James Grimmelmann: In the U.S., there are two primary trends regarding “fair use.” The first is transformative use, which involves creatively adapting the work of others. The second is copying materials without artistic expression, such as research archives and search engines; while these systems feed on copyrighted data, their output does not compete with the original works. However, generative AI straddles both trends. It not only feeds on copyrighted works but also produces expressive derivative content. As a result, generative AI does not fit neatly into either of the existing fair use scenarios.

 

Jason M. Schultz: Assessing whether the use of copyrighted data for AI training constitutes fair use requires considering the AIGC itself. Even if a company solely relies on New York Times web pages to train its AI system, it might generate millions of unique expressions, with only 0.1% regarded as potentially infringing. A crucial debate centers on the difference between AI companies automatically scraping data for training and obtaining specific authorization from copyright holders. Two main concerns arise from this issue: First, competition. To encourage competition among different AIs, we cannot limit access to training data exclusively to the wealthiest companies. However, acquiring licenses from all copyright holders can be extremely costly. Second, bias. Using a broader range of data for AI training helps prevent biased language. For instance, in the U.S., the left opposes AI using its data for training, while the right is more accommodating. If a licensing regime is implemented, could training data become dominated by right-wing perspectives? Furthermore, the AIGC largely depends on the prompts, making it challenging for courts to determine whether AI has truly copied a specific book. While users can force AI to copy, it is uncertain how many would actually do so. This leads to another question: if the AIGC does exhibit substantial similarity, who should be held responsible—the user or the AI service provider?

 

James Grimmelmann & Jason M. Schultz: There is no one-size-fits-all answer to this question, as the purposes and contexts in which users employ AI vary. We can only analyze the question on a case-by-case basis. Copyright law should be reasonably tolerant of private space. In the case of the Beijing Internet Court, where the AIGC is uploaded into the public domain, the nature of the issue changes, especially when it involves unfair competition. In sum, generative AI is still in its early stages of development, and there are no definitive answers to copyright issues. We must rely on more judicial cases to expand our understanding. A well-designed legal system should enable both large and small companies to flourish, guarantee equal access to AI for everyone, and promote fair competition. If the law only permits big companies to negotiate with other big companies, it does not resolve the problem, and artists will not receive fair compensation. Moreover, the alleged financial losses suffered by copyright holders might result more from increased competition, rather than just someone stealing their works through ChatGPT.

​

 

Chinese Experts

 

Ge Zhu: In the Chinese “AI Text-to-Image” case, the plaintiff used a large AI model to generate an image called “Spring Breeze Has Brought Tenderness.” The defendant, a poetry author, used the image as an illustration while publishing his poetry. The plaintiff claimed that the defendant removed the signature watermark from the image and uploaded it to social media, violating the plaintiff’s right of authorship and information network dissemination rights. The court ruled that the image possesses identifiable differences and the plaintiff’s original intellectual investment, which meets the definition of “works” in China’s Copyright Law. The image was considered a work of art, and the copyright therefore belonged to the plaintiff, since the AI model itself could not be considered the author.

 

Regarding the intellectuality element, it requires the reflection of a natural person’s intellectual investment. In generating the image, the plaintiff made significant intellectual contributions, such as character design, selecting prompt words, arranging the order of prompt words, and choosing images that met expectations. As for the originality element, it mandates that the work be completed independently by the author and exhibit original expression. Determining whether the AIGC reflects the author’s personalized expression must be assessed on a case-by-case basis. In the present case, the plaintiff arranged and selected expressive details, such as image elements, layout, and composition, based on his aesthetic preferences and personal judgment, all a reflection of his own will. The large AI model functioned as the author’s creative tool, akin to a paintbrush or a camera.

 

Furthermore, applying the law in novel and challenging cases necessitates balancing various interests, including the interests of both parties, the groups they represent, the value choices of legislators, and social and public interests. The present case incentivizes people to use new tools to create, aligning with the legislative purpose of copyright law. As creators increasingly adopt AI tools, the income of software developers may also rise, creating a virtuous cycle that positively affects industrial development. Regarding public interest, it is difficult to distinguish between AIGC and human creations under existing technical conditions. If human creations are protected while AIGC is not, it could lead to negative incentives in society, discouraging individuals from using new tools or hiding the use of AI, potentially infringing upon the public's right to know.

 

Qian Wang: I must respectfully disagree with the presiding judge. Most of the so-called “problems” and “challenges” attributed to generative AI are not real. The only valid question is whether using copyrighted works to train AI falls under fair use. Generative AI does not pose any challenges to the idea/expression dichotomy and substantial similarity. First, let’s consider the idea/expression dichotomy. In the context of AI, we are discussing whether the prompts are considered “ideas” in relation to AI-generated images. If they are, then they are not protected. Our focus is not on determining if the prompts constitute a work or expression, but rather on whether the images generated by AI based on these prompts qualify as expressions of users under copyright law.

 

Suppose an art school teacher, who is also a poet, writes a poem spontaneously and asks a class of 30 students to each create a drawing based on the poem. While the poem written by the art teacher is undoubtedly a “work,” it serves as an idea in relation to the students’ drawings. The poem cannot dictate the composition of each student’s artwork, as they will interpret the poem according to their own thoughts and use their creativity to generate corresponding images. Regardless of how intricate and sophisticated the words used to describe the image may be, they cannot determine its composition. Similarly, even if the prompts are detailed, AI-generated images do not constitute a work. For instance, I once entered an English poem describing a sunset scene as a prompt into two large models, resulting in entirely different images. The description in the poem was detailed enough, yet one could write another 1,000 lines without obtaining identical results. If selecting and inputting prompts is considered a creative act, why does one creative act produce so many varied expressions? The only explanation is that the text, in relation to the picture it describes, serves merely as an idea, not an expression.

 

Second, AI does not challenge substantial similarity, which adopts an objective criterion for assessment. It focuses solely on the similarities between the plaintiff’s work and the alleged infringing content. Whether the content in question was generated by AI or created by humans is irrelevant. Furthermore, the U.S. Copyright Office is unlikely to register the image involved in the Beijing Internet Court case, according to the Office’s four rulings and guidelines. The Office has not accused any applicant of forgery, but instead seeks to clarify whether AIGC can be registered as a work. In the “Théâtre d'Opéra Spatial” case, the Office did not dispute the applicant’s claim that more than 600 prompts were used. However, it still determined that this was not a human-created work because it was autonomously generated by AI.

 

Guobin Cui: Firstly, regarding the idea/expression dichotomy, if the creator only provides prompts, although the prompts themselves may constitute a written work, the images generated by AI based on these prompts typically do not contain the creator’s original expression. It is only after the creator selects an image and then repeatedly modifies the expressive details or compositional elements through prompts or other methods that the expressive aspect of the image can be considered original. The difference between my opinion and Professor Qian Wang’s is that he believes even in the latter situation, originality is still not present.

 

Secondly, I concur with Professor Qian Wang regarding the likelihood of the U.S. Copyright Office granting copyright protection to the image in the Beijing Internet Court case. The Office has denied protection to many AI-generated images, asserting that they lack originality. The image in the Chinese case involves even fewer prompts and modification details than its U.S. counterparts. For instance, in the “Théâtre d'Opéra Spatial” case, after the creator selected the image, he first established the larger framework, then modified the details, and repeatedly used traditional tools such as Photoshop to refine the content. The entire process took more than 80 hours. However, the U.S. Copyright Office still believes it lacks originality, which I consider an overly strict standard. If this standard were applied to the image in the Chinese case, it is impossible that the Office would recognize its originality.

 

Finally, concerning infringement and fair use, both Professors Grimmelmann and Schultz generally agree that using copyrighted works for AI training may constitute fair use. Professor Schultz primarily considers two aspects: competition and neutrality. First, imposing licensing requirements for AI training processes could hinder fair competition, not only between companies but also between countries. Second, if some copyright holders agree to license their works while others do not, it may result in a biased position in the AIGC. Moreover, the two U.S. professors seem to suggest that fair use may apply more readily to non-commercial purposes, while purely commercial purposes warrant further examination. However, I argue that even purely commercial purposes should be considered fair use. If commercial AI companies are required to pay substantial licensing fees for all training data and identify every individual’s contribution, it could lead to market failure and increased social costs, which are entirely unnecessary. Of course, if the AIGC infringes upon copyright, the legal liability of the content at the output stage should be investigated. This issue is separate from whether the use of data during the training stage constitutes fair use. They are two distinct matters.

 

 

 

Discussion

 

Jason M. Schultz: I agree that when prompts are more constructive and creative, the connection between them and the AIGC becomes stronger, thus facilitating the “transmission” of original expression. However, if only a few prompts are provided and the AI is left to complete the work, achieving this “transmission” effect becomes challenging. This intricacy is what makes the thought/expression dichotomy so interesting.

 

James Grimmelmann: Judge Ge Zhu raised an intriguing and thought-provoking point concerning the motivations and incentives for people to use AI in producing artistic works. The objective distinction between AIGC and human-created works is minimal. If copyright law were to declare that all AIGC is not protected, many individuals might be tempted to use AI while lying about and denying its involvement in their final products. Although I am unsure if this issue can ever be fully resolved, a legal system that establishes stark distinctions between rights in human-generated and AI-generated content could indeed create negative incentives.

 

Guobin Cui: I agree with Professor Schultz. When AI generates an initial image and a human creator uses detailed prompts to modify specific features of that image repeatedly, the creator may indeed make original contributions to the final image. Professor Qian Wang posits that text cannot define the expressive elements in images. However, this view is not always accurate, particularly when prompts are so detailed that they precisely define numerous pixel-level features of an image. In fact, any digital image can be described through words and defined by a computer program. Program code is akin to written expression. Thus, it would be incorrect to assert that text-based prompts can never contribute to the expression in an image. However, as previously mentioned, in most cases, a single round of prompts does not result in original contributions to AI-generated images. I also agree with Professor Grimmelmann that a failure to protect any AIGC could lead to negative societal attitudes toward AI usage. We should not just assume that original works cannot be produced using AI tools. In the context of AI’s deep integration with commonplace tools like Photoshop, users are almost certain to make personalized modifications to the AIGC. In such cases, it becomes meaningless to emphasize that AIGC cannot be protected by copyright law.

 

Qian Wang: Firstly, in the “Théâtre d'Opéra Spatial” case, I have no objections to the idea that AIGC may become a protected work after being modified by individuals using Photoshop. The U.S. Copyright Office also does not state that AIGC processed through Photoshop cannot be registered. In fact, the Office only required the applicant to waive their rights to the pure AIGC, but the applicant refused, resulting in the eventual failure to register.

 

Secondly, regarding the formation of a copyrighted work through multiple rounds of modifications, I have conducted experiments with Stable Diffusion and MidJourney. I initially prompted them to draw Chinese-style girls, and the images generated by the two AIs were entirely different. Next, I asked them to add glasses to the girl, which both generative models accomplished. However, Stable Diffusion also added a third hand to the girl, making the image look quite unsettling. The crux of the matter lies in the third step. When I specifically requested to reduce the height of the girl’s glasses frame to 2/3 of the original, both AI systems failed to achieve this. This is because AI currently cannot comprehend user instructions as humans do. It can only generate new images based on its own training and algorithmic rules. Consequently, no matter how many rounds of modifications there are, the user lacks control over the content generated in each round. In other words, each round of AIGC remains a black box, and humans cannot predict the final outcome, regardless of the number of rounds.

 

Thirdly, Professor Guobin Cui suggested that an image can be fully described by dividing it into numerous grids on the screen and then detailing its features at the pixel level. I recall when I first learned computer science, I input the Mona Lisa into the computer. My method involved entering coordinates rather than describing the painting in human natural language. These numerical values were communicated solely to the computer, unlike the AI text-to-image scenario we are discussing today. If we were to use natural language to describe an image, it would never perfectly match the image generated by AI, no matter how detailed. The only exception would be if AI evolves to the level depicted in the movie “Inception,” where it can infiltrate the human brain and accurately replicate a completed image. However, in this situation, it should be referred to as “replicative AI” rather than “generative AI.” Moreover, in the U.S., despite AIGC not being protected by copyright law, the number of AI users has not decreased. This suggests that the protection of AIGC through copyright does not necessarily influence the willingness of users to use generative AI.

   

Angela Zhang: Judge Ge Zhu argues that granting copyright to AIGC can encourage people to use AI. However, if AI becomes the primary tool for creation, human original works may diminish, potentially leading to data scarcity. This is because training large models still relies on works created by humans. Studies have shown that feeding a large model only on AI-generated data can cause its performance to decline over time. In the long term, promoting human-created content is vital for both artistic creation and AI development. What are your thoughts on this?

 

Jason M. Schultz: To cultivate an ideal creative economy, we should raise the bar for originality in AIGC. Currently, it is too easy to produce content with generative AI, so people do not need additional rewards and incentives to use it. In some cases, this may stifle creativity. New laws and regulations should establish thresholds based on human creations, granting certain rights to the AIGC that meets specific standards. This approach would reserve space for human creativity while still acknowledging AI-generated works.

 

Guobin Cui: We must maintain our faith in art. Content generated solely by AI cannot rival the work of genuine artists. Encouraging artists to use AI is just the beginning; continuous refinement and improvement are necessary during the creative process to give the work a soul. If an artist cannot achieve this, their work, if replaced by AI, cannot be considered true art. In response to Professor Qian Wang’s skepticism about AI’s ability to modify specific features of selected images, such issues as the sudden appearance of a third hand or the inability to adjust the height of glasses are technical problems that can be resolved by users. If a third hand appears, it might be due to unconstrained prompts. If the height of glasses cannot be modified, it may be because the user didn’t add brackets after the keyword “glasses” to provide machine-readable prompts or didn’t install the appropriate plug-in. Stable Diffusion is an open-source software with numerous developers creating plug-ins that enable humans to modify specific content in images using text, graphic commands, or keyboard operations. The possibilities are vast, and as AI technology advances, the distinction between text-based and Photoshop button-based modification methods will become irrelevant. In the future, seamless integration between the two will be achieved, marking an inevitable trend. Professor Qian Wang’s example of AI’s inability to modify specific features of selected images does not prove that text prompts cannot specifically or completely modify an AI-generated image. It simply indicates that existing AI technology has not yet reached its full potential, or it may stem from users’ inaccurate understanding of the technical possibilities of AI plug-ins.

 

Ge Zhu: I’ll respond to the question of incentives. First, AI enables individuals without traditional artistic skills to enter the art market and showcase their creativity. Second, many artists have incorporated large AI models into their toolset, potentially replacing repetitive tasks in their workflow. Regarding the issue of value, people may prefer handmade items despite their higher cost. As AI-generated works become more common, handcrafted works will become increasingly scarce and valuable. In addition, there are currently some barriers to using AI software, and users should be encouraged to invest more time and effort into learning these tools. Lastly, according to the mainstream view in China, originality is an all-or-nothing matter. Based on existing standards for “works,” a significant number of AI-generated images can meet the “originality” requirement, as the focus is on human input.

 

Qian Wang: I’ll respond with three concise points. First, my previous example was not meant to insult AI, but to illustrate the practical application of AI technology. Second, when discussing “tools,” we refer to “creative tools,” not “tools” as in “workers as tools for capitalists to make money” or “AI as a tool for humans to transform the world.” A creative tool should not participate in the content creation decision-making process; otherwise, it cannot be considered a creative tool. Third, whether AIGC can be protected as a work is unrelated to whether originality is an all-or-nothing question or a matter of degree. Any analysis of originality and intellectual investment is meaningless without considering the idea/expression dichotomy. For instance, E=mc², despite Einstein’s intellectual input and originality, is not a work because it falls within the realm of an “idea.”

 

 - The End - 

 

Angela Zhang: Thank you to all our esteemed speakers for their insights, and to the audience for their participation! This event marks the beginning of the Generative AI Governance Summit Dialogue Series. We hope you will continue to support us in our future endeavors!

 

Linghan Zhang: Today’s discussion was incredibly engaging. I’d like to extend my gratitude to all the speakers for their valuable insights and stimulating conversations. I eagerly anticipate future summit dialogues being just as productive as today. Thank you all!

​