Chinese court deems AI-generated image has copyright – Assessing the possibly over-hasty ‘spring breeze’ case

This Kat has been receiving numerous inquiries from colleagues and friends, particularly those well-versed in copyright law, about an AI-generated image copyright case in China. After reading the original judgment of this high-profile case carefully, this Chinese Kat brings her take on the case.

Background

Case reference: [(2023) Jing 0491 Min Chu No 11279], date: 27 November 2023. The decision, published by Zhichan Ku (知产库) (pdf scanned version, in Chinese only), can be read here,

The plaintiff is Mr. Lee; the defendant is Ms. Liu. This Kat will address the plaintiff as ‘AA’ and the defendant ‘BB’.

AA posted an AI-generated image entitled ‘The Spring Breeze Brings Tenderness’ (the image in question) on Xiaohongshu, a social media platform some call the ‘Chinese Instagram’. In the lower right corner of the image, a string of numbers is marked, which constitute AA’s assigned ID at Xiaohongshu.

Later, AA found an article on another platform posted by BB that contained the image in question without AA’s authorization or AA’s Xiaohongshu ID. AA sued BB at the Beijing Internet Court (the Court) for alleged infringements to both authorship and the right of communication to the public on information networks.

How the image in question was generated? As written in the judgment

The image in question was generated from an open-source software, Stable Diffusion (AA did not use the web version but the .exe version), on 24 February 2023.

Through Stable Diffusion, AA implemented the following key steps (directly translated from the original judgment, tried by this Kat):

1. AA selected and downloaded specific model packages, e.g., ‘AsiaFacemix-pruned-fix.safetensors’;

2. He used the following positive prompts:

(ultra photorealistic: 1.3), extremely high quality highdetail RAW color photo, in locations, japan idol, highly detailed symmetrical attractive face, angular simmetrical face, perfectskin, skin pores, dreamy black eyes, reddish-brown plaits hairs, uniform, long legs, thighhighs, soft focus, (film grain, vivid colors, film emulation, kodak gold portra 100, 35mm, canon50 f1 2), Lens Flare, Golden Hour, HD, Cinematic, Beautiful Dynamic Lighting

3. He used the following negative prompts:

(3d, render, cg, painting, drawing, cartoon, anime, comic: 1.2), bad anatomy, ad hands, text, error, missing fingers, extra digit, fewer di gits, cropped, worst quality, signature, watermark, username, blurry, artist name, (long body) , bad anatomy, liquid body, ma formed, mutated, badproportions, uncoordinated body, unnaturalbody, disfigured, ugly, gross proportions, mutation, disfigured, deformed, (mutation), (child: 1.2) , b&w, fat, extra nipples, minimalistic, nsfw, lowres, badanatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, disfigured, kitsch, ugly, oversaturated, grain, low-res, deformed, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, floating limbs, disconnected limbs, mal formed hands, blur, out of focus, long neck, long body, ugly, disgusting, poorly drawn, childish, mutilated, mangled, old, surreal, text, b&w, monochrome, con joined twins, multiple heads, extra legs, extra arms, meme, elongated, twisted, fingers, strabismus, heterochromia, closed eyes, blurred, watermark, wedding, group, dark skin, dark-skinned female, tattoos, nude, lowres, badanatomy, badhands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry

The negative prompts were all directly copied from the content shared by someone else in a forum. The only part edited by AA himself was ‘3d, render, cg, painting, drawing, cartoon, anime, comic: 1.2’ (highlighted above).

4. AA changed the number of iterations to 33, the height of the image to 768, the prompts guidance coefficient to 9, and the random seed to 2692150200, and got the outcome below:

5. AA changed the weight function of the model ‘lord-hanfugirl-vl-5. Safetensors’ at the ‘Additional-Networks’ to 0.75, and obtained the following image:

6. AA Changed the random seed to 2692150199 and got the image below:

7. With further positive prompts of ‘shy, elegant, cute, lust, cool pose, teen, viewing at camera, masterpiece, best quality’, the image in question was generated:

It shall be noted that the image in question was not completed in one go. Rather, as stated in the judgment (which has also been translated by this Kat):

After obtaining the first image by inputting prompts and setting the relevant parameters, the plaintiff continued to add prompts, modified the parameters, and made constant adjustments and corrections, and finally obtained the image in question.

Does the image constitute a work?

The Court: Yes.

‘Works’ defined by Article 3 of the Copyright Law of China (CRL) refers to ‘intellectual achievements that are original in the fields of literature, art and science, and can be expressed in a certain form’ (emphasis added). The parts of the bold text are the four key components of the works.

On the latter two key components of works, the Court unhesitatingly determined:

In this case, judging from the appearance of the image in question, it is no different from the photographs and paintings that people usually see. Obviously, it (the image in question) belongs to the field of art and is of a certain form of expression.

The first two key components also received the Court’s affirmation. This Kat excerpts some key points relevant from the judgment below.

On the ‘intellectual achievements’ requirement:

A work should reflect the intellectual input of natural persons.

The Stable Diffusion model can be applied to a large number of images and their corresponding text descriptions on the Internet. It can generate images that match the text according to the textual instructions by using the correspondence between the semantic information (contained in the text) and the pixels (contained in the images).

The image generated is not a ready-made image called up by a search engine, nor is it a combination of various elements predefined by the software designer. In layman’s terms, the role or function of the Stable Diffusion model is similar to that of human beings who have acquired some abilities and skills through learning and accumulation. It can generate corresponding images based on text description input by humans, replacing humans in drawing lines and colours, so as to present human beings’ creativity and ideas in a tangible way.

In the entire process, from the plaintiff’s conception to the final selection of the image in question, the plaintiff has made a certain intellectual input, such as designing the presentation of the characters, choosing the prompts, arranging the order of the prompts, setting the relevant parameters, and selecting the image that meets his expectations. The image in question reflected the plaintiff’s intellectual input, so the requirement of “intellectual achievement” has been satisfied.

On the originality:

Of course, not all intellectual achievements are works, but only those that are ‘original’. Generally speaking, ‘originality’ requires that a work is created independently by the author and reflects the author’s individual expression. ‘Mechanical intellectual achievements’ should be excluded. For example, works completed following a specific sequence, formula, or structure will lead to the same results for different people. Such works, due to their uniqueness in expression, lack originality.

As for the use of AI to generate images, the matter of whether or not it reflects the personalised expression of the author should not be generalized, but needs to be judged on a case-by-case basis. Generally speaking, when people use Stable Diffusion to generate images, the more different their demands and the clearer and more specific their descriptions of the elements in the image and layout composition are, the more personalised the expression of the person.

In this case, from the point of view of the image in question, it embodies a recognisable difference from the prior work. From the point of view of the generation process of the image in question, on the one hand, the plaintiff did not draw specific lines and did not even instruct the Stable Diffusion model with 100% clarity on how to draw specific lines and colours. It can be said that the lines and colours that constitute the image in question were basically ‘drawn’ by the Stable Diffusion model, which is different from what people have used before.

However, the plaintiff designed the figure and its presentation, as well as other elements of the image, through the prompts. He also set the parameters for the layout and composition of the image, reflecting the plaintiff’s choice and arrangement. On the other hand, the plaintiff, through the input prompts, set the relevant parameters, obtained the first image, continued to add prompts, modified the parameters, continued to adjust and amend, and finally obtained the image. This adjustment and amendment process also reflects the plaintiff’s aesthetic choice and personalized judgment.

In the trial, the plaintiff generated a different image through the change of individual prompts or parameters. It became clear that, when using the model for creation, different people can input new words, set new parameters, and generate different content. Therefore, the image in question did not belong to ‘mechanical intellectual productions’.

What kind of work is the image in question?

To answer this question, the Court gave, once again, a direct rather than profound answer:

Article 4 of the Regulations for the Implementation of the Copyright Law of the People’s Republic of China stipulates that, a work of fine art refers to a work of plastic art, flat or three-dimensional, which is composed of lines, colours or other means and has aesthetic significance, such as a painting, calligraphy, sculpture, etc. In this case, the image in question is an aesthetic work of art composed of lines and colours, which belongs to the category of works of fine art.

Is the plaintiff entitled to the copyright of the work?

The Court: Yes.

Per Article 11 of the CRL, which stipulates three types of authors (who own the copyright), i.e., natural persons, legal persons or unincorporated organisations, the Court opined that:

At this stage, generative AI models do not have free will and are not subjects in law.

In essence, people still use tools to create, that is, humans make intellectual input in the entire creation process, not AI models.

Notably, the Court added that, per the principle of good faith and the need to protect the public’s right to know, the plaintiff should have prominently marked the fact that he had used AI technology or a model to generate the image. In that regard, since the plaintiff had tagged ‘AI illustration’ on his posting of the image in question at Xiaohongshu, the Court considered this to be adequate. That requirement for information disclosure aligns with Article 17 of a departmental regulation that came into effect this January, titled Provisions on the Administration of Deep Synthesis of Internet-based Information Services (互联网信息服务深度合成管理规定) (accessible here, only in Chinese):

Providers of deep synthesis services should, if they offer services that could potentially cause public confusion or misidentification, prominently mark the generated or edited content in a reasonable position or area to inform the public about the deep synthesis situation. (Further enumerations omitted by this Kat)

Did BB’s accused conduct constitute copyright infringement?

The Court: Yes.

In light of the evidence admitted, the Court ruled that BB’s unauthorized use of AA’s image in question and removing AA’s ID had infringed on AA’s right of authorship and the right of communication to the public on information networks. Therefore, the defendant had to apologise to the plaintiff and pay him a compensation of 500 RMB (approx. 65 EUR).

Comments

The excitement at starting reading the ruling was soon replaced by a sense of ‘This is not quite right’. Particularly when reading the following:

… The plaintiff submitted a video to the Court demonstrating the (re-)creation process of the image in question.

In this Kat’s interpretation, precisely the same image will be generated per the same inputs (e.g., the same prompts, parameters, and random seeds to the same value). A quick search on how Stable Diffusion works (e.g., here and here) supports such observation. Given this, isn’t the image in question exactly a ‘mechanical intellectual achievement’, which the Court had, in the earlier paragraphs of the ruling, excluded from copyright protection? There appears to be a clear contradiction, one that shakes the foundation of the ruling yet was not elaborated on by the Court.

The Court used a common analogy in explaining the lack of human input in the creation process: cameras:

Nowadays, the camera functions of smartphones are becoming increasingly powerful and easier to use. However, as long as smartphone photos reflect the photographer’s creative skills, they still constitute a photographic work and are protected by copyright law. As technology develops and the relevant tools become smarter, the human input decreases. This, nevertheless, does not affect our continued use of the copyright system to encourage the creation of works.

However, Stable Diffusion operates through a mechanism vastly different from cameras. Imagine a large black box filled with numerous images, each coded with a specific set of information, including values and random seeds. In an abstract sense, AA’s process of ‘creation’ was akin to reaching into a general area of this black box and getting a pre-prepared image located in that area. Regardless of what exactly the hand gesture is, it does not 'hold a pen' and thus does not ‘draw’ an image in the black box.

Anyone, not exclusively AA, could have accessed the same image. AA was fortunate to first reach that particular image (though the assumption of AA’s ‘first-reach’ was not evident-proved but an assumption made ‘in the absence of contrary evidence’). AA obtained that final image because there was only one corresponding outcome to a specific set of inputs following the predetermined path formed by the Stable Diffusion algorithm.

Surly, serendipity and strokes of genius in human creation are beautiful and not rare ways to create works. Yet, in this Kat’s view, when it comes to Stable Diffusion, an individual’s ‘creativity’ in generating images is extremely insignificant, and AA should not be deemed the author.

To what extent, or which parts of the image-generation-procedure, did the Court deem, or, should the Court deem, as ‘protectable creation’? A reasonable answer could hardly be AA’s inputs per se, nor the final product. Since the ruling did not provide an in-depth analysis addressing those questions, implied information between the lines is the only trace we could weigh:

Currently, a new generation of generative AI technology is increasingly being used. The Stable Diffusion model and models with similar functions can claim beautiful images based on text descriptions. Many people, including those without drawing skills, are trying to use these new models to generate content and present their ideas and designs in a tangible way, greatly improving the efficiency of creating images. It should be said that generative AI technology has changed the way people create; that is the same impact that many technological advances have had throughout history. The development process of technology is the process of gradually outsourcing human work to machines.

Encouraging creativity is widely-recognized as the core purpose of the copyright system. Only by correctly applying the copyright system, using appropriate legal means, and encouraging more people to use the latest tools to create, can it be more conducive to the creation of works and the development of AI technology. Under this background and technological reality, the images generated by AI, as long as they can reflect human’s original intellectual input, should be recognized as works and enjoy protection under copyright law.

Clearly, the Court valued and encouraged human participation in the human-AI-creation scenarios. It was seemingly the case to the Court that the image in question is deemed as not completely detached from human intellectual activities, but a human’s intellectual activity assisted by an AI; Accordingly, the line between an idea and its expression is less distinct once the human’s inputs/participations/interventions go beyond certain thresholds, reaching to another level where human’s inputs and AI’s processing together, via interactions, merged as the expression that is protectable under the copyright law.

The Court points on the threshold of originality here, while engaging, may not be relevant to the Stable Diffusion-generated images that are ‘mechanical’. An immediate reductio ad absurdum scenario could be, that the sweepstakes nature of obtaining copyrights for works like ‘The Spring Breeze Brings Tenderness’ may easily lead to a competition of speed in generating images and author-claiming in a batch, which raises concerns about the impact on parts of the human creativity that are worthy of protection.

Considering the significantly higher efficiency of generation and lower costs of ‘human-AI works’ compared to those traditional works from human beings, does setting (or lowering) the threshold of copyright protection for the former category of work truly foster human creativity? Or does it risk stifling innovation and potentially overloading courts with legal disputes? As ‘human-AI works’ become more prevalent and sophisticated, do the traditional concepts and logic of copyright law need to be re-evaluated and adapted? And how should generated benefits be fairly distributed to realize interest-balancing and sustainable developments? Is copyright necessarily the only appropriate tool to reflect and regulate the considerable nuance featured in the different scenarios of the AI-involved human creation process?

This case is very much more of a start rather than an end, regarding which, the discussions invoked will keep growing. On a side note, this Kat wishes to answer a message from an Italian friend of hers who asked: ‘Do you think that the ruling reflects a general approach of the Chinese court or that it depends on the peculiarities of the specific case?’ This Kat would say the answer is the latter suggestion. This case does open up more questions than answers and raises substantial issues. 

Interestingly, when it comes to the ‘spring breeze’, there is an exceptionally well-known phrase in China: ‘the spring breeze of reform (改革春风)’, which has often referred to the economic reform and opening-up that began in the late 1970s. Those reforms led to significant changes in China’s economy and society, like a revitalizing spring breeze. The similar expressions of the ‘spring breeze’ here, i.e., the image in question and the proverb, somehow coincidently, seem to resonate with the Court’s recognition and welcome of the emerging technologies. However, in this specific case, it is fair to say that the ‘spring breeze’-intended ruling is slightly over-hasty.

Photos credits: Zhichan Ku