COPYRIGHT PROTECTION OF ARTIFICIAL INTELLIGENCE GENERATED WORKS FROM AN INDUSTRIAL PERSPECTIVEChu MengI. INTRODUCTIONWith the emergence of Midjourney and ChatGPT, the role of artificial intelligence (AI) has shifted from assisting in creation to content generation. Its high efficiency in creativity and realistic effects that are indistinguishable from human creation have quickly made it a favorite in the industry. As AI in the stage of assisting creation is considered an extension of human wisdom, the works thus created are mere ‘works created by humans using technological tools’. However, generative artificial intelligence completely changes the status of artificial intelligence and humans in the creative process. With just a few prompts, accompanied by a selection of models, styles, image sizes, and qualities, Midjourney can generate aesthetically appealing images. By inputting a subject for creation, ChatGPT can build the overall framework of an article. Nowadays, we only need to give our creative ideas to artificial intelligence, and it can create ideal works according to our wishes. Without exaggeration, generative artificial intelligence is leading a new pattern of creation that liberates human hands.The emergence of new patterns of creation has brought about challenges to copyright law, one of which is how to define the copyrightability and ownership of artificial intelligence generated works. In recent years, there have been numerous studies on this topic, but a consensus has yet to be reached. For example, there is debate regarding whether ‘originality’ should be determined objectively or subjectively, whether copyright should belong to the software designer, the owner, or be jointly held by both, and whether neighboring rights and fruit theory should be considered. Furthermore, many of these studies lack responsiveness to new technological scenarios. Unlike artificial intelligence as an assisting tool, generative artificial intelligence possesses ‘emergent capability’. It has a range of human-like ‘intelligent abilities’, including common sense reasoning, question answering, translation, and others. Therefore, constructing legal regulations in light of new technological scenarios becomes a new challenge presented to us by the era.In July 2023, the Interim Measures for the Management of Generative Artificial Intelligence Services was promulgated, which, as stated in article 3, proposes a regulatory concept that combines inclusive and prudent approaches with classification and hierarchical supervision, emphasizes the development and security of generative artificial intelligence, and promotes innovation and governance under the rule of law. These principles should serve as the starting point for rethinking copyright regulations in the era of generative artificial intelligence. In other words, the discussion on copyright protection for generative artificial intelligence should revolve around the attributes of the ‘law of industry’ within the copyright law, and assess the specific paths to promote the development and innovation of the copyright industry. Guided by this approach, this article will discuss whether there are deficiencies in the existing laws, and propose reform solutions at the legislative level, taking into account the lagging nature of the law and the needs of industry development.II. REGULATORY INADEQUACY REGARDING AI-GENERATED WORKSThere is regulatory inadequacy in current laws regarding the copyright protection of artificial intelligence generated works, which is evident in four aspects. The first is the inapplicability of originality judgment without the existence of natural persons. The second is the delayed application of ownership rules. The third is the statutory type of neighboring rights and disputes around its desirability. The fourth is the insufficiency of fruit theory on the explanation of intangible works.A. The Inapplicability of Originality Judgment Without Existence of Natural PersonsCopyright protection is premised on the existence of original works, with ‘originality’ being a combination of ‘independent creation’ and ‘minimal creativity’. A widely disseminated viewpoint suggests that if artificial intelligence generated works cannot be distinguished from works created by humans, they can be protected by copyright law due to their originality. This article does not hold this viewpoint. In fact, the concept of ‘originality’ in copyright law presupposes the involvement of natural persons, which can be inferred from the literal and historical interpretation of copyright law.Modern copyright law follows the approach of ‘defining creative acts based on originality’ and ‘defining authors based on creative acts’. The concept of ‘creative act’ plays a bridging role, determining the extension of the concepts of ‘originality’ and ‘author’. In China, according to article 3 of the Regulations on the Implementation of the Copyright Law, ‘creative acts’ refer to ‘intellectual activities that directly produce literary, artistic, and scientific works’. Therefore, the key issue lies in defining ‘intellectual activity’. From a definition perspective, ‘intellect’ refers to the ability of humans to comprehend and understand objective things, apply knowledge and experience to problem-solving, and includes memory, observation, imagination, thinking, judgment, etc. Technically speaking, the core of artificial intelligence’s creative implementation lies in algorithms, following the information processing intelligence upgrading pipeline of ‘data-information-knowledge-wisdom-enlightenment’. However, it does not possess genuine intellect. It is evident that intelligence is inherent to humans, not artificial intelligence. The first artificial intelligence case in China which refused to grant copyright protection based on the argument that ‘literary works should be created by natural persons’ was a correct application of literal interpretation of the law. In the process of the third amendment to the Copyright Law in 2022, artificial intelligence creation had already emerged, but the legislation did not provide a specific response, which indicates that the element of ‘human’ in the concept of ‘originality’ continues to be indispensable.In the process of generating artificial intelligence creations, the main role of artificial intelligence is not to assist in creation but to directly produce works that possess minimal creativity in their form. Even though the prompts inputted can be quite common and lack originality, such as ‘beautiful lovely girl, long hair, blue eyes, with a hat’, Midjourney is capable of sketching out unique pictures of artistic value based on such simple prompts in terms of composition, lines, and color choices. This transformation from creative ideas to tangible works is entirely achieved by Midjourney itself, without any creative contribution from a natural person. According to current laws, such AI-generated outcomes cannot be recognized as works with originality.B. The Delayed Application of Ownership RulesSome scholars propose that artificial intelligence creations can be protected by copyright law from the perspective of distinguishing between the judgment of a work and the rules of copyright ownership. They argue that copyright ownership can be determined based on copyright protection for ‘commissioned works’ and ‘works of a legal person’. According to the Copyright Law, works that are created under the supervision of a legal person or non-legal person organization, represent the will of the legal person or non-legal person organization, and are subject to the responsibility of the legal person or non-legal person organization, are considered as works authored by the legal person or non-legal person organization. For commissioned works, copyright ownership is determined by the contract between the commissioning party and the commissioned party. If the contract does not provide explicit provisions or if there is no contract, the copyright belongs to the commissioned party. Formally speaking, artificial intelligence creations are carried out under the organization of the user, triggered by the user, and represent the will of the user. If infringement is established, the user who publishes the infringing content should bear the liability for infringement. In this context, it seems reasonable to consider the users as the copyright owners when they are legal persons. As for non-legal person users, since artificial intelligence is creating under the instructions of the users, a ‘virtual contract of the commissioned work’ can be established between the machine author and the human, and the transfer of rights to humans can be realized based on the rule of ownership for commissioned works. However, the above views should still be subject to scrutiny, as they violate the fundamental legal principles regarding the rules for determining ownership of works.The delayed application of ownership rules refers to considering the issue of copyright ownership only after the existence of a copyright-protected work is determined. Both of the aforementioned provisions, based on their literal meaning, have the prerequisite of the ‘existence of a work’ for their application. Since artificial intelligence creations do not meet the requirement of originality and cannot be termed as ‘works’, there is no possibility of applying the provisions on copyright ownership. Furthermore, the application of ‘works of legal persons’ is premised on the creative efforts of natural persons behind them. As to contracting parties, the commissioning party and the commissioned party must meet the basic requirements of legal subjects under civil law, for which artificial intelligence lacks eligibility. Last but not least, the fact that artificial intelligence cannot reach an agreement with users through negotiation provides further evidence for this conclusion.C. Statutory Types of Neighboring Rights and Disputes Regarding Its DesirabilityIn addition to the author’s right, the Copyright Law of China also separately regulates neighboring rights to protect the interests of investors and disseminators of works. There is still debate in the academic community regarding whether the subject matter of neighboring rights protection should possess originality. Affirmative viewpoints deny the distinction between copyright and neighboring rights while opposing viewpoints argue that the ‘lack of originality’ standard is a logical premise of the neighboring rights system. If the subject matter of neighboring rights needs to possess originality, the same presupposition of natural persons would also be applicable to neighboring rights, which means artificial intelligence generated works would not be eligible for the protection of neighboring rights. Therefore, the issue left to be explored is whether the current neighboring rights system can be applicable to artificial intelligence generated works if the subject matter of neighboring rights does not need to meet the requirement of originality.Unlike the author’s right system which adopts an open attitude towards types of works, the types of neighboring rights are strictly limited to layout design rights, performer’s rights, sound and video recording rights, and rights of the broadcasting organization, without a cover clause. If works generated by artificial intelligence do not exist in the form of layout design, performance, sound and video recording, or broadcasting, such as being solely in the form of text or images, they cannot seek neighboring rights protection. Even if it is argued that ‘the random creative behavior of artificial intelligence corresponds to the performance of performers according to a script’, the Regulations on the Implementation of the Copyright Law define ‘performers’ as ‘actors, performance units, or other performers of literary or artistic works’, which implicates a human being behind and does not include artificial intelligence. Moreover, the historical development of neighboring rights also shows that even if it does not require human creative behavior, the corresponding outcomes are still produced by human beings. Neighboring rights protection does not, as some scholars have claimed, ‘evade the issue of the subjectivity of artificial intelligence’. On the contrary, there is still a presupposition of labor input by natural persons. Furthermore, due to the controversy surrounding the nature of neighboring rights, the academic community has skepticism about the neighboring rights system, with the main reasons as follows. Firstly, the introduction of neighboring rights into a separate system was not due to differences between subject matters, but rather copyright holders’ concern that granting neighboring rights would lead to a decrease in their royalty income. As a result, a new system of neighboring rights is established, which emphasizes the concept of prioritizing copyright protection over neighboring rights protection. However, a similar relationship also exists between original works and derivative works. Since derivative works can be protected under the same legal framework as original works, it is hard to understand why neighboring rights cannot. Secondly, whether it is the personal experience of the performer that constitutes the essence of performance, the arrangement and editing of sound sources that reflect certain ideas in sound and video recordings, or the collection, selection, and processing of information that forms broadcasting signals and programs, all of these aspects demonstrate originality. In other words, neighboring rights and copyright have the same qualitative nature, and the concept of neighboring rights can find its position within the copyright system. Therefore, building copyright protection for artificial intelligence generated works on the precarious neighboring rights system would not be a wise move.D. The Insufficiency of Fruit Theory on the Explanation of Intangible WorksGiven the regulation insufficiency of artificial intelligence generated works under the current system of copyright law, some scholars have proposed an alternative approach by suggesting the application of fruit theory from civil law to provide protection for artificial intelligence generated works. Fruits refer to the products derived from an original object, including natural fruits that arise from the natural attributes of the object, such as fruits from planted trees or milk products from livestock, and legal fruits that arise from a legal relationship with the original object, such as rent from renting out a house or dividends from shares according to capital. Since artificial intelligence does not have a legal personality, the works generated by it based on human input exhibit characteristics of being ‘derived from an object’ in their appearance, thus resembling the basic form of fruits. This is the primary reason why scholars have proposed fruit theory for artificial intelligence generated works.In the author’s view, fruit theory does not provide sufficient explanation for intangible objects like artificial intelligence generated works. The specific reasons are as follows. Firstly, artificial intelligence generated works are the result of a series of human interventions, such as inputting data, training algorithms, selecting prompts, and initiating the creative process. They are not naturally derived from the attributes of artificial intelligence and thus do not possess the characteristics of natural fruits. Secondly, apart from the fact that legal fruits require the use of the original object by others, the income and remuneration obtained from using one’s own property do not fall under the category of ‘legal fruits’; since the legal fruit is expected to generate regular income based on the use of an object, artificial intelligence generated works may not fulfill this requirement, especially in cases where payment is based on usage rather than at regular intervals. Thirdly, just as the fruit of a tree or the rent of a house, a key characteristic of the fruit is that its value should not far exceed that of the original object. However, artificial intelligence generated works may not necessarily meet this condition. The painting ‘Portrait of Edmond de Belamy’ which was sold at Christie’s auction house in New York for over $430,000 in 2018 serves as a good illustration. The reason behind this high value lies in the scarcity of this specific artificial intelligence generated work, indicating that the value can indeed surpass that of artificial intelligence software. Finally, these three reasons can ultimately be attributed to the differences between real property rights and intellectual property rights, which stem from the opposing nature of natural and human-made objects, as well as the distinction between abundance and scarcity. In fact, copyright law governs derivative works with a totally different logic compared with fruit theory, by emphasizing the creative input of derivative authors, while ensuring that the exercise of rights of derivative works does not infringe upon the rights of the original work as a balance. This is fundamentally different from the provision in article 321 of the Civil Code, which deals with ‘legal fruits’ and allows for fruits to be obtained according to agreements or customary practices if no agreement is reached. Such distinctions were intentionally made by the legislator. Since its application to intangible objects would be impractical, it is evident that fruit theory lacks explanatory power for artificial intelligence generated works.III. JUSTIFICATIONS FOR THE PROTECTION OF AI-GENERATED WORKS FROM AN INDUSTRIAL PERSPECTIVEIn the fast-paced development of technology, the issue of the lack of responsiveness to law has become increasingly prominent. The current regulation insufficiency of artificial intelligence is just one manifestation of this situation. However, the absence of regulations is merely a descriptive statement and not a value judgment. Whether copyright law should respond to artificial intelligence generated works or continue with its current approach of addressing hot issues with a lukewarm response should be considered based on legitimate considerations. The ultimate answer should be found by examining the nature of copyright law as a ‘law of industry’ and determining which approach is more conducive to realizing the values of copyright law.A. The Dual Value of Copyright Law in Industrial Development: Incentive and OrderFor a long time, incentive theory has been regarded as the foundation of copyright protection. Since creative works possess intangible attributes and can be disseminated at very low costs, but the act of creating such works is not an overnight endeavor, those who distribute the works at lower costs could potentially drive the creators out of the competitive market if no protection is given to the latter. In this scenario, creators may lose their motivation to create, leading to a shortage in the supply of creative content in the market. Therefore, it is necessary to provide authors with some economic incentives to encourage them to engage in creative activities that contribute to cultural and social prosperity. Copyright is one of such incentive mechanisms. By granting copyright holders exclusive rights for a specified period, creators can enhance their income, creating the conditions for a professional approach to creation. The professionalization of creation further normalizes employment in the creative industry. Through copyright law provisions concerning works of legal persons, works of employment, cinematographic works, commissioned works, and other ownership rules, the incentive theory takes on an additional dimension, incentivizing organizational creative activities. As a result, the copyright industry continues to grow and thrive.Incentive theory has a certain explanatory power in establishing rights to intangible property, although ‘the connection drawn by the orthodox account between IP rules and innovation is less strong and direct than commonly believed.’ Due to a lack of sufficient empirical data, the incentive theory often becomes a form of rhetoric, employed by copyright holders when the existing allocation of benefits is affected by technological changes, in order to achieve distribution schemes more favorable to their own interests. Exercising ‘protection of investment’ under the guise of ‘incentivizing creation’ has to some extent become the new norm. However, there is an essential difference between incentivizing creation and protecting investment. Just as Paul Goldstein argues, proponents of the doctrine of incentivizing creators are pessimists about copyright, willing to extend copyright protection only to the extent necessary for the incentive. On the other hand, proponents of the doctrine of protecting investors are optimists about copyright, fixated on the half-filled glass of copyright water and eagerly waiting to fill it further to maximize investment returns. Complete protection for investors often results in the reduction of rights for users, which will inevitably disparage the principle of balance and rationality of copyright law.Apart from its incentive value, copyright law holds another significant value for industrial development, which is order. The order has a dual meaning of both form and substance. In form, the order represents a stable and consistent state, a mode of thinking that involves addressing similar issues with similar solutions. This mode of thinking encourages us to consider the adaptability of existing legal frameworks when faced with problems arising from new technologies, rather than starting from scratch. In substance, order implies that people can expect specific outcomes when they act in accordance with certain rules. Only through this can behavior expectations be maintained, thus promoting the stability of societal life. The value of order provided by copyright law has four dimensions.Firstly, the value of order provided by copyright law is industry-oriented. Being a natural inclination of humans, ‘creation’ has a history much longer than that of copyright law. It was not until the development of the printing press which facilitated the rise of publishers that copyright law emerged to maintain order in the market transactions of works. Although the historical event of the ‘literary property debate’ resorted to the ‘personalities of authors’ to argue for perpetual property rights, it still served industry interests, and creators’ relevantly weak status did not change. In other words, the demand of the interests of investors in the industry is the original driving force of copyright legislation. Nowadays, with the increasing organization and industrialization of the creative process, it has become normal for enterprises to be copyright holders. Without a doubt, copyright law is essentially a system that recognizes and allocates the benefits generated by the marketization of works.Secondly, the value of order provided by copyright law is directed at intangible works with aesthetic appeals. Unlike subject matters protected by patent law or trademark law, the value of works does not lie in their utility or in satisfying a particular desire, but rather in the pleasure derived from their aesthetic value, their allure to individuals, and their fulfillment of spiritual and cultural needs. Following the approach of defining rights through subject matters, the boundaries of copyright law, trademark law and patent law are relatively clear, reflecting the inherent order of intellectual property law.Thirdly, the value of order provided by copyright law is a set of organized thought patterns. While remedies offered by anti-unfair competition law have the characteristics of vague scope, case-by-case approach, and lack of rule guidance, copyright law sets clearer boundaries for the behavior of copyright holders, disseminators, and the public. Reverting from copyright law, which possesses a higher degree of order, to anti-unfair competition law, often implies regression unless there are substantial reasons for doing so.Lastly, the value of order provided by copyright law is also future-oriented. In response to new business models in the copyright industry brought about by technological developments, copyright law usually responds in a gradual manner. If the application of new technology undermines creative incentives, new rights are added to the copyright system as a response. The ‘making available right’ introduced to adapt to internet development serves as an illustration. If the application of new technology compromises dissemination efficiency, alternative solutions are sought through licensing arrangements, such as enlarging the scope of statutory licenses, improving the operation of collective management organization, and establishing general public licensing rules to address the impact of user remixing under participatory culture. These measures of improvement imply the fundamental assumptions of protecting copyright and achieving balance. In other words, abandoning copyright protection has never been a desirable option. Copyright has become a way of thinking, which itself is an important source of order.B. The Impact of Entitlement on the Realization of Incentive Value in the Copyright LawWhether AI-generated content should be protected by copyright requires a discussion centered on both incentive and order values. In this discussion, the author will adopt a prospective viewpoint, examining how different legal provisions may affect the behavior of future and unspecified parties from an economic analysis perspective, and thus evaluate their impact on the value of copyright law. The discussion begins with the incentive value.Firstly, if AI-generated content were to be placed in the public domain, the market space for human-created works would dwindle, thus undermining the incentive for human creativity. Research suggests that with increased acceptance and improved productivity of AI, the Chinese AI painting market is likely to experience explosive growth after 2022, expanding from 10 million yuan in 2021 to 15.466 billion yuan by 2026. Accompanied by this growth, a change in the landscape of the creative market will emerge, where individuals gradually shift from purchasing licenses to use copyrighted works to creating AI-generated works that better suit their needs or directly using AI-generated content that is not protected by copyright. This shift will inevitably decrease the sales market for human-created works. Moreover, a new market is also emerging where AI trainers are required to obtain authorization and pay for their reproduction of copyrighted works. Copyright lawsuits related to AI training data infringement, involving software such as Midjourney, Stable Diffusion, and ChatGPT, have already been widely reported. These indicate the desire of the creative industry for this emerging market. In commercial practice, Google and Universal Music Group are already negotiating on licensing voices for AI-generated songs. If AI-generated content is certain to be unprotected by copyright, while the need for AI training to obtain authorization from copyright holders remains in dispute, or receives affirmative feedback in court decisions, rational trainers of AI will be more inclined to use AI-generated content for data training to mitigate the risk of copyright infringement. This would result in the loss of licensing fees for human creators in the emerging market, which is not conducive to realizing the incentive value.Secondly, following the above line of thought, if trainers are more inclined to use AI-generated content for data training, the prevalence of low-level repetition in the creative market will inevitably occur, which is detrimental to the prosperity of society and culture. It is undeniable that AI does enhance the creative performance of individual users to a certain extent. With just a few prompts, AI can provide highly divergent answers, sparking creators’ inspiration and overcoming creative barriers. However, if we consider human creativity as a whole, it is difficult to arrive at the same conclusion. Creativity, as a human trait, has its biological basis. Abilities like pattern recognition, association, and synthesis cannot be easily obtained by machines. Creativity is also based on social psychology, which emphasizes the role of intrinsic motivation. Factors in the social environment, such as working under supervision and using external material rewards for motivation, can weaken individual creativity. Since AI generates content as a result of model training and data recombination, it lacks the desire and intrinsic drive for creation. Moreover, in order to ensure that AI-generated content follows users’ instructions correctly, human supervision is usually needed in the AI training process. With the combination of low-level intrinsic motivation and high-level extrinsic motivation, we can safely reach the conclusion that AI itself does not possess a high degree of creativity. Thus, the extent of creativity in its outputs depends more on the quality of inputted works. In fact, the main reasons for the development of AI-generated content and its quality improvement lie in the development of pre-trained large models and the enhancement of data quality. If human creative drive is reduced and AI trainers lack the necessary incentive to input human-created works, the society we live in will be filled with works with low-level repetition. This is not beneficial to the development of the cultural industry, nor will it lead to the prosperity of cultural endeavors.Thirdly, providing copyright protection for AI-generated content can create scarcity, thereby incentivizing the use of AI technology and benefiting the industry, which leads to a virtuous cycle. Some scholars have pointed out sharply that at the time of the creation of business intelligence software such as ChatGPT, there were no explicit regulations in copyright law regarding AI-generated content, which precisely indicates that this emerging industry does not require incentives offered by copyright law. This viewpoint might seem convincing at first sight, but unpersuasive after second thought since it overlooks an important fact — the responsive nature of the law. Law is responsive as a superstructure in society, and industry development always precedes changes in rules. Since we cannot regulate a new industry before its emergence, we cannot argue for the adaptability of current rules based on the fact that the industry has developed prior to regulations. In fact, the argument that not granting copyright protection to AI-generated content would significantly inhibit innovation incentives in the AI industry still holds a dominant position. Furthermore, it should be noted that from the perspective of enriching human spiritual life, what we need to incentivize is not just AI-generated content itself, but AI-generated content with high quality. The higher the quality, the greater the scarcity, and the more it can stimulate people’s desire for exclusivity. With the development of AI technology, people are increasingly inclined to use intelligent tools for activities such as generating avatars, creating product posters and wallpapers, etc. The motivation behind this not only lies in the pursuit of aesthetics and efficiency but also encompasses a desire for scarcity. Since it is highly unlikely for different users to generate the same work based on the same prompts, users often expect a certain level of scarcity and exclusivity in the generated content. Once this expectation is undermined by new rules stating that ‘AI-generated content is not entitled to copyright protection’, anyone can freely duplicate and distribute it. Therefore, the unique connection between users and AI-generated content will no longer exist, and such content will lose its irreplaceability and value as a medium for the expression of unique emotion and individuality. The higher the users’ desire for scarcity, the greater the probability that they will be reluctant to use AI-generated tools due to the lack of copyright protection. This fracture in the incentive chain, where high-quality production is no longer supported by high-standard consumption, will ultimately hinder the progress of AI-generated technology.In conclusion, from the perspective of incentive theory, granting copyright protection to AI-generated works is a more desirable option.C. The Impact of Entitlement on the Realization of Order Value in the Copyright LawIndeed, it is an oversimplification of reality to equate the function of copyright solely with incentivizing creativity. Individual creativity is driven by multiple motivations, and copyright plays a relatively marginal role. Taking this factor into consideration, the conclusion regarding the impact of granting rights on the realization of incentive value in copyright law may be somewhat weakened. Nevertheless, considering the order value of copyright law, granting rights remains a more preferable choice.Order refers to a regulated and organized state. When relatively certain conclusions can be drawn based on given facts, a sense of legal order will reach us. As the last source of defense in upholding social fairness and justice, judicial authorities need to base their judgments on judicially determined facts, ensuring that the determined legal truth aligns with objective truth, and only in this way can fair justice and order be realized. This requires that the determination of legal facts should not primarily be the result of allocation of the burden of proof, but should strive to restore the truth as much as possible.As to copyright protection of AI-generated content, the determined legal truth refers to the truth related to the production process of the subject matter deduced by the court based on evidentiary rules, while the objective truth refers to the actual production process of the subject matter. It can be anticipated that when human creative contributions are the criterion for providing copyright protection, defendants in copyright infringement lawsuits will tend to argue that the plaintiff’s work is AI-generated and should not be entitled to copyright protection. Suppose there are two different rules for allocating the burden of proof: the first requires the plaintiff to prove that there is a human contribution in the original expression of the work, while the second requires the defendant to prove that there is no human contribution in the original expression of the work. In the first scenario, even if the plaintiff can provide evidence such as drafts or different layers in graphic software, as long as the entire creative process is not fully recorded, we still cannot ascertain whether the provided information is the result of human creation or copied from AI-generated content. The burden of proof for the plaintiff becomes extremely excessive. In the second scenario, the defendant lacks the necessary information regarding the generation process of the plaintiff’s work and is likely to bear the consequence of failure to fulfill the burden of proof. As a result, the final judgment will largely depend on the allocation of the burden of proof, and little correlation will stand between the determined facts and the objective facts. This clearly deviates greatly from the value of order required by copyright law.Continuing along the same line of thought, the value of the presumption of authorship rule will also be deconstructed with the emergence of AI-generated content. Article 12 of the Copyright Law of China states, ‘The natural person, legal person, or non-legal person organization indicated as the author on the work shall be presumed as the author and has corresponding rights unless proven otherwise.’ This is the legislative expression of the presumption of authorship rule. This rule encompasses dual meanings of ‘presumption of authorship’ and ‘presumption of rights’. The latter meaning was added in the third amendment to the Copyright Law, the purpose of which is to determine whether the work claimed by the plaintiff satisfies the conditions for copyright protection, i.e., its essence lies in the presumption of the subject matter to be protected. In other words, as long as the plaintiff provides the external manifestation of the object, no further evidence is needed to prove that the claimed work possesses originality. In the case of preliminary evidence and the absence of contradictory evidence, the subject matter requirements are met, the claimed rights are established, and these rights exist in the claimed subject matter. The presumption of authorship rule follows the simple logic of ‘presuming the process based on the result’, namely, if the author’s name is attached to the work at the time of registration or publication, it can be presumed that the person is the author and copyright is vested in the work. This practice relieves the burden by not retracing the creative process, thus aligning with economic rationality. More importantly, as the others have limited access to the work prior to its registration and publication, the correctness of the presumption of authorship can be reasonably ensured. As for the question of whether the subject matter possesses originality, the accused party can provide evidence from many perspectives, including listing materials in the public domain or subject to copyright protection of third parties, resorting to the doctrine of scènes à faire and mix of ideas and expressions, etc. In other words, the presumption does not impose an excessive burden on the accused party. However, once AI-generated content is deemed incapable of constituting a work, the foundation of the presumption of authorship rule, which relies on the presumption of process based on the result, no longer exists. As a consequence, neither of the dual meanings of the presumption of authorship rule remains achievable, which not only burdens the litigants with more legal complexities but also fundamentally changes the functioning mechanism of the copyright system, undermining its inherent value of order.The impact of denying rights to AI-generated content on the value of order extends beyond the mechanism and touches upon people’s fundamental recognition of the copyright system. In 2018, the French team Obvious sold a portrait created using AI for $432,000. In 2022, Jason M. Allen won first prize in the digital arts/digitally manipulated photography category at the Colorado State Fair Fine Arts Exhibition with an AI-generated artwork titled ‘Théâtre D’opéra Spatial’. These demonstrate that we can hardly distinguish AI-generated creations from human creations, regardless of whether copyright protection is granted to AI-generated content or not. If AI-generated content is not protected by copyright, it will be natural for people to question why human works, which may not manifest more originality, individuality, or value in form, can enjoy copyright protection. If anyone can freely copy AI-generated works without legal risk, why should they pay for the use of human-created works? As long as copyright law continues to insist on the premise of a natural person’s original contribution for protection, while readers evaluate the value of works based on their appearance, these questions will only increase as AI technology matures. Over time, the concept of respecting copyright will be eroded, and the humanistic foundation upon which copyright law is built will face more challenges.IV. THE COPYRIGHT ATTRIBUTION SCHEME OF DEEMING AI USERS AS AUTHORSWith the potential harm to both incentive value and order value, which in turn hinders the development of the copyright industry in mind, the reasonable approach is to grant copyright protection to AI-generated content. Similarly, based on a comprehensive consideration of the value of copyright law and comparative analysis of different copyright attribution schemes, and guided by the Kantian concept of ‘subjectivity of human being’, treating AI users as ‘deemed authors’ not only aligns with principles of efficiency and fairness but also can be seamlessly integrated with existing laws. Therefore, it is a more preferable copyright attribution scheme.A. Theoretical Foundation: The Directness of Users as Value-Bearing EntitiesThe rationale behind considering AI users as authors can be derived from the production process of AI-generated content, which can be divided into three stages: data input, algorithm training, and output of the work. AI software designers and owners are mainly involved in the former two stages, while AI software users play a major role in the output stage. Similarly, human creation also goes through three stages: input, processing, and output. Language learning, extensive reading, and accumulation of life experiences constitute the input process, learning writing patterns and acquiring new knowledge based on the inputted content form the processing process, and applying various creative methods to express and present the acquired knowledge in expressive works represent the output process. No work is created out of thin air; it is always a product of authors who have been trained through specific thought patterns in a particular historical context. In this continuous process, not only the works existing in the input stage, the thought patterns existing in the processing stage, but also the works existing in the output stage, deserve to be incentivized. However, the mechanism of copyright law selects the authors who directly produce the works as the direct targets of the system. Since the time-limited copyright protection encourages works to eventually enter the public domain, this mechanism also indirectly incentivizes earlier stages of production. From an analogical perspective, the above reasoning also applies to AI-generated content. In other words, users should logically be the starting point for copyright protection of AI-generated content.Compared with users, AI software designers and owners are not the direct bearers of the incentive mechanism, as they do not directly determine the output of AI-generated works. As we do not attribute copyright ownership to the developers of Photoshop just because they provide creative tools, the creator and owner of AI software are not suitable as the targets of incentives as well. These two categories of entities are not suitable to be copyright owners of AI-generated works for three more specific reasons. Firstly, copyright is not the primary concern for AI software designers and owners, especially for generative AI that extends beyond the creative domain to areas such as strategy generation and emotional communication. The role of incentive theory is almost irrelevant in these cases. Secondly, there are multiple channels for these entities to recover costs due to their control over algorithms and devices. For example, owners can generate revenue by renting out AI devices, while designers can generate revenue by licensing AI software with a fee. Even providing software for free can help expand their user base, thus serving commercial purposes. Therefore, no additional incentive is needed. Thirdly, products of the model derived from training data, including ‘AbyssOrange’ and ‘Counterfeit’ in Stable Diffusion, as well as ‘Chilloutmix realistic model’, etc., are all models of style, which do not represent specific expressions that copyright law seeks to incentivize; conversely, they primarily fall within the category of ideas, which are inherently uncopyrightable. A recent commercial practice of Getty Images in its application of AI tools also reflects the same view. By indicating that ‘content generated through the tool will not be added into existing Getty Images and iStock content libraries for others to license’, ‘further, contributors will be compensated for any inclusion of their content in the training set’, Getty Images is to some extent acknowledging the copyright of users.The copyright attribution scheme that considers users as authors also aligns with the value of order of copyright law. The fundamental principle of copyright, i.e., ‘acquiring the license and paying remuneration for dissemination of works’, is established to regulate the circulation of works in the market through the control of scarcity and the granting of benefits. In other words, copyright law exercises control over the distribution order of works in the market. It is evident that the logical starting point for the industrialization and market circulation of works does not solely reside with the authors themselves, but rather with the entities responsible for pushing the works into the market. These entities predominantly refer to publishers in the age of the printing press, and network content providers who upload works for subsequent dissemination in the age of online publishing. In the era of AI creation, it primarily denotes users of artificial intelligence who decide to introduce works into market circulation. These entities are the direct beneficiaries of the commercialization of works, highly valuing the scarcity of works, and showing a willingness to meticulously construct distribution channels for the purpose of achieving profits. It is precisely from the perspective of industrialization that the value of order and the value of incentive are further unified.B. Practical Foundation: Direct Contributions of Users to Creative ExpressionWhile the originality within AI-generated content often does not originate from human creativity, the ultimate output of AI creations is inseparable from the direct contributions of users. This forms the practical foundation for users’ copyright entitlements. Specifically, the contribution of users exists in various aspects, such as inputting and optimizing prompts, pressing the ‘start’ button, selecting the outputted work, and making adjustments to the work.Firstly, the choice of prompts. Prompts are a method of guiding or stimulating artificial intelligence models to perform specific tasks using natural language. The term ‘prompt’ in English corresponds to ‘prompting’, which reflects the guiding role of the prompt’s inputter in the outcomes of the AI-generated content. In practice, the output obtained based on initial prompts is often too generic and fails to meet user needs. In such cases, further optimization of the prompts is required. This optimization can involve refining the question, specifying formatting requirements for AI-generated outputs, or providing answer samples for the AI to reference, which aims at enhancing the quality of the output. The degree of proficiency of users in formulating prompts varies, which inevitably leads to differences in matching degrees between the final output and the intended requirement. Without exaggeration, the ability to formulate effective prompts has become one of the most crucial productivity factors in the AI era, which directly determines the quality of AI-generated results. Regardless of whether the prompts used by users are creatively original, the significance of their instructions in shaping the final AI-generated outcomes should not be underestimated.Secondly, the act of pressing the ‘start’ button. Despite the continuous enhancement of various capabilities in AI due to technological advancements, its instrumental nature has not fundamentally changed. A pivotal reason for this is the lack of autonomous will and thought in artificial intelligence. In the realm of creation, while artificial intelligence has acquired the potential for creativity through the input of vast amounts of data and algorithmic training, encompassing abilities such as summarization, reasoning, translation, transformation, and paraphrasing, the fundamental fact remains that it lacks the subjective initiative to create. It is the users who have the subjective intention of using AI for content generation and issue generating instructions to the AI, that effectively control the ‘start’ button. This process, though seemingly easy and straightforward, constitutes an indispensable aspect of the production of AI-generated works.Thirdly, the selection of outputted work. The process of AI generation, from production to market entry, involves users’ selection, facilitated by the mechanisms provided by AI software. Taking Midjourney as an example, it can generate four images at one time based on a single prompt. If users find the results unsatisfactory, they can generate again and again until they achieve the desired outcome. This process embodies the user’s selection of AI-generated content. ‘Selection’ is an act of copyright significance. Creative endeavors fundamentally involve the selection and recombination of existing elements such as text, lines, colors, and creative techniques within the public domain. From this perspective, all creative acts can be understood as a form of ‘compilation’. Just as the selection by editors of submitted works and the choice by database operators of included information determine what works and information are presented to the public, the logic behind the protection of works of compilation similarly applies to the establishment of copyright claims of AI users. In fact, even in the renowned copyright registration dispute over the comic work ‘Zarya of the Dawn’, while the US Copyright Office denied copyright to each individual image on the basis of the absence of human authorship, it still recognized Kashtanova’s claim to the copyright of work of compilation based on the selection and arrangement of images. Though as a famous scholar has noted, the grant of copyright requires an examination of whether relevant subjects have exercised their free will to determine the expressive elements of a work, the same scholar also points out that this ‘determination’ is not limited to deciding every detail of the work but can encompass contributions from other sources. Following the same logic, while the originality in AI-generated content mainly stems from AI, the determination of expressive elements of the work suitable for market entry remains within the domain of the user. Since discussing copyright issues outside the context of the market is meaningless, considering the user as the logical starting point is appropriate.Fourthly, the adjustments to the work. Given that AI-generated works often fall short of fully meeting users’ requirements, making adjustments to the generated content is frequently necessary, in order to reflect their own creative intentions and aesthetic standards. In the first highly scrutinized US case on copyright protection of AI-generated content, Kashtanova also asserted her role in the creation of derivative work. However, her claim was not supported due to the minute nature of the adjustments and a lack of sufficient evidence to substantiate her claim. If comprehensive evidence is presented, demonstrating substantial human contributions, establishing the user as a derivative author is not an insurmountable task. In fact, creative tools like Stable Diffusion have already enabled users to exercise precise control, exemplified by the ‘local control’ feature. This function allows users to control specific elements such as the outline, depth of field, character poses, and scene composition of an image. For instance, a user can upload a preferred image to Stable Diffusion, which recognizes its outline and depth of field, and upon clicking ‘apply’, the AI-generated image automatically replicates similar contours and depth. In the creative industry, it has become popular to leverage the randomness inherent in tools like Midjourney for inspirational purposes, followed by utilizing the precise control capabilities of Stable Diffusion to fine-tune the visual effects of the generated content. This creative approach underscores the user’s authority over the output results and undoubtedly serves as compelling evidence of authorship.C. The Basis of Consensus: Ease of Incorporation into Existing LawsThe enforcement of the law comes with costs, the amount of which can be influenced by both the extent to which the incentives offered by the law align with human behavior and the level of understanding and acceptance people have toward the system. The efficiency of granting copyright protection to AI is already analyzed in the former parts of this article and will not be reiterated. This part will highlight that considering users as authors can minimize public adaptation costs to the greatest extent, enhance the acceptance of the system, and contribute to the formation of consensus.Some scholars have explored the cost of implementing the copyright system, noting its advantages of establishing various legally determined forms of usage, with which copyright holders and readers can identify to the maximum extent information related to the rights, thereby reducing transaction costs. The same argument also applies to the rules for determining copyright ownership within copyright law. Prior to the emergence of AI technology, a comprehensive set of rules for determining copyright ownership already existed, which mainly consists of three components. Firstly, there are rules for determining the author, including rules of presumed authorship and the ‘deemed authorship’ principle for works of legal persons. Secondly, there are rules that attribute copyright ownership based on authorship, as reflected in the presumption that ‘copyright belongs to the author’. Thirdly, there are exceptions to determine copyright ownership, primarily encompassing rules for specific types of works such as works of employment, audiovisual works, commissioned works, derivative works, works of compilation and collaborative works. These exceptions can be further classified into two categories. The first category involves rules for derivative works, works of compilation and collaborative works, which are unique due to the involvement of multiple authors or multiple works. The second category pertains to rules for works created in the course of employment, audiovisual works, and commissioned works, primarily reflecting the protection of the interests of investors.With the above rules in mind, there are two possible institutional approaches to incorporating AI factors in order to improve existing rules: one is changing the definition of ‘author’; the other is creating exceptions specifically for AI-generated works. Current research predominantly leans towards the latter approach. However, this design has several drawbacks outlined as follows. Firstly, it lacks pertinence. As discussed earlier, the definition of ‘author’ focuses on the creative process, while special ownership rules primarily emphasize the protection of the interests of investors. While it is true that the application of AI technology heavily relies on the contribution of investors, the direct impact of AI on copyright stems from its alteration of the creative landscape. Since AI directly assists human creators in their specific creative activities, focusing on the ‘creation’ aspect is crucial to enhance the pertinence of institutional design. Secondly, it still leaves significant gaps in the system. Directly stipulating the ownership of AI-generated works does not achieve a comprehensive solution. For instance, if the copyright holder of an AI-generated work creates the work under commission, or if the work produced by AI is incorporated into an audiovisual work as a component, since establishing ownership rules alone cannot modify the underlying premise of natural persons for commissioned works and audiovisual works, the commissioning party and investors are ineligible to enjoy copyright ownership without further agreement. Such an approach in determining ownership based on the involvement of creative tools will ultimately disrupt the expectation of corresponding investors, leading to an inadequate response of the system to real-world scenarios.Conversely, if the approach of considering users directly utilizing AI creative tools as authors is adopted, the issues raised above can be better addressed. For example, in the aforementioned scenarios, by considering the users of AI creative tools as authors, the prerequisite for applying ownership rules for audiovisual works and commissioned works is satisfied. The same also applies to collaborative works. For example, if the first and second parts of a work are generated by individuals A and B respectively, using AI tools, A and B can each be recognized as collaborative authors, and the ownership rules for collaborative works as outlined in current copyright law can be applied for individual portions of the work, as well as for joint exercise of copyright in the collaborative work as a whole. This approach of considering users as authors not only enables convenient integration and incorporation with the existing legal systems but also highlights the contributions of users in the creative process, which aligns with the basic consensus on copyright ownership established through long-term practice. Compared to creating exceptions specifically for the ownership of AI-generated works, this approach is a more feasible solution.V. CONCLUSIONAs a new mode of content production, AI ‘creativity’ has brought revolutionary changes to content producers, production methods, and production efficiency, and has become a representative of advanced productive forces. Regardless of the efficiency and quality of AI-generated content, its essential nature as a tool remains unchanged; it simply presents itself in a more advanced form. Based on the fundamental values of incentive and order in copyright law, granting copyright protection to AI is necessary and beneficial to the promotion and healthy development of the AI copyright industry. Given that AI does not possess a legal personality, it cannot be the copyright owner of the generated works. The legal approach of considering the user of AI-generated content as the author not only promotes the two fundamental values of copyright law but also reflects the user’s direct contribution to the originality of the content. Moreover, it minimizes the adaptation costs of the public to the legal system. Taking all factors into consideration, we can reach the conclusion that this legal approach offers a more feasible solution at the current stage.
