In December 2023, the Beijing Internet Court held a public hearing to hear the first known case of AI painting in the country in the era of large models. The case goes like this: Someone saw an AI-generated picture of a goddess on the Internet. It was beautiful and touching. After all, it was generated by AI and would not come to ask for copyright, so he forwarded it directly to his own self-media platform.
There is indeed no way for AI to claim copyright, but the person who controlled the AI to generate this picture of the goddess is here, and the court's verdict is also here - the Beijing Internet Court made a judicial determination on "whether the AI painting constitutes a work within the meaning of copyright law."It was determined that the plaintiff owned the copyright of the AI painting and the defendant was sentenced to pay RMB 500 in compensation.
The pictures cited in the case may have copyright risks, so we used the midjourney tool to reproduce them with AI.
Although the amount of the verdict is small, the special feature of this case is that this drawing was generated on a computer through a "stablediffusion model", which means that it is not the plaintiff's painting, but a work automatically generated by an AI program based on "prompt words" given by humans.
As soon as this judgment came out, it immediately triggered discussions in the legal circles and the artificial intelligence industry circle. Many relevant people and scholars put forward their own different opinions on the judgment in different forms.
What is the "stable diffusion model"?
The "Stable Diffusion Model" was released in 2022 and is a deep learning artificial intelligence model. After training, it can gradually perform denoising operations on random Gaussian noise to obtain image samples.
The model is mainly used to generate images from text, although it can also be applied to other tasks such as image inpainting, image expansion, and image-to-image or text-to-image generation under the guidance of prompt words.
As far as the technical route is concerned, stable diffusion is a variant of the diffusion model and a further development of the "latent diffusion model".
The picture is generated by AI and comes from the platform midjouney
The stable diffusion model is trained from a large number of pictures and their corresponding text descriptions on the Internet. This model can use the correspondence between the semantic information contained in the text and the pixels contained in the picture according to the text instructions to generate pictures that match the text information. The new pictures generated are not calling existing ready-made pictures through search engines, nor are they arranging and combining various elements preset by the software designers.
Generally speaking, the role or function of this model is similar to that of human beings who have acquired some abilities and skills through learning and accumulation. It can generate corresponding pictures based on the text descriptions input by humans, draw lines and colors on behalf of humans, and present human text creativity and ideas in a tangible way.
Unlike other cloud-based AI services, the pre-trained stable diffusion model is open source, and the source code can be publicly downloaded and installed and run on your local computer. Therefore this tool has been widely used in the production of commercial images.
Why does the copyright of images created by AI belong to the owner?
In December 2023, the Beijing Internet Court issued a judgment (2023) Beijing 0491 Minchu No. 11279, confirming that the copyright of the AI drawing work belonged to the plaintiff.
The reasons for the above judgment are:
When the plaintiff released the picture involved in the case, it had been marked as "AI illustration", and the plaintiff could use the stable diffusion model to restore the generation process of the picture based on the prompt words and parameters set by itself.
Of course, the process of producing the specific disputed picture cannot be restored, because the generation of the picture itself is random. In the absence of contrary evidence, it can be concluded that the "Spring Breeze Brings Tenderness" image involved in the case was generated by the plaintiff using generative artificial intelligence technology.
The court held that from the conception of the picture involved in the case to the final selection of the picture involved, the plaintiff made a certain amount of intellectual investment in the entire process, such as designing the presentation of the characters, selecting prompt words, arranging the order of the prompt words, setting relevant parameters, selecting which picture meets expectations, etc. The pictures involved in the case reflected the plaintiff’s intellectual investment, so the pictures involved in the case met the requirements of “intellectual achievements”.
According to the court’s comments, the plaintiff’s instructions to the stable diffusion model in this case were similar to “commissioning others to create” according to the prompt words it input. If the stable diffusion model is a specific human being, then the plaintiff in this case is clearly not the author of the drawing.
However, the stable diffusion model is not a human being, nor a legal person, nor a subject under copyright law. Therefore, it cannot enjoy copyright. The copyright belongs to the plaintiff Li who "told the AI how to draw".
Similar to this case, Shenzhen Nanshan Court stated in a judgment in October 2018: "The article involved was generated by the plaintiff's creative team using Tencent's writing robot (Dreamwriter). Its external performance complies with the formal requirements of a written work, and its content reflects the selection, analysis, and judgment of relevant stock market information and data that morning. The article has a reasonable structure, clear expression logic, and a certain degree of originality."
The court determined that the creative process technically "generated" by Tencent's writing robot met the protection conditions for literary works under the Copyright Law and was a literary work protected by my country's Copyright Law. Therefore, the defendant was ordered to compensate the plaintiff for economic losses and reasonable rights protection costs of RMB 1,500.
What exactly is copyright?
The legal copyright of works includes personal rights such as publication, signature, modification, etc., as well as property rights such as reproduction, rental, performance, and adaptation. Property rights can be licensed to others to exercise, or can be transferred to others in whole or in part, but personal rights belong to the author and cannot be transferred or authorized.
That is to say, if the copyright of a work is determined to belong to a certain rights holder, he has the right to prohibit others from copying the work, and also has the right to prohibit the use of the work for training other AI models, unless there is a contrary agreement.
The "Copyright Law of the People's Republic of China" defines "works" as - refers to intellectual achievements that are original and can be expressed in a certain form in the fields of literature, art and science. Works by citizens, legal persons or unincorporated organizations, whether published or not, shall enjoy copyright in accordance with this Law.
The protected works in the U.S. Copyright Act of 1976 (Copyright Act of 1976) must be "fixed in any material form now or created in the future. The original work can be perceived, copied or disseminated in other ways, directly or with the help of mechanical devices." They mainly include written works, musical works, dramatic works, pantomime works and dance works, pictures, paintings and sculptures, film works and other audio-visual works, sound recordings, and architectural works. Oral works are not fixed and are not protected. The scope of copyright protection is limited to the form of expression of the work and does not extend to its ideas.
In essence, copyright protects human expression, that is, both humans and human organizations (legal persons) can become the "authors" of a work. Animals, plants or other non-living entities other than these two cannot become "authors" under copyright law.
A painter who randomly splashes paint on the canvas can own the copyright of the post-modern art style work, because although the entire creative process is random and unrepeatable, the expression process of "painting" is completed by "human beings". However, a photographer who deliberately throws a camera with adjusted shutter and aperture parameters to a monkey cannot claim to own the copyright of the monkey’s photos, because he only provided the scope and conditions for creation, but did not complete the photographic creation itself.
In the same thought experiment, a trainer who hands brushes and paints to an elephant cannot claim copyright in the "work" painted by the elephant. Although he does "train" the elephant to paint, he cannot control the final process and expression of the elephant's painting.
Don't think this is an exaggeration;There are real cases of "monkeys taking photos".
In 2001, British outdoor photographer David J. Slater accidentally got a "selfie" of a crested black macaque while visiting North Sulawesi National Park in Indonesia.
The photo was immediately shared by many media outlets around the world, setting off a copyright war between Wikipedia and Slater.
Slater claimed that he owned the copyright to the photo, and Wikipedia made the image available for public download, causing it to lose huge royalties. However, Wikipedia said that no one owned the copyright to the photo because the photo was taken by a monkey.
In December 2014, the United States Copyright Office (USCO) stated that works created by non-human beings are not protected by U.S. copyright law.
Selfie of a macaque, source: MediaWiki
Regarding the copyright issue of AI-generated images discussed in this article, U.S. Federal District Judge Beryl A. Howell dismissed AI entrepreneur Stephen Thaler's lawsuit against the U.S. Copyright Office on August 18, 2023. She ruled that artistic works generated by AI are not protected by copyright, and emphasized that human creation is "an important component of a valid copyright claim."
Contrary to the Beijing Internet Court, in this case, although the plaintiff also "provided instructions and directed his artificial intelligence to create works" and "the artificial intelligence is completely controlled by (him) and the artificial intelligence only operates under (his) instructions," the U.S. court still determined that "the author's human nature is a basic requirement of copyright."
Photo of US court verdict, source: MediaWiki
The author believes that the reason for the different judgments of different courts lies in the different value judgments on whether the subject who completes the copyright creation must be a "person", as well as the different understanding of the true inner principles of AI tools such as the stable diffusion model.
In the logic of some cases, the Diffusion model is equivalent to a hammer or a paintbrush in a human hand. The person behind the control of the tool is still the specific human being who operates the tool.
But in the logic of other cases, humans using AI tools are like throwing a paintbrush to an elephant or a monkey with a camera. Although humans issue various "instructions," the process of "creation" is not completed by humans.
Potential work copyright issues with artificial intelligence learning models
Artificial intelligence models such as the stable diffusion model require a large amount of original data or pictures for training. The pictures output by the plaintiffs in these cases are the results of a huge number of original pictures for training.
On April 11, 2023, the Cyberspace Administration of China released the "Measures for the Management of Generative Artificial Intelligence Services (Draft for Comments)", proposing that in-depth artificial intelligence products or services should respect intellectual property business ethics, and set up "barriers" on privacy, intellectual property, training data, unfair competition, etc., and specifically clarified that pre-training and optimized training data used for generative artificial intelligence products must not infringe.
Just a few months ago, in January 2023, GettyImage, one of the world's largest image distributors, sued the creator team of the Stable Diffusion Model, accusing the latter of misusing millions of its copyrighted images as training data without authorization - because the model generated images with distorted but still clearly legible Getty image watermarks (this watermark would appear on the preview of unauthorized images).
However, the machine learning model is based on Bayesian operations, and its own mathematical principles determine the black box and uninterpretability of the process. Humans cannot observe the specific steps of final image generation. Unless the above-mentioned "accident" occurs, we cannot judge which specific data the generated image is based on, let alone whether the data is legally authorized.
Just like I can't describe the specific reading in my life that the language fragments and knowledge used to write this article come from - then, the authors of the data set I used to write this article - the authors of primary school textbooks, can they claim the right to adapt this article from me? And let's go one step further. If I read a pirated book intentionally or unintentionally, how can the original author of the book claim the rights he deserves from me?
A deeper future
When new productivity and production tools appear, it is often people's perceptions and rules that need to be changed.
Current deep learning model algorithms are in a closed environment. The data set provided by the training programmer and the actual data encountered, the model distribution in the closed environment remains unchanged. In actual human life, the environment we live in is open and random, and it is impossible to exhaust all data and possibilities. This is why the expression of computers and programs may be essentially different from human expression.
The law itself is just an abstract human consensus. We have seen different rulings on whether AI works can obtain copyright. Some judges are obviously more conservative on the issue of whether to recognize the copyright of AI works.
As to whether works created by human beings using "parameters" and "conditional restrictions" should be given copyright, and whether the author of these pictures is the program itself, or the human who just entered a few prompt words, it is probably not an easy question to answer and has a fixed answer.
After all, code is iterating rapidly in ways that humans cannot see, computing power is improving at a rapid pace, and the standards of authors and works are still highly controversial in the academic world.
We believe that in the foreseeable future, these disputes themselves may also become a footnote and motivation for human technology to continue to advance.
Planning and production
Author丨Jiang Yifan, lawyer at Beijing Kangda (Shenzhen) Law Firm
Review｜Zhao Hu, lawyer at Beijing Zhongwen Law Firm
Planning丨Xu Lai
Editor丨Lin Lin