Copyright Fair Use in the Face of Technological Developments: Staying Ahead or Limping Behind? – Part 2

Specific Technological Developments and Fair Use – Two Case Studies

At the time of writing, there are over a dozen lawsuits since the start of 2023 against companies such as OpenAI and Stability AI for using copyright-protected works in the training of various AI systems,¹For example, Raw Story Media, Inc v OpenAI Inc, Case 1:24-cv-01514 (SDNY, 22 August 2024); UMG Recordings Inc, et al v Uncharted Labs Inc, Case 1:24-cv-04777 (SDNY, 24 June 2024); The Intercept Media, Inc v Open AI, Inc, Case 1:24-cv-01515 (SDNY, 6 June 2024); The New York Times Company v Microsoft Corporation, Case 1:23-cv-11195 (SDNY, 30 May 2024); Zhang, et al v Google LLC, et al, Case 3:24-cv-02531 (ND Cal, 24 April 2024); Nazemian v Nvidia Corp, Case 3:24-cv-01454 (ND Cal, 8 March 2024); Tremblay v Open AI, Case 3:23-cv-03223 (ND Cal, 16 February 2024); Andersen v Stability AI Ltd, Case 23-cv-00201-WHO (ND Cal, 30 October 2023); Getty Images (US), Inc v Stability AI, Inc, Case 1:23-cv-00135 (D Del, 3 February 2023). but there has been no judicial ruling on the merits as yet. Part 1 has provided an overview of fair use case law. This Part will focus on the analysis of whether two particular technological uses qualify as fair use – the common thread between these two uses is that both rely on TDM (text and datamining) and machine learning. Generally, TDM is an umbrella term referring to “computational processes for applying structure to unstructured electronic texts and employing statistical methods to discover new information and reveal patterns in the processed data.”²Brief of Digital Humanities and Law Scholars as Amici Curiae in Support of Defendant-Appellees at 5, Authors Guild v Google, Inc, 804 F.3d 202 (2nd Cir. 2015) (No. 13-4829). It can refer to any process using computers that creates metadata derived from something that was not initially conceived of as data; it can be used to produce statistics and facts about copyrightable works, and to render copyrighted text, sounds, and images into uncopyrightable abstractions.³Matthew Sag, “Copyright Saftety for Generative AI” (2023) 61 Houston Law Review 295 at 305-306.

1. Search Engines

Over the past two decades, search engines have emerged as a significant technology that may qualify as a transformative fair use, making images and information that would otherwise be protected by copyright searchable on the Internet. By 2020, Google has amassed a global market share of 87%, and “has become ubiquitous, often used as a transitive verb”.⁴Chuck Price, “17 Great Search Engines You Can Use Instead of Google”, Search Engine Journal (5 April 2020). Back then, other search engines, such as Bing, Baidu, Wiki, Twitter, DuckDuckGo and Startpage, while gaining popularity, still lag far behind. However, by 2024, with the rapid development of AI, the possibility of combining AI with search engines is fast becoming a reality, such as GPTGO.ai and a new Gemini model customised for Google Search. On 31 October 2024, OpenAI launched a search feature within ChatGPT, which offers numerous up-to-the-minute information such as sports scores, stock quotes, news, and the weather, powered by real-time web search and partnerships with news and data providers.⁵Hayden Field, “OpenAI launches ChatGPT search, competing with Google and Microsoft”, CNBC (31 October 2024).

Based on prevailing case law, in assessing fair use in the context of internet search engines, courts have relied heavily on the first fair use factor. In Kelly v Arriba Soft Corp,⁶336 F.3d 811 (9th Cir. 2003). it was held in 2003 that the now-defunct search engine Arriba’s creation and use of thumbnail versions of a professional photographer’s copyrighted images was fair use because the “smaller, lower-resolution images … served an entirely different function than [the] original images”.⁷Kelly v Arriba Soft Corp, 336 F.3d 811 at 818 (9th Cir. 2003). The Ninth Circuit was of the view that the original images served an artistic or aesthetic purpose but the thumbnail images, which were provided in response to a user’s search query, were incorporated into the search engine’s overall function “to help index and improve access to images on the internet and their related web sites”.⁸Ibid at 818–819. Arriba’s use was found to be transformative and the use of the thumbnail images also “benefit[ed] the public by enhancing information-gathering techniques on the internet”.⁹Ibid at 820. In the Google Images litigation, the Ninth Circuit in 2007 held that Google’s use of thumbnail images in its search engine is “highly transformative” and that “a search engine provides social benefit by incorporating an original work into a new work, namely, an electronic reference tool”, and thus Google’s use of the thumbnail images was fair use.¹⁰Perfect 10, Inc v Amazon.com, Inc, 508 F.3d 1146 at 1163–1165 (9th Cir. 2007).

In 2015, the Second Circuit similarly found that fair use protected the Google Books search engine, which employs digital, machine-readable copies of millions of copyright-protected books scanned by Google.¹¹Authors Guild v Google, Inc, 804 F.3d 202 at 220–221 and 223 (2nd Cir. 2015) (“Google Books”). The Google Books search engine enables searching for a specific term, and then provides “snippets,” or a part of a page, for users to read. The Court held that both functions involve a “highly transformative purpose of identifying books of interest to the searcher”.¹²Ibid at 218. Again the consideration of public benefit was foremost in the minds of the judges as the Court unanimously held that the search function “augments public knowledge by making available information about Plaintiffs’ books without providing the public with a substantial substitute” [emphasis in original].¹³Ibid at 207. Finally, the Court concluded that Google’s commercial motivation did not significantly outweigh these transformative uses. The search engine makes possible a new type of research known as “text mining” or “data mining”, whereby users can search across the corpus of books to determine the frequency of specified terms across time. In 2019, the Ninth Circuit distinguished these search engine cases where fair use was found, from the closed-universe search engine Digs – a service provided by Zillow, an online real estate marketplace – that does not “crawl” the web, deciding that it was not fair use there.¹⁴VHT, Inc v Zillow Group, Inc, 918 F.3d 723 at 739–740 (9th Cir. 2019). The court cautioned that:¹⁵Ibid at 742.

What we divine from these cases is that the label ‘search engine’ is not a talismanic term that serves as an on-off switch as to fair use. Rather, these cases teach the importance of considering the details and function of a website’s operation in making a fair use determination.

In summary, not all uses of works in search engines are considered fair use. Courts will still have to analyse each fair use factor on a case-by-case basis. Cases such as HathiTrust and Google Books were based on a different technological paradigm of non-expressive fair use. Generative AI systems like ChatGPT and Midjourney today, however, “produce much more than information about expression; they are now the engines of new content creation.”¹⁶Matthew Sag, “Copyright Saftety for Generative AI” (2023) 61 Houston Law Review 295 at 308. For new search engines that combine generative AI with search features, such as the new ChatGPT search, the fair use evaluation becomes more complicated. The next section attempts to understand some of the implications.

2. Generative AI

(a) Computational Data Analysis Exception

It should be noted that a TDM exception known as the computational data analysis (CDA) exception has been enacted in the Copyright Act 2021. Under the CDA exception, five stringent conditions must be fulfilled, and they include “lawful access” to the copyrighted content and that the making of a copy cannot be used for any other purpose except for identifying, extracting or analysing information/data and using that to improve the functioning of a program in relation to that type of information/data.¹⁷For an analysis of the computational data analysis exception, see David Tan & Thomas Lee Chee Seng, “Copying Right in Copyright Law: Fair Use, Computational Data Analysis and the Personal Data Protection Act” (2021) 33 SAcLJ 1032. In implementing Proposal 8 of the Copyright Review Report,¹⁸Ministry of Law & Intellectual Property Office of Singapore, Singapore Copyright Review Report (17 January 2019). section 243 of the Copyright Act 2021 introduces a specific exception for reproduction of works made for the purpose of “computational data analysis” provided that a number of conditions in section 244 are met. “Computational data analysis” is defined non-exhaustively as “using a computer program to identify, extract and analyse information or data from the work” – which is synonymous with TDM. The UK has already in place a TDM exception – albeit narrower than the Singapore version – that “a person who has lawful access to the work may carry out a computational analysis of anything recorded in the work for the sole purpose of research for a non-commercial purpose”.¹⁹Copyright, Designs and Patents Act 1988 (c 48) (UK) s 29A. For an analysis of other TDM exceptions, see Rossana Ducato and Alain M Strowel, “Ensuring Text and Data Mining: Remaining Issues with the EU Copyright Exceptions and Possible Ways Out” (2021) 43 European Intellectual Property Review 322; Martin Senftleben, “Compliance of National TDM Rules with International Copyright Law: An Overrated Nonissue?” (2022) 53 International Review Intellectual Property & Competition Law 1477; Thomas Margoni and Martin Kretschmer, “A Deeper Look into the EU Text and Data Mining Exceptions: Harmonisation, Data Ownership, and the Future of Technology” (2022) 71 GRUR International 685.

It is generally difficult to prove wholesale copying of millions of works as the various GAIAs do not disclose the training datasets, and one would have to proceed on a classic substantial similarity analysis in respect of each output text/image vis-à-vis the original work.²⁰David Tan, “Generative AI and Copyright – Part 1: Copyright Infringement” (2023) SAL Prac 24. Two salient issues are relevant: (a) whether the use of copyright-protected works for machine learning (“input”); and (b) the works created from natural language commands (“output”) are infringing copyright. To determine if there is liability, one needs to understand how the CDA exception and fair use provision in the Copyright Act 2021 may be relevant to the input and output scenarios. In Singapore, copyright law can provide a defence for such infringing uses if these uses fall under either the CDA exception²¹Copyright Act 2021 (2020 Rev Ed) ss 243–244. or the fair use provision.²²Copyright Act 2021 (2020 Rev Ed) ss 190–191. It should be noted that these permitted uses are independent of one another, and more than one exception may apply if the relevant conditions are met.²³Copyright Act 2021 (2020 Rev Ed) ss 184. Lastly, any contract term is void to the extent that it purports, directly or indirectly, to exclude or restrict the CDA exception.²⁴Copyright Act 2021 (2020 Rev Ed) ss 187(1)(c).

For the CDA exception, the five conditions to be satisfied include the user proving that the copy is made for the purpose of CDA and not for any other purpose; the user not supplying the copy to any person other than for the purpose of verifying the results of the CDA carried out by the user; the user having lawful access to the material (the first copy) from which the copy is made; and that the first copy not being an infringing copy. All five conditions must be satisfied. The Singapore legislation gives an example that “X does not have lawful access to the first copy if X accessed the first copy by circumventing paywalls” and that the use of images to train a computer program to recognise images, such as facial recognition software, as a permissible purpose. Furthermore, it is stated in the Act that “X does not have lawful access to the first copy if X accessed the first copy in breach of the terms of use of a database”.²⁵Copyright Act 2021 (2020 Rev Ed) s 244(2)(d).

For machine learning purposes, the scraping of the Internet for text and images will often circumvent paywalls or violate the terms of use, hence failing the “lawful access” requirement under s 244(2)(d) of the Copyright Act 2021. Furthermore, the making of a copy, which will involve the conversion of authorial works into a machine-readable format or, in some GAIAs, data storage, will not be for the sole purpose of analysing the data to improve the functioning of the AI in relation to that data;²⁶Copyright Act 2021 (2020 Rev Ed) s 244(2)(b). it will be for the purpose of generating new works based on that data, which is an impermissible purpose.

(b) Fair Use

There is a wealth of scholarship on training data and fair use.²⁷See, eg, Andrew W Torrance and Bill Tomlinson, “Training is Everything: Artificial Intelligence, Copyright and ‘Fair Training’” (2023) 128 Dickinson Law Review 233; Matthew Sag, “Copyright Saftety for Generative AI” (2023) 61 Houston Law Review 295; Mark A Lemley and Bran Casey, “Fair Learning” (2021) 99 Texas Law Review 743; Benjamin L W Sobel, “Artificial Intelligence’s Fair Use Crisis” (2017) 41 Columbia Journal of Law & the Arts 45. It was pointed out by Matthew Sag that US courts “have agreed that copying without permission is fair use, and thus noninfringing, in the context of software reverse engineering, plagiarism detection software, and the digitization of millions of library books to enable meta-analysis and indexing.”²⁸Sag, ibid at 304. Sag notes that:

The fair use status of nonexpressive use is not a special exception for the technology sector. Rather, the rationale for allowing for-profit and academic researchers to derive valuable data from other people’s copyrighted works is a necessary implication of the fundamental distinction between protectable original expression and unprotectable facts, ideas, abstractions, and functional elements.²⁹Ibid at 304-5.

In considering fair use as applied to generative AI systems, two fair use factors that are likely to carry the greatest weight in the analysis are: (a) what is the purpose/character of the use, namely whether the use by generative AI is “transformative”, that is, whether it changes the purpose or the nature of the original work in some way; and (b) what is the impact of the generative AI’s use on the market, that is, whether it threatens the livelihood of the original creator by competing with their works or the licensing market for their works. While I conclude that the use of copyrighted works as training data is generally not fair use, there are other scholars who argue that such a use amounts to “fair training” and is therefore fair use.³⁰Eg Andrew W Torrance and Bill Tomlinson, “Training is Everything: Artificial Intelligence, Copyright and ‘Fair Training’” (2023) 128 Dickinson Law Review 233, 245, 253-254.

In respect of the first factor, Authors Guild v HathiTrust³¹Authors Guild v HathiTrust 755 F.3d 87 at 92 (2nd Cir. 2014). is instructive – the issue was whether the digitisation of copyrighted works by 13 universities and other organisations in creating the HathiTrust Digital Library (HDL) without authorisation may constitute fair use. The US Second Circuit Court of Appeals found that the first factor weighed in favour of fair use as HDL’s enabling of full-text search “serves a new and different function from the original” and is socially beneficial.³²Ibid at 97. See also William F Patry, Patry on Copyright, vol. 4 (West, Online, 2015) at §10:21 (observing that the use in Authors Guild v HathiTrust is “socially beneficial, serves a different purpose than the original, and is in no way substitutional”). Additionally, the dealing was found to carry a “non-profit educational” purpose as the HDL was a project started by educational and non-profit institutions targeted at providing greater access to works without any “purely commercial” motive.³³Authors Guild v HathiTrust 755 F.3d 87 at 90–91 (2nd Cir. 2014).

The Ninth Circuit’s decision in Kelly v Arriba Soft Corp³⁴336 F.3d 811 (9th Cir. 2003). is also useful in understanding how the evaluation of the third factor could be applied to generative AI uses. There, it was held that the use of entire copyrighted works was necessary in situations involving search engines since copying only a part of the copyrighted work would create practical difficulties for users, thereby reducing the usefulness of the search engine. In the same vein, even if entire works were copied by web robots in the TDM context, it could be reasoned that such a taking is reasonable, considering the different purpose of the dealing (that is, to identify patterns in vast amounts of raw data); thus, the third factor might not necessarily weigh against fair use. But if the purpose of generative AI is to analyse the specific expression of particular artists, and then replicate portions of that expression in response to a text prompt, then it does not appear to be a different purpose.

In the latest US Supreme Court’s decision on fair use, the majority observed that “whether the purpose and character of a use weighs in favour of fair use is, instead, an objective inquiry into what use was made, that is, what the user does with the original work.”³⁵Andy Warhol Foundation for the Visual Arts, Inc v Goldsmith, 598 US 508 at 545 (2023). In that case, the use was Andy Warhol Foundation’s commercial licensing of Warhol’s Orange Prince (which was based on Lynn Goldsmith’s original photograph) to appear on the cover of Condé Nast’s special commemorative edition. The purpose of that use was to illustrate a magazine about Prince with a portrait of Prince, and an infringing work that portrays Prince somewhat differently from Goldsmith’s photograph (yet has no critical bearing on her photograph) was insufficient for the first factor to favour Andy Warhol Foundation, given the specific context of the use. The majority emphasised:³⁶Ibid at 546.

To hold otherwise would potentially authorize a range of commercial copying of photographs, to be used for purposes that are substantially the same as those of the originals. As long as the user somehow portrays the subject of the photograph differently, he could make modest alterations to the original, sell it to an outlet to accompany a story about the subject, and claim transformative use.

These observations are especially pertinent for images produced by GAIAs such as DALL·E, Stable Diffusion or Midjourney. If a user was looking for an image for illustrative purposes for a magazine, book, annual report or marketing brochure, and provides specific text prompts to a generative AI system to produce such an image – as opposed to licensing one directly from the original author – then the first factor is unlikely to weigh in favour of fair use. The most recent US Circuit Court of Appeals decision on fair use clearly illustrates this point. Deciding that the first factor weighed in favour of fair use despite the defendant’s reproduction of the artist’s Dog Art painting series online as part of educational art kits so that students could learn at home during the pandemic, the unanimous opinion states: “As noted, the art kits had educational objectives, while the original works had aesthetic or decorative objectives. The purpose of Mix Creative’s specific use thus was not “substantially the same” as that of the original works, and there was little threat that the art kits would serve as substitutes for the originals.”³⁷Keck v Mix Creative Learning Center LLC, 116 F.4th 448 at 455 (5th Cir. 2024). This stands in contrast to the facts in AWF v Goldsmith, where the purpose of the Foundation’s “specific use” was, according to the Supreme Court, “substantially the same” as the purpose of the original photograph. Both the original and the adaptation that the Foundation licensed to Condé Nast were “portraits of Prince used in magazines to illustrate stories about Prince.”³⁸Ibid.

The application of the fourth factor is also highly dependent on the finding of the first factor. The US Supreme Court in Campbell v Acuff-Rose Music Inc had emphasised the close linkage between the first and fourth factors, in that the more the copying is done to achieve a purpose that differs from the purpose of the original, the less likely it is that the copy will serve as a satisfactory substitute for the original.³⁹Campbell v Acuff-Rose Music, Inc, 510 US 569, 591 (1994). The Second Circuit noted that even if the purpose of the copying was for a valuably transformative purpose, such copying might nonetheless harm the value of the copyrighted original if done in a manner that resulted in widespread revelation of sufficiently significant portions of the original as to make available a significantly competing substitute.⁴⁰Authors Guild v Google, 804 F.3d 202, 223 (2nd Cir. 2015). Generally, copyright-protected works copied for data mining purposes will require extensive processing and analysis before knowledge is derived and shared. Miners must ensure that they do not reveal significant portions of the original copyrighted works to the public. Although one could argue that data mining could limit the rights owners’ expansion into a potential market (e.g. a lost opportunity to license the works⁴¹Authors Guild, Inc v HathiTrust, 755 F.3d 87, 99 (2nd Cir. 2014) (this was an argument the plaintiffs raised).) since markets are dynamic and change over time to meet new demands, the US Circuit Courts have universally dismissed this argument where only a small portion of the original works was revealed to the public. In the Google Books litigation, the Second Circuit held that “a mere revelation of 16% of the text of plaintiffs’ books overstates the degree to which snippet view can provide a meaningful substitute.”⁴²Authors Guild v Google, 804 F.3d 202, 223 (2nd Cir. 2015). In generative AI scenarios where a significant portion of an original work is reproduced in an output in response to a user’s text prompt, then one may more confidently discern a substitutive impact.

I have previously written:

Programme developers, and the organisations that invest significant resources in generative AI, will have the benefit of access to two generous but distinct provisions in the Singapore Copyright Act 2021 as a defence to their unauthorised uses of copyright-protected works. However, they must be mindful that the judicial interpretation of the fair use provision in Singapore likely remains guided by paradigmatic cases where an open-universe search engine making of a digital copy for the purpose of enabling a search for identification of books containing a term of interest involves a highly transformative purpose, and where copying from an original for the purpose of criticism, commentary, caricature, parody, or pastiche would be transformative.⁴³David Tan, “Generative AI and Copyright – Part 1: Copyright Infringement” (2023) SAL Prac 25 at para 23.

ChatGPT, Stable Diffusion, Midjourney and many other comparable GAIAs, including the new Gemini 1.5 announced by Google at the time of writing, are not search engines. A number of them are highly successful commercial enterprises, with Stability AI valued at US$1bn, and some charging a user fee for their services. It is likely that there is little transformative purpose to be found as the AI would be accessing and reproducing the creative expression in these works in the outputs, that is the works would have been appropriated for their creative elements rather than their underlying facts. Similarly, Sag concludes: “If [large language models] just took expressive works and conveyed that same expression to a new audience with no additional commentary or criticism, or no distinct informational purpose, that would be a very poor candidate for fair use.”⁴⁴Matthew Sag, “Copyright Saftety for Generative AI” (2023) 61 Houston Law Review 295, 312-313. According to Professor Rebecca Tushnet at Harvard Law School, an overhaul of copyright law is not needed as “current copyright law is clear when it comes to those employing the technology for creative use.”⁴⁵Colleen Walsh, “How to think about AI: Delving into the legal and ethical challenges of a game-changing technology”, Harvard Law Bulletin, Summer 2023, 21 at 22. The fair use provision ain’t broke, and it is poised to be able to handle the challenges of new technology. Bring it on!

This article is adapted from a presentation by Professor Tan titled “AI Learning and Copyright Law: Developments in Singapore” at the United States-Asia Comparative Copyright Roundtable in December 2024 at Waseda University in Tokyo.

The post Copyright Fair Use in the Face of Technological Developments: Staying Ahead or Limping Behind? – Part 2 appeared first on The Singapore Law Gazette.