Legal Boundaries of AI-Generated Content and Artistic Creation

More than 20 years ago, the internet challenged the intellectual property legal system, and the Digital Millennium Copyright Act passed in the US in 1998 provided an early legal response to the development of internet technology. Today, generative artificial intelligence (hereinafter referred to as “Generative AI”) again impacts the field of intellectual property. In May 2023, China passed the “Interim Measures for the Management of Generative AI Services” as a legal response to this issue. Around the intellectual property issues related to Generative AI, the Hong Kong University’s Huang Qianheng China Law Research Center and the Policy and Regulation Working Group of the China Artificial Intelligence Industry Development Alliance held a summit dialogue on January 26. During the dialogue, Zhu Ge, Deputy Chief Judge of the Comprehensive Trial Division of the Beijing Internet Court, Professor Cui Guobin from Tsinghua University Law School, Professor Wang Qian from East China University of Political Science and Law, Professor Jason M. Schultz from New York University School of Law, and Professor James Grimmelmann from Cornell University Law School engaged in a heated discussion on copyright issues concerning AI-generated content.

The first case of “AI-generated images” in China has drawn international attention.

According to Zhu Ge, the presiding judge of the first instance of the “AI-generated image” copyright case, the plaintiff generated an image using a large AI model, titled “The Gentle Spring Breeze Brings Warmth.” The defendant, an original poetry author, used this image as an illustration when publishing original poetry. The plaintiff claimed that the defendant removed the signature watermark from the image and uploaded it to social media platforms, infringing on the plaintiff’s right to attribution and the right of communication through information networks. If the plaintiff had not generated the image using AI, the case would appear quite common and could be handled very simply. In determining the work, the court found that the image in question was a flat artistic work composed of lines and colors, possessing identifiable differences, and the generation process reflected the plaintiff’s original intellectual contribution, meeting the definition of a work under Chinese copyright law. Regarding the attribution of rights, according to Chinese copyright law, copyright generally belongs to the author, but the author is limited to natural persons, legal persons, and unincorporated organizations; the AI model itself cannot be the author. The plaintiff configured the AI model based on their needs and selected the image, which was directly generated through the plaintiff’s intellectual input, reflecting their personalized expression. Thus, the court concluded that the plaintiff was the author of the image in question and enjoyed the copyright of that image. After establishing the composition of the work and the attribution of rights, similar to other cases, the court determined that the defendant infringed on the plaintiff’s right to attribution and the right of communication through information networks.

Discussing the trial’s approach, Zhu Ge stated that four elements can be inferred from the definition of works in Chinese copyright law: limited domain, form of expression, intellectual achievement, and originality. The image in question evidently meets the first two criteria; the discussion centers around the intellectual achievement and originality criteria. Intellectual achievement requires the reflection of human intellectual input, and AI models are not civil subjects. In what aspects might intellectual input be reflected? Because it is a new type of creative tool, based on the generation process of the image in question, the court believed that from the conception of the image to the final selection, the plaintiff made certain intellectual contributions, including designing the presentation of characters, selecting prompt words, arranging the order of prompt words, and selecting images that met expectations, fulfilling the intellectual achievement requirement. Originality is the core criterion for defining a work, requiring that the work be completed independently by the author, reflecting the author’s original expression. Whether the use of Generative AI can reflect the author’s personalized expression requires case-by-case judgment. Generally speaking, when people use AI models to generate images, the demands posed are diverse; the more specific the description of the layout, the more it can reflect human personalized expression. The plaintiff made arrangements and selections regarding the visual elements, layout, and other expressive details of the image based on their aesthetic choices and personal judgments, and their intent was manifested through the image in question. At this point, AI serves as the author’s brush or camera, a tool for creation.

Zhu Ge believes that the application of law involves scientifically interpreting the law, sometimes requiring certain rules and methods to fill legal gaps. In the legal application of new and difficult cases, interest balancing is generally incorporated, prioritizing the interests of both parties and their respective groups, as well as considering the legislator’s value choices and social public interests. The ruling in this case aligns with the legislator’s value choices. The purpose of copyright law is to encourage creation; under certain conditions, works generated based on AI can motivate people to create using new tools. Additionally, encouraging users to form a positive cycle of use, profit, investment, and reuse can have a positive impact on emerging industries. The court also considered the interests of the public. Under existing technological conditions, it is challenging to detect and distinguish whether a work is AI-generated; if treated differently, failing to protect AI-generated works could send a negative incentive to society, leading people to refuse to use new tools or conceal their usage, infringing upon the public’s right to know.

It is challenging to distinctly separate rights between manually created and AI-generated works.

Regarding the motivations and incentives for AI creation mentioned by Zhu Ge, Wang Qian noted that in the US, despite many AI-generated works not being protected by copyright law, the number of users employing AI has not decreased. This indicates that whether copyright protection is granted to AI-generated works does not directly impact users’ willingness. Grimmelmann also responded to this topic, stating that there is very little objective distinction between AI-generated works and human-created works. If copyright law stipulates that all AI-generated works are not protected, it may entice many individuals to deny their use of AI. He is uncertain whether this issue can be resolved, but he believes that if a legal system sets a stark distinction between the rights of manually created and AI-generated works, it could indeed produce negative incentives.

In response, Zhu Ge stated that first, AI allows individuals without artistic skills to enter the market for fine art and showcase their creativity. Secondly, many artists have incorporated AI models into their toolbox, which may replace repetitive labor for artists. Furthermore, regarding value, many people may prefer handmade items, even if they are more expensive, but there remains a demand for them. In the future, if the market is flooded with AI-generated works, handmade paintings will become scarcer and more valuable. Additionally, currently, the use of AI software has certain thresholds, and users should be encouraged to invest more effort in learning. Finally, according to mainstream views in China, originality is a binary issue. Based on existing standards for works, a considerable portion of AI-generated works can meet the originality requirements because China focuses on human input.

Regarding whether the US Copyright Office might grant copyright to the works in this case, Grimmelmann pointed out that there is not much difference in how the two countries’ laws handle this issue. Currently, related cases in the US often encounter problems during the copyright registration phase. These cases can be seen as a test, establishing broader copyright law precedents. The creators involved either claim to register the AI itself as the author or have very incomplete disclosures regarding the input instructions and AI participation, failing to adequately reflect the importance of human involvement in the creative process. The nature of the “AI-generated images” case is entirely different; it is a specific infringement lawsuit with very detailed disclosures regarding the generation process and input instructions. Therefore, there may not be any essential differences between the two jurisdictions. If this case occurred in the US, it might yield similar judicial outcomes. Although in the “Space Opera” case, an AI-generated work containing over 600 instructions was rejected by the US Copyright Office, the creator did not disclose the specific content of these instructions, nor did they disclose the original sketches generated by the AI. The US Copyright Office and courts encourage creators to actively disclose their level of involvement; otherwise, insufficient evidence may prevent the work from being protected by copyright.

Cui Guobin mentioned that the US Copyright Office has rejected copyright protection for many AI-generated works on the grounds of lacking originality, while the modification details of prompt words in the Chinese case may be less than those in some American works. For instance, in the “Space Opera” case, after the creator selected the work, they first determined the broad framework and then modified the details, repeatedly using traditional painting tools like Photoshop to refine the content, a process that took a considerable amount of time; yet the US Copyright Office still deemed it lacked originality. If the aforementioned case occurred in the US, under stringent standards, the US Copyright Office would likely conclude that the image in question lacked originality.

Wang Qian also believes that based on the rulings and guidelines of the US Copyright Office, it would not be possible to register the images in the “AI-generated images” case. The US Copyright Office does not question the applicant’s integrity; rather, it seeks to clarify the recognition of works generated by AI. In the “Space Opera” case, the US Copyright Office did not refute the applicant’s claim of using over 600 prompt words, but still believed that this “is not a human-created work” because the content is generated autonomously by AI. The US Copyright Office does not consider that content formed after processing AI-generated content through Photoshop cannot be registered; they require the applicant to declare the waiver of rights to the AI-generated content, but the applicant is unwilling to do so, which ultimately leads to a failure to register. Currently, AI cannot understand user instructions as humans do; it can only generate new images based on its training and algorithm rules. Therefore, no matter how many rounds the work has been modified, each round remains a black box; the user always lacks control over the generated content and cannot predict the final outcome. When a person describes this work using natural language, no matter how detailed, it will never fully align with the image generated by AI, unless AI develops to a level akin to the movie “Inception,” accurately “copying” the artwork conceived in the human brain. But by then, “Generative AI” should be renamed “Copying AI.”

Cui Guobin stated that the Stable Diffusion software used by the plaintiff in the “AI-generated images” copyright case is open-source, and many individuals are developing various plugins that enable users to modify specific content in the images through text, graphic instructions, or keyboard operations, providing a vast operational space. As AI technology continues to evolve, continuing to emphasize the essential differences between modification methods based on text instructions and button-based modifications in Photoshop may be meaningless. Achieving seamless integration between the two is an inevitable trend for the future. The examples of AI failing to modify the specific features of selected images do not prove that using text prompts cannot specifically or thoroughly modify an AI output image; this is due to the existing AI technology not being fully utilized or the user’s inaccurate understanding of the potential of AI plugin technology.

Revisiting the Dichotomy of Thought and Expression

In copyright law, the principle of the dichotomy between thought and expression divides works into two aspects; copyright law only protects the original expression of ideas and does not protect the ideas themselves. Regarding the challenges that Generative AI poses to the dichotomy theory, Grimmelmann pointed out that currently, creators can output rich content by simply inputting simple instructions. If the traditional dichotomy is upheld, it implies that copyright law only protects the instructions themselves since only the instructions represent the creator’s expression, and overly brief instructions may not receive copyright protection. When the instructions are more creative and constructive, the connection between the instructions and the AI-generated work is often tighter. If only a small number of prompt words are provided, allowing AI to complete the remaining work, it becomes difficult to achieve the effect of conveying original expression. The invention of the Gutenberg printing press significantly reduced the production costs of literary works, but it took over two centuries for the world’s first copyright law to be enacted. Now, humanity has just entered the era of Generative AI, and we know very little about how to utilize AI to promote creativity, how to design incentive mechanisms, and the potential risks of AI-generated content, which makes the current situation full of uncertainty.

According to Schultz’s analysis, copyright law theory posits that its purpose is to encourage people to create. The widespread application of Generative AI has significantly lowered the costs and thresholds of creation. Assuming the core of art lies in “conception,” while “expression” is merely a mechanical execution process, then under the premise that AI can greatly simplify the expression process, it is necessary to revisit the dichotomy of thought and expression. In a Generative AI-dominated environment, do people really need copyright protection as an incentive to create? This remains questionable.

However, Wang Qian believes that most of the so-called problems and challenges posed by AI do not necessarily exist; the key lies in whether using works to train AI constitutes fair use, which does not challenge the dichotomy. In the context of AI, the focus is on whether the prompt words input by users are merely “thoughts” relative to the images generated by AI. If so, they are not protected. The discussion here is not whether the prompt words themselves constitute works or expressions, but whether the images generated by AI based on the prompt words represent the user’s expression under copyright law. Suppose an art college teacher, who is also a poet, writes a poem on the spot and asks all students to draw a picture based on this poem. The poem written by this teacher is undoubtedly a work, but this poem is, in relation to the paintings created by these students, merely a thought. It is clear that this poem cannot determine the composition of each painting; the students can interpret the poem according to their thoughts and generate corresponding specific images using their creativity. No matter how complex and detailed the verbal description of the image, it cannot fully determine its composition. Similarly, even if the prompt words are sufficiently detailed, the generated images cannot constitute the user’s work. Wang Qian pointed out that he once input a poem describing a sunset scene into two AI models, and the generated images were completely different. This poem’s description was already detailed enough; if it were insufficient, one could write a thousand more lines. However, inputting a thousand lines of poetry into two models would still not yield the same result. If selecting and inputting prompt words is deemed a creative act, why can so many different expressions emerge? This is because words, in relation to the images they describe, are merely thoughts, not expressions.

Regarding the dichotomy of thought and expression, Cui Guobin stated that if the creator only inputs prompt words, although the prompt words themselves may constitute a written work, the images generated by AI based on the prompt words usually do not contain the creator’s original expression. Only after the creator selects the image and then repeatedly modifies the expressive details or compositional elements through prompt words or other means, can originality be established in the expressive part of the image.

“Substantial similarity” is an important rule for determining infringement in copyright law, indicating that the defendant’s work and the plaintiff’s work express the same thought to the extent that it excludes the possibility of the defendant’s independent creation. Grimmelmann stated that copyright law judges infringement based on the principle of substantial similarity. If the creator’s expression of thought is embodied in the instructions, then the differences between the instructions should be compared. Although the content generated by the same instruction through AI may differ, two completely different instructions may generate highly similar content.

Wang Qian believes that AI has not challenged the standards for determining substantial similarity. The comparison of substantial similarity adopts an objective standard, focusing solely on the degree of similarity between the plaintiff’s work and the alleged infringing content, regardless of whether the alleged infringing content is generated by AI or created by humans.

Discussing Infringement and Fair Use on a Case-by-Case Basis

Fair use refers to using a work in a certain way according to copyright law without needing the copyright owner’s consent or paying them remuneration. Grimmelmann stated that there are two major trends in fair use in the US. One is transformative use, which creatively adapts others’ works; the other is copying others’ materials for purposes that lack artistic expression, such as research archives and search engines. These systems include many copyrighted works, but the products they reproduce do not compete with the original works. However, Generative AI merges both. It not only widely absorbs copyrighted works for training but also generates derivative works with expressive qualities. Therefore, Generative AI is not entirely applicable to either of the above situations, yet it is related to both.

Schultz stated that when assessing whether fair use is applicable, it is necessary to evaluate the final generated content. The focus of the debate is on the difference between AI companies automatically crawling data for training and obtaining specific authorization from copyright holders. This involves two issues: the first is competition. To promote healthy competition among different AI tools, we cannot allow only the wealthiest companies in the world to access training data. However, the costs of purchasing copyright are extremely high. The second issue is bias. If AI can use vast amounts of data for training, it can better prevent the generation of biased content. Different political factions in the US have varying attitudes towards AI using their data for training; if a licensing system is implemented, will the training data be dominated by the views of a particular faction? Additionally, if AI-generated content indeed constitutes substantial similarity, should the user or the AI service provider be held responsible?

Grimmelmann and Schultz both agree that there is no one-size-fits-all answer to the issue of “fair use.” The purposes and circumstances of users employing AI vary, requiring specific analysis of each case. Copyright law should afford appropriate leniency to private spaces. In relevant cases, when AI-generated content is uploaded to the public domain, especially when the case involves unfair competition factors, the nature of the issue changes. In summary, Generative AI is still in its early stages of development, and many copyright issues lack clear answers, relying on more judicial precedents to enrich people’s knowledge base. A reasonable system should not depend on large-scale suppliers, helping both large companies and startups thrive, ensuring that everyone has equal rights to use AI and encouraging fair competition. If the law only allows large companies to deal with other large companies, it neither resolves substantial issues nor enables creators to receive reasonable compensation. Furthermore, the economic interests of copyright holders may be harmed because more people can compete with them, not just because someone has stolen their work through ChatGPT.

In response to the remarks of the two experts, Cui Guobin commented that both American experts tend to recognize that the training use of data by AI may constitute fair use. Professor Schultz’s viewpoint mainly considers competition and neutrality from two aspects. First, requiring licenses for the AI training process may hinder fair competition; this involves not only competition between companies but also competition between countries. Secondly, if some copyright holders agree to license while others do not, it may lead to bias in the AI-generated content. Additionally, fair use may apply more to non-commercial purposes, while purely commercial purposes require reevaluation. However, in Cui Guobin’s view, purely commercial purposes can also be seen as fair use. If AI suppliers are required to pay high copyright fees for all training data and confirm each person’s contribution, it could lead to market failure and increased social costs, which is unnecessary. Of course, if the AI output content infringes, legal responsibility should be pursued in the output process, which is a different issue from whether the use behavior in the training process constitutes fair use.

Leave a Comment Cancel reply