Artificial Intelligence (AI) technology has reached a point where the contents it generates are ostensibly analogous to expressions - a skill once dominated by humans. AI technology at the current level can already generate various types of contents that are analogous to different kinds of human expressions, such as literary, musical or pictorial/graphical works. For example, ‘The Next Rembrandt’, a 3D printed painting is generated by AI technology made solely from Rembrandt’s works.
This project raises a critical question: can AI-generated contents be considered as works?
Artificial Intelligence (AI) technology has reached a point where the contents it generates are ostensibly analogous to expressions - a skill once dominated by humans. AI technology at the current level can already generate various types of contents that are analogous to different kinds of human expressions, such as literary, musical or pictorial/graphical works. For example, ‘The Next Rembrandt’, a 3D printed painting is generated by AI technology made solely from Rembrandt’s works. The team has examined the entire collection of Rembrandt’s work, and analysed a broad range of materials like high-resolution 3D scans and digital files, which were upscaled by deep learning algorithms to maximize resolution and quality. This extensive database was then used as the foundation for creating the painting. In short, it is an AI-generated ‘artwork’ that consists of over 148 million pixels based on 168,263 painting fragments from all 346 of Rembrandt’s paintings, and the whole process took 18 months.[1]
Whether AI-generated contents can be considered as copyrighted works is a crucial question, as the answer could help us ‘identify those (AI-generated) contents that fail the originality test and exclude them from copyright protection.’[2] In other words, it helps to clarify ambiguities and conflicts and to focus on the real question: the copyrightability of AI-generated contents that can pass the originality test. This is a logical corollary of that longstanding copyright theory, which is supported by China as well.
The understanding of originality varies from country to country. In United Kingdom (UK) and Hong Kong, originality is not concerned with whether the work is inventive, novel, or unique, it only cares about if it is originated from the author and is not copied from elsewhere. Specifically, the author must have exercised the requisite ‘labour, skill, or effort’ in producing the work. This is the so-called ‘Sweat of the Brow’ principle.[3] But after the Infopaq case,[4] the originality standard adopted by the European Union (EU) requires that the work must be the ‘author’s own intellectual creation’ rather than merely a work of ‘labour, skill, or effort’.[5] The former requirement, according to precedents of the Court of Justice of the European Union (CJEU),[6] is actually requiring the authors to stamp the work with their ‘personal touch’,[7] which apparently is a much higher requirement than the ‘Sweat of the Brow’ principle. The United States (US) currently holds an originality position somewhere in between these two extremes, requiring that the work should be considered original plus at least some minimal degree of creativity.[8]
In terms of China, according to Article 2 of the Regulations for the Implementation of the Copyright Law (RICL), the term ‘works’ as referred to in the CLC means intellectual creations with originality in the literary, artistic or scientific domain, insofar as they can be reproduced in a tangible form.[9] It is clear from this article that ‘originality’ is required for a creation to be considered as a ‘work’ in China. The historical reason why the CLC failed to mention the ‘originality’ requirement is to avoid confusion with the patent law concepts such as ‘novelty’ or ‘inventive step’,[10] but after the concept of ‘originality’ was introduced with the promulgation of the RICL in 2002, the inconsistent interpretation of the ‘originality’ requirement by Chinese courts became the new problem.[11] Article 15 of the 2002 Interpretation of the Supreme People’s Court (SPC) Concerning the Application of Laws in the Trial of Civil Disputes over Copyright provides that, if a work was created by different authors on the basis of the same topic, the authors shall enjoy independent copyright if the expression of the work is completed independently and is creative.[12] This is the first clear indication that China is following a ‘Sweat of the Brow’ plus standard. Nevertheless, apart from Article 15 of the aforesaid SPC Interpretation, there is no further explanation with regard to the meaning of ‘creative’ provided anywhere in the related laws and regulations. Yet considering the fact that China is a civil law jurisdiction and that the CLC does distinguish copyright and neighbouring rights, it is believed that China’s originality standard is more akin to but not as demanding as those of other civil law jurisdictions such as Germany.[13] But in judicial practice, the requirement of creativity could be very low,[14] making it resemble the US standard.
Devoid of its own originality theory, the most controversial part of the debate surrounding the AI-generated contents in China is then the standard in determining originality. Some scholars believe that when assessing the originality of a work, the court should only focus on the end result, not the creative process. Therefore, they uphold an objective standard in assessing the originality of AI-generated contents. If the lowest degree of creativity can be found on the work, it is then considered original.[15] Some other scholars uphold a subjective standard which believes that the creative process is decisive when assessing originality rather than the end product.[16] Behind the debate is the fight for the ultimate say over the work between the author and the reader: is it the actual contribution of the author in creating the work that matters, or the imaginary contribution of the author surmised by the reader?
Opinions differ about the question. In a 2006 case, the Henan Higher People’s Court ruled that the silhouette of a stone tablet is not original, as ‘there is not enough evidence to prove that the tablet maker has conceived the silhouette and used his skill to shape the tablet in a certain way.’[17] In the Film v. Baidu case — the first AI and copyright case in China that concerns a data report generated by AI program — being ruled on 25 April 2019, the Beijing Internet Court also investigated the AI generative process of the report when determining originality.[18] Prof. Buccafusco commented that ‘an author is a human being who intends to produce one or more mental effects in an audience by an external manifestation of behaviour. A writing is any medium through which the mental effects are to be conveyed’,[19] which is also clearly in support of the subjective standard. However, in the 1951 Alfred Bell v. Catalda Fine Arts case, the US Court of Appeals for the Second Circuit held that ‘[A] copyist’s bad eyesight or defective musculature, or a shock caused by a clap of thunder, may yield sufficiently distinguishable variations. Having hit upon such a variation unintentionally, the ‘author’ may adopt it as his and copyright it.’[20] It supports the idea that the author’s intention is irrelevant in determining originality. The next question is: which method is better, the subjective or the objective standard?
This article is in support of the subjective standard. In the Alfred Bell v. Catalda Fine Arts case, the Second Circuit has drawn an analogy between copyright and patent.[21] But this analogy, too, is flawed, as patent rights are bound by a different set of rules and focusing on the ‘inventiveness’ of the discovery. Even if a scientific discovery was the result of an accident, granting patent rights to the first discoverer – from a utilitarian perspective – is beneficial to society. But if we are to grant copyright to unintentional variations, the justification is just not strong enough. Specifically, if we merely base on the end product to determine originality, then we may not be able to distinguish between the work of nature and the work of humans, if no evidence concerning the creative process can be garnered. Without interrogating the creative process, the consequences could include the ridiculous scenario where a stone with a unique shape done by natural erosion, if claimed to be carved by a human being, will be copyright protected. Similarly, AI-generated content will be even more thorny, as the quality is good enough to mislead the public about its origin, if the creative process is not taken into account. Considering the fact that machines could create ‘works’ in a much faster pace than humans in the future,[22] it is crucial to trace back to the creative process of AI-generated works and compartmentalize them by the originality test.
The prerequisite of the following discussion is the fact that we are still in a ‘weak AI’ era that ‘computing machines carrying out apparently sophisticated tasks but within a narrowly-defined role that they are programmed to navigate.’[23] In other words, the machine is not ‘strong’ enough to create contents on its own and to go beyond the programmed tasks yet. Wang Qian characterized AI-generated contents as ‘the result of the application of algorithms, rules and templates’,[24] and the learning process of AI as ‘the course of identifying rules.’[25] Wang thus concluded that, compared with human, the intelligent machine followed a fundamentally different rationale in generating content.[26] If applying the same algorithm set on different terminals on the same object, will render the same result, then no matter how ‘creative’ the AI-generated content appears, it is unoriginal, as the result is definite.[27] Even in the case of contents generated by deep learning AI techniques,[28] they are still derived from a process that aims to apply a specific algorithm on a heap of data to secure the ‘best results’.[29] This is a logical corollary: if different applications using the same algorithm were employed to analyse the same set of data and the results are identical, then there is no room for copyright to step in due to the merger doctrine, as the ways to express the idea are extremely limited.[30] Nevertheless, we should not jump to a conclusion that AI-generated contents are not copyrightable in such an early stage, as the degree of creativity required by the originality test of the CLC is comparatively low, and that in different levels of abstraction there could be more than enough possibilities that could satisfy the copyrightability requirements. Hence the results – albeit not as perfect as the expected end result – can be more than enough. For example, in the Generative Adversarial Networks (GANs) model,[31] which can be trained to produce images by text to image synthesis, there can be multiple results generated that match the description.[32] Moreover, even if we assume that the machine will and can only produce ‘the’ end result, we still cannot preclude the possibility that machines and algorithms could evolve and generate quality contents independently in the future. Yet the most formidable hurdle for human beings is to recognize that machines and humans can be equal subjects in both practical (that machines can express feelings and ideas)[33] and legal terms (that machines can be made a subject that can enjoy ownership).
Obviously, we can draw the conclusion that AI shall be treated as a tool that can help humans to create works more efficiently, but not as an equal subject. Considering the fact that we are still in a ‘weak AI’ era, it is suggested that, in terms of originality, the subjective standard shall be followed so that we could distinguish between the AI-generated works and the human-authored works. Putting machines on the author seat will be against several fundamental justifications of IP, and it is also a fact that machines cannot act in the human way that is expected by the law.
(This piece is an excerpt from He, T. (2020), “The Sentimental Fools and The Fictitious Authors: Rethinking the Copyright Issues of AI-generated Contents in China,” 27 Asia Pacific Law Review issue 2 (forthcoming).
[1] See ‘Microsoft AI creates ‘new’ Rembrandt painting’ Netimperative (2016) <http://www.netimperative.com/2016/04/microsoft-ai-creates-new-rembrandt-painting/> accessed 27 April 2019.
[2] Qian Wang, ‘On the Legal Determination of AI-generated Contents in the Copyright Law’ (2017) 5 Science of Law 148, 150 (王遷, ‘論人工智能生成的內容在著作權法中的定性’, 《法律科學》, 2017年第5期, 第150頁).
[3] Lionel Bentley et al., Intellectual Property Law (5th edn Oxford University Press, 2018) 96-98.
[4] Infopaq International A/S v Danske Dagblades Forening, Case C-5/08 [2009] ECR I-6569 (ECJ).
[5] Ibid, 37.
[6] Football Dataco v. Yahoo! UK, Case C-604/10 [2012] 2 CMLR (24) 703 (ECJ); Eva-Maria Painer v Standard Verlags GmbH and Others, Case C-145/10 [2012] ECDR (6) 89 (ECJ).
[7] Ibid, 87-93; Football Dataco v. Yahoo! UK, ibid, 38.
[8] Feist Publications v. Rural Telephone Service, 499 U.S. 340 at 341-344 (1991).
[9] The Regulations for the Implementation of the Copyright Law, 2013, Article 2.
[10] Chengsi Zheng and Michael Pendleton, Copyright Law in China (CCH International, 1991) 108.
[11] For example, with regard to a case ruled in 2018 related to the originality of the relay of sports games, the court of first instance held that the ‘selection and arrangement of the camera in recording the game and forming a viewable new screen’ amount to an original work, whereas the second instance held the opposite opinion. See Tian Ying Jiu Zhou v. Sina.com and LETV, first instance [2014] Beijing Chaoyang District People’s Court, CMCZ. No. 40334; second instance [2015] Beijing Intellectual Property Court, JZMZZ. No. 1818.
[12] SPC Interpretation Concerning the Application of Laws in the Trial of Civil Disputes over Copyright, 2002, Article 15.
[13] Jinchuan Chen, ‘The Two Non-negligible Factors in Determining Originality Standards’ (2018) 6 China Copyright 27, 28 (陳錦川, ‘確定“獨創性”標準不能忽視的兩個因素’, 《中國版權》, 2018年第6期, 第28頁).
[14] Unlike German Copyright Law that distinguishes between ‘photo’ and ‘photography works’ and provides that the former is covered by the neighboring rights, the CLC only provides ‘photography works’ under its Article 3, which provides a list of copyright works. This kind of arrangement will in fact render random photo shoots copyright works and will dramatically water down the originality standard. See Yide Ma, ‘The Copyright Determination of Reproduced Photographic works’ (2016) 4 Chinese Journal of Law 137, 138-40 (馬一德, ‘再現型攝影作品之著作權認定’, 《法學研究》, 2016年第4期, 第138-140頁).
[15] See Jiming Yi, ‘Can AI-generated Contents Be Considered Works?’ (2017) 5 Science of Law 137, 139 (易繼明, ‘人工智能創作物是作品嗎？’, 《法律科學》, 2017年第5期, 第139頁); See also Handong Wu, ‘The System Design and Legal Regulation in the AI Era’ (2017) 5 Science of Law 128, 131 (吳漢東, ‘人工智能時代的制度安排與法律規制’, 《法律科學》, 2017年第5期, 第131頁); Chen Li, ‘Xie Wan Qiao and Originality’ (2005) 8 Electronic Intellectual Property 58, 58 (李琛, ‘謝綰樵與獨創性’, 《電子知識產權》, 2005年第8期，第58頁); Shan Sun, ‘The Predicament and Future Path of Copyright Protection over AI-generated Contents’ (2018) 11 Intellectual Property 60, 63 (孫山, ‘人工智能生成內容著作權法保護的困境與出路’, 《知識產權》, 2018年第11期，第63頁)
[16] Wang suggested that we should focus on the ‘contents that not only similar in form to the human-created works, but its creative process also satisfies the originality test.’ See Wang (n 22) 150; See also Bo Yuan, ‘On the Falsification of AI Copyright in Literature’ (2018) 6 Electronic Intellectual Property 20, 26 (袁博, ‘論文學領域人工智能著作權之證偽’, 《電子知識產權》, 2018年第6期，第26頁).
[17] Yan Yong v. The Administrative Department of Cultural Heritage and Tourism of Yong Cheng City [2006] Henan Higher People’s Court, YFMSZZ. No. 7.
[18] Beijing Film Law Firm v. Baidu [2018] Beijing Internet Court, J0491MC. No. 239.
[19] Christopher Buccafusco, ‘A Theory of Copyright Authorship’ [2016] 102 Virginia Law Review 1229, 1260.
[20] Alfred Bell & Co. Ltd. v. Catalda Fine Arts, Inc. et al, 191 F.2d 99 (2d Cir. 1951).
[21] ‘Many great scientific discoveries have resulted from accidents.’ Ibid, fn 25.
[22] Kalin Hristov, ‘Artificial Intelligence and the Copyright Dilemma’ [2017] 57 IDEA 431, 434.
[23] Jeremy Cubert and Richard Bone, ‘The law of intellectual property created by artificial intelligence’, in Woodrow Barfield and Ugo Pagallo (eds), Research Handbook on the Law of Artificial Intelligence (Edward Elgar, 2018) 415-16.
[24] Wang (n 2) 150.
[25] Ibid, 151.
[26] Ibid, 152.
[27] Ibid, 152.
[28] Yann LeCun, Yoshua Bengio and Geoffrey Hinton, ‘Deep Learning’ (2015) 521 Nature 436, 436: ‘Deep-learning methods are representation-learning methods with multiple levels of representation, obtained by composing simple but non-linear modules that each transform the representation at one level (starting with the raw input) into a representation at a higher, slightly more abstract level.’
[29] Wang (n 2) 152.
[30] BUC Int’l Corp. v. International Yacht Council Ltd., 489 F.3d 1129, 1143 (11th Cir. 2007): ‘The merger doctrine operates as an exception to the normal idea-expression dichotomy. The doctrine holds that, when there are so few ways of expressing an idea, not even the expression is protected by copyright.’
[31] See Ian Goodfellow et al., ‘Generative Adversarial Nets’ (2014) Advances in neural information processing systems 2672: ‘In the proposed adversarial nets framework, the generative model is pitted against an adversary: a discriminative model that learns to determine whether a sample is from the model distribution or the data distribution.’
[32] Scott Reed et al., ‘Generative Adversarial Text to Image Synthesis’ (2016) Generative adversarial text to image synthesis. arXiv preprint arXiv:1605.05396.
[33] For example, the Japanese government in a report indicated that, ‘AI-generated content is not considered to be a work because it is not a “creative expression of thought or emotion (Article 2 Section 1 of the Japanese Copyright Act)”, and it is considered that a copyright does not occur either.’ See Intellectual Property Strategy Headquarters, ‘Intellectual Property Strategic Program 2016’ Prime Minister’s Office of Japan (2016) <http://www.kantei.go.jp/jp/singi/titeki2/kettei/chizaikeikaku20160509.pdf> accessed 27 April 2019.