With regard to the criteria of “originality”, the Beijing court held that a copyrightable work should be independently created by its author and reflect the author’s personalized expressions. The court went further, stating that “as to whether the use of AI to generate pictures could reflect the author’s personalized expressions, it needs to be decided on a case-by-case basis”. The court held that, in this case, although Plaintiff did not physically draw the specific lines (using his own hands), Plaintiff designed the character styles and arranged the final layout and composition of the picture, by trying different prompt words, negative prompt words and various tech parameters. The judges seemed to be especially impressed with the fact that after Plaintiff obtained the first picture by entering around 150 prompts, negative prompt words and relevant parameters, Plaintiff continued to add more prompt words and keep changing the tech parameters until he received the final subject picture that he was happy with. The court thus concluded that, this entire process of adjustment and rearrangement reflected Plaintiff’s aesthetic choices and personal judgment. Therefore, the court found that the subject picture is not merely a “mechanical intellectual creation” but possessing originality of the author.
The creation process of Plaintiff’s subject picture
2. Whether Plaintiff is the Copyright Owner of the Subject Picture.
The court first ruled out the possibility that an AI service itself could be considered as an author of a copyrightable work because an AI is not a human being.
Then, the court held that neither the developers/providers of AI services could be considered as the author in this case because such AI providers/developers neither had the intent to create the subject picture nor did they actually participate in the subject picture creation process. Further, based on the “CreativeML Open RATL++-M License” of Stable Diffusion posted on GitHub.com, the AI developers already waived their rights, if any, in the AIGC output—that they “do not claim rights to the output content”. As such, the court held that, because the subject picture was generated as a result of Plaintiff’s intellectual input and reflected Plaintiff’s personalized expressions, Plaintiff is the author of the subject picture.
3. Whether Defendant Should Be Held Liable.
The court finally found Defendant liable for infringing Plaintiff’s copyright in the subject AI-generated picture, on the grounds that Defendant removed Plaintiff's user ID and the watermark of Little Red Book from Plaintiff’s AI-generated picture and reposted it on social media platform without authorization.
Beijing Internet Court’s decision in LI v. LIU seems to contradict with the recent U.S. decisions on the copyrightability of AIGC output (e.g., “Zarya of the Dawn”, “A Recent Entrance to Paradise” and “Theatre D’opera Spatial”), in which both the U.S. Copyright Office and the U.S. courts have denied copyright protection to AIGC outputs that lack human authorship. Yet the difference between the Chinese case LI v. LIU and the U.S. cases is not that Chinese courts believe that non-humans could be “authors” or that Chinese copyright law does not require “human authorship” in copyrightable works. Rather, in LI v. LIU, Beijing Internet Court seems to make a distinction between a straightforward AIGC output, where the human author simply takes and uses the AIGC output “as is” without any creative involvement and an AIGC output, where the human author keeps experimenting and adding various prompts, negative prompts and tech parameters until he receives the final satisfactory piece. In the later scenario, the Beijing court was actually deeming the subject work as “AI-assisted work”, where Plaintiff has exercised aesthetic choices and personal judgment in the final representation of the work.
It should also be noted that in the court’s hearing, Plaintiff has demonstrated that he could receive the exact same AI picture that he received earlier and claimed copyright for by inputting the exact same prompt instructions he selected (i.e., over 150 prompts, negative prompts and tech parameters). It would be interesting to speculate whether the Chinese judges would come to the same conclusion—recognizing the copyrightability of the subject AI picture if the AIGC output turns out to be unpredictable—producing various AI pictures each time. Would the Chinese judges change their rationale because the human authors do not have “control” in the AIGC output? Again, where to draw the line between an AIGC output that is not qualified for copyright protection and an AIGC output that is subject to copyright protection remains a challenge for the AI industry to watch in the years to come.
Thanks to Wang Mo and Zhao Yibing for their contributions to this article.