Fair Learning
Introduction
The challenge handed to the musician was peculiar and daunting: Take a five-second sample of a randomly selected song and, with just a moment’s notice, transform it into a full-length piece composed in the style of a completely different artist.11See infra notes 2–5 and accompanying text. Close On this occasion, the musician rose to the challenge with such aplomb that it took the internet by storm, earning praise and recognition from media outlets across the globe.22See, e.g., Will Knight, This AI-generated Musak Shows Us the Limit of Artificial Creativity, MIT Tech. Rev. (Apr. 26, 2019), https://www.technologyreview.com/s/613430/this-ai-generated-musak-shows-us-the-limit-of-artificial-creativity/ [https://perma.cc/D7FS-42VA] (describing MuseNet’s ability to deftly accomplish these unique musical pairings); Devin Coldewey, MuseNet Generates Original Songs in Seconds, from Bollywood to Bach (or Both), TechCrunch (Apr. 25, 2019, 3:31 PM), https://techcrunch.com/2019/04/25/musenet-generates-original-songs-in-seconds-from-bollywood-to-bach-or-both [https://perma.cc/2MVD-2GQA] (same); Jon Porter, OpenAI’s MuseNet Generates AI Music at the Push of a Button, Verge (Apr. 26, 2019, 9:11 AM), https://www.theverge.com/2019/4/26/18517803/openai-musenet-artificial-intelligence-ai-music-generation-lady-gaga-harry-potter-mozart [https://perma.cc/8JBD-2NL6] (same). Close A Mozart concerto played in the style of Katy Perry?33See Porter, supra note 2 (describing this ability). Close No problem. Lady Gaga’s Poker Face in the style of a recent Bollywood hit? Why not?44See id. (describing this ability). Close Seemingly no pairing of styles, no matter how clashing, proved too difficult.55The artist couldn’t take on literally every pairing, of course, but was nonetheless capable of a stunning diversity of combinations. See Christine Payne, MuseNet, Open AI Blog (Apr. 25, 2019), https://openai.com/blog/musenet/ [https://perma.cc/LG4V-8DS5] (describing some of the limits of musical diversity inherent to MuseNet). Close
The artist capable of such a tour de force? MuseNet—the type of unique, futuristic-sounding name common of many contemporary artists.66See Porter, supra note 2 (describing MuseNet). Close But, on closer inspection, MuseNet’s name wasn’t a reference to the internet (or even to the Terminator series’ dystopian Skynet). Rather, MuseNet was a neural
network. The musician, in other words, was a robot,77Here, and for the remainder of this piece, we use the term “robot” loosely. See Bryan Casey & Mark A. Lemley, You Might Be a Robot, 105 Cornell L. Rev. 287, 295–96 (2020) (explaining the difficulties of explicitly defining the term robot and opting to include various forms of artificial intelligence (AI) in the definition). Close not a person.88A number of scholars have begun to address the copyrightability of creative works made by machines, following the lead of Pam Samuelson, who did it thirty-four (!) years ago. Pamela Samuelson, Allocating Ownership Rights in Computer-Generated Works, 47 U. Pitt. L. Rev. 1185 (1986); e.g., Clark D. Asay, Independent Creation in a World of AI, 14 Fla. Int’l. U. L. Rev. 201 (2020); Bruce E. Boyden, Emergent Works, 39 Colum. J.L. & Arts 377, 378 (2015); Annemarie Bridy, Coding Creativity: Copyright and the Artificially Intelligent Author, 2012 Stan. Tech. L. Rev. 5, 2; Katherine B. Forrest, Copyright Law and Artificial Intelligence: Emerging Issues, 65 J. Copyright Soc’y USA 355 (2018); Jane C. Ginsburg & Luke Ali Budiardjo, Authors and Machines, 34 Berkeley Tech. L.J. 343 (2019); James Grimmelmann, There’s No Such Thing as a Computer-Authored Work—And It’s a Good Thing, Too, 39 Colum. J.L. & Arts 403 (2016); William T. Ralston, Copyright in Computer-Composed Music: HAL Meets Handel, 52 J. Copyright Soc’y USA 281 (2005); Shlomit Yanivsky-Ravid, Generating Rembrandt: Artificial Intelligence, Copyright, and Accountability in the 3A Era—The Human-Like Authors Are Already Here—A New Model, 2017 Mich. St. L. Rev. 659, 664; Enrico Bonadio & Nicola Lucchi, How Far Can Copyright Be Stretched? Framing the Debate on Whether New and Different Forms of Creativity Can Be Protected, 2019 Intell. Prop. Q. 115. A Chinese court held in 2020 that an AI-written article is protected by copyright (and owned by the company that owns the AI). Paul Sawers, Chinese Court Rules AI-Written Article Is Protected by Copyright, VentureBeat (Jan. 10, 2020, 1:54 P.M.), https://venturebeat.com/2020/01/10/chinese-court-rules-ai-written-article-is-protected-by-copyright/ [https://perma.cc/4RQ9-QXEF]. Other scholars have focused not on how copyright law affects AI, but how AI can be used in implementing copyright regimes, particularly fair use. See, e.g., Dan L. Burk, Algorithmic Fair Use, 86 U. Chi. L. Rev. 283, 284–85 (2019) (questioning the ability of algorithms to incorporate fair use); Peter K. Yu, Can Algorithms Promote Fair Use?, 14 Fla. Int’l U. L. Rev. 329, 352 (2020) (considering whether and how algorithms can promote fair use). Our focus here is different than either thread. We are interested, not in how copyright might apply to the outcome of an AI, but in how it affects the process of training and using that AI. Close
Thanks to rapid advances in a subfield of computer science known as “machine learning” (ML),99Although the term was first coined in 1959, see Arthur L. Samuel, Some Studies in Machine Learning Using the Game of Checkers. I., in Computer Games I 335 (David N.L. Levy ed., 1988), the field did not see a significant inflection point in progress until around 2010. Close feats of robot ingenuity like the one displayed by MuseNet have become regular fixtures of the news. But feats of ML prowess aren’t limited to displays of creativity. There have also been similarly impressive advances in a host of other industry and social contexts, ranging from transportation,1010See, e.g., Winnie Hu, Driverless Cars Arrive in New York City, N.Y. Times (Aug. 6, 2019), https://www.nytimes.com/2019/08/06/nyregion/driverless-cars-new-york-city.html [https://perma.cc/6DGY-M3ZD] (describing recent advances in self-driving cars). Close to media curation,1111See, e.g., Mike Isaac, In New Facebook Effort, Humans Will Help Curate Your News Stories, N.Y. Times (Aug. 20, 2019), https://www.nytimes.com/2019/08/20/technology/facebook-news-humans.html [https://perma.cc/CG2J-BD7W] (describing how “Facebook has long relied on algorithms to select news stories for its users to see”). Close to medical diagnostics,1212See, e.g., Nicola Davis, AI Equal with Human Experts in Medical Diagnosis, Study Finds, Guardian (Sept. 24, 2019), https://www.theguardian.com/technology/2019/sep/24/ai-equal-with-human-experts-in-medical-diagnosis-study-finds [https://perma.cc/S7RR-TK7R] (explaining how AI can assist in medical diagnoses). Close to insurance risk mitigation.1313See, e.g., Jason Pontin, How AI-Driven Insurance Could Reduce Gun Violence, Wired (Feb. 27, 2018, 6:00 AM), https://www.wired.com/story/how-ai-driven-insurance-could-reduce-gun-violence/ [https://perma.cc/7WAT-XZAS] (proposing an AI system to assess gun insurance costs). Close
The vast potential of ML systems is matched only by their appetite for data. To perform, they must first learn how—generally, through a process of trial-and-error of epic proportions. And in order to create the right conditions for this learning process, engineers must begin by collecting and compiling enormous databases of exemplary tasks for machines to practice on, known as “training sets.”1414For a more thorough explanation of the term, see Training and Test Sets: Splitting Data, Google, https://developers.google.com/machine-learning/crash-course/training-and-test-sets/splitting-data [https://perma.cc/3EYU-7K3J]. For this Article’s purposes, the term “training set” encompasses both training sets (subsets of datasets used to train ML models) and test sets (subsets of datasets used to test a trained ML model). Close
Enter copyright law. Creating a training set of millions of examples almost always requires, first, copying many more millions of images, videos, audio, or text-based works. Those works are almost all copyrighted.1515There are occasional exceptions. A weather prediction AI, for instance, might train only on factual data provided by the National Weather Service, and that data is not subject to copyright. But as we discuss below, most training data sets are built on copyrighted works. Close Were a human to copy the sheer volume of songs that MuseNet did they’d be looking at serious consequences. (Just ask the founders of Napster;1616Napster Loses Net Music Copyright Case, Guardian (July 26, 2000), https://www
.theguardian.com/technology/2000/jul/27/copyright.news [https://perma.cc/8XXP-M9T9]. Close or Cox Cable, which was ordered to pay $1 billion because it didn’t terminate some of its users who shared copyrighted files.)1717See Chris Eggertsen, Labels & Publishers Win $1 Billion Piracy Lawsuit Against Cox Communications, Billboard (Dec. 19, 2019), https://www.billboard.com/articles/business/legal-and-management/8546842/cox-1-billion-piracy-lawsuit-labels-publishers [https://perma.cc/
XXM6-9AMW] (describing the verdict in the Cox suit and the plaintiffs’ claim that Cox failed to exercise reasonable measures to prevent copyright infringement). Close But thanks in large part to copyright’s fair use doctrine, robots that do the same have traditionally been granted broad latitude. As recently as 2016, James Grimmelmann observed that “[q]uietly, invisibly almost by accident, copyright has concluded that reading by robots doesn’t count.”1818James Grimmelmann, Copyright for Literate Robots, 101 Iowa L. Rev. 657, 658 (2016). Close
Today, however, that truism is no longer quite so true. After decades of allowing—or even just plain ignoring—machine copying, copyright owners and courts have begun to loudly and visibly push back against the copyright system’s permissive attitude towards machine copying.1919See infra notes 21–29 and accompanying text. Close True, a years-long saga between Google and the Author’s Guild (hereafter, the Google Books Cases) offers hope to robot readers.2020See infra subpart II(A). Close But countervailing, even contradictory, moves in other decisions2121See, e.g., Fox News Network, LLC v. TVEyes, Inc., 883 F.3d 169, 174 (2d Cir. 2018) (distinguishing the Google Books case); Associated Press v. Meltwater U.S. Holdings, Inc., 931 F. Supp. 2d 537, 543–44, 550 (S.D.N.Y. 2013) (rejecting Meltwater’s fair use defense). Close have thrown the legality of machine copying into question. Complicating matters more, the nature of machine copying has also changed as the use of datasets has expanded from narrower “text data mining” (TDM) systems that read existing content to more sophisticated systems like MuseNet that actively learn from it. Given the doctrinal uncertainty and the rapid development of ML technology, it is unclear whether machine copying will continue to be treated as fair use.
There are reasons to think courts in the future won’t be so sympathetic to machine copying. Fair use doctrine in the last quarter century has focused on the transformation of the copyrighted work. ML systems, however, rarely transform the databases they train on; they are using the entire database, and for a commercial purpose at that. Courts may view that as a kind of free riding they should prohibit, particularly when the companies training ML models tend to be giant multinationals and the owners of individual photographs and books are often small, sympathetic plaintiffs. And many of those plaintiffs may be motivated to sue, either by the extraordinary statutory damages copyright law offers them or because they don’t want their work used to train an AI that might someday replace their job or might use the data in undesirable ways.2222We may learn more about the future of the transformative use doctrine this year as the Supreme Court takes up Google v. Oracle. Google LLC v. Oracle Am., Inc., No.18-956 (U.S. argued Oct. 7, 2020). Aside from copyright, the bulk collection of training data also implicates a host of other laws that could prove formidable barriers in the future. See, e.g., Kashmir Hill & Aaron Krolik, How Photos of Your Kids Are Powering Surveillance Technology, N.Y. Times (Oct. 11,
2019), https://www.nytimes.com/interactive/2019/10/11/technology/flickr-facial-recognition.html [https://perma.cc/87G7-68VJ] (discussing how Illinois’s Biometric Information Privacy Act of 2008 could have massive liabilities implications for companies using the photos of Illinois inhabitants); Orin Kerr, Norms of Computer Trespass, 116 Colum. L. Rev. 1143, 1165 (2016) (discussing how the Criminal Fraud and Abuse Act creates a chilling effect for those deploying the tools necessary to collect training data). Close
Further, these uncertainties arrive at a time when copyright’s attitude toward robotic readership is under increasing fire in a court of equal importance to those established by Article III—the court of public opinion.2323See, e.g., Joshua New, Copyright Law Should Not Restrict AI Systems from Using Public Data, Ctr. for Data Innovation (Oct. 14, 2019), https://www.datainnovation.org/2019/10/copyright-law-should-not-restrict-ai-systems-from-using-public-data/ [https://perma.cc/QCM3-37ER] (describing the backlash to AI’s use of publicly available photographs to train its algorithms). Close Hardly a week now passes without headlines from media outlets, thought leaders, or advocacy organizations decrying new ML systems that push data usage norms to the limits.2424See, e.g., Gregory Bobillot, ‘Techlash’—How Big Tech Is Influencing Your Thinking, Fin. Times (May 10, 2018), https://www.ft.com/video/3339f59e-f760-4bc7-b359-3899fabbd190 [https://perma.cc/4M9J-8WLJ] (describing how big tech influences users’ mood and thinking); Rachel Botsman, Dawn of the Techlash, Guardian (Feb. 10, 2018), https://www.theguardian.com/commentisfree/2018/feb/11/dawn-of-the-techlash [https://perma.cc/DPS5-CEGD] (discussing the backlash against AI); Emails Show How Amazon Is Selling Facial Recognition System to Law Enforcement, ACLU of Nor. Cal. (May 21, 2018), https://www.aclunc.org/news/emails-show-how-amazon-selling-facial-recognition-system-law-enforcement [https://perma.cc/7APP-2APX] (describing the racial and privacy concerns surrounding law enforcement’s use of Amazon’s facial recognition technology); John Rubino, ‘Tech Wreck,’ ‘Techlash,’ ‘Techmageddon’—Whatever You Call It, Wall Street Is Terrified of It, Seeking Alpha (Mar. 29, 2018, 4:58 AM), https://seekingalpha.com/article/4159827-tech-wreck-techlash-techmageddon-whatever-call-wall-street-terrified [https://perma.cc/QG9U-QSCV] (discussing the financial threat on Wall Street from
big tech companies dominating markets); Eve Smith, The Techlash Against Amazon, Facebook
and Google—And What They Can Do, Economist (Jan. 20, 2018), https://www.economist
.com/briefing/2018/01/20/the-techlash-against-amazon-facebook-and-google-and-what-they-can-do [https://perma.cc/G3HX-H3BT]; The Techlash Has Just Begun, Axios (Jan. 10, 2018), https://www.axios.com/the-techlash-1515609266-e27ca299-0031-460a-96f1-db842ec88121.html [https://perma.cc/446C-KRZD] (discussing the danger from Amazon, Facebook, and Google’s growing control of consumer data). Close Once a relatively obscure topic, debates over ML and copyright law are now the subject of New York Times pieces headlined How Photos of Your Kids Are Powering Surveillance Technology2525See, e.g., Hill & Krolik, supra note 22 (noting that “[m]illions of Flickr images were sucked into a database called MegaFace. Now some of those faces may have the ability to sue”); Olivia Solon, Facial Recognition’s ‘Dirty Little Secret’: Millions of Online Photos Scraped Without Consent, NBC News (Mar. 17, 2019), https://www.nbcnews.com/tech/internet/facial-recognition-s-dirty-little-secret-millions-online-photos-scraped-n981921 [https://perma.cc/G4TQ-DWUM] (highlighting a growing and potentially problematic trend of photographs being used without the permission of either the photographers or the subjects of the images in AI training datasets created for facial recognition purposes). Close and takedown campaigns of “deepfake”2626Bobby Chesney and Danielle Citron describe “deepfakes” as ML-based “[t]echnologies for altering images, video, or audio (or even creating them from scratch) in ways that are highly realistic and difficult to detect. See Bobby Chesney & Danielle Citron, Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security, 107 Calif. L. Rev. 1753, 1757 (“We use that label here more broadly, as shorthand for the full range of hyper-realistic digital falsification of images, video, and audio.”); see also Samantha Cole, We Are Truly Fucked: Everyone Is Making AI-Generated Fake Porn Now, Vice: Motherboard (Jan. 24, 2018, 12:13 PM), https://motherboard.vice.com/en_us/article/bjye8a/reddit-fake-porn-app-daisy-ridley [https://perma.cc/7U9T-XFLQ] (“[T]echnology[] allows anyone with sufficient raw footage to . . . convincingly place any face in any video.”). Close videos satirizing celebrities such as Kim Kardashian.2727As Bobby Chesney and Danielle Citron note, “Pornographers have been early adopters of the technology, interposing the faces of celebrities into sex videos.” See Chesney & Citron, supra note 26; see also id. at 1793–94 (floating copyright law as a potential, albeit imperfect, mechanism for curbing deepfakes abuses); Alex Engler, Fighting Deepfakes When Detection Fails, Brookings (Nov. 14, 2019), https://www.brookings.edu/research/fighting-deepfakes-when-detection-fails/ [https://perma.cc/73NG-K8NT] (same). Close The public reacted with fury when it learned that Clearview AI had collected over three billion photos from the internet to build a facial recognition database it licensed to law enforcement.2828See Beryl Lipton, Records on Clearview AI Reveal New Info on Police Use, MuckRock (Jan. 18, 2020), https://www.muckrock.com/news/archives/2020/jan/18/clearview-ai-facial-recogniton-records/ [https://perma.cc/9WFE-DU6Y] (discussing the revelation that Clearview AI uses open source images to assist law enforcement with facial recognition). Close Prominent copyright professor Tim Wu tweeted that it “should be the target of a class-action copyright lawsuit.”2929@superwuster, Twitter (Jan. 18, 2020, 7:26 AM), https://twitter.com/superwuster/status/1218524978225741824?s=20 [https://perma.cc/HRV8-WFTV]. Close This unique confluence of commercial, normative, and doctrinal factors has teed up what may well be one of the most important legal questions of the coming century: Will copyright law allow robots to learn?
In this Article, we argue that ML systems should generally3030The word “generally” here is intentional. As we discuss below, machine copying should not be permissible in every conceivable instance. Fair use in the machine learning context, for example, should be sensitive to the purpose of the ML system and what it eventually produces as output. Close be able to use databases for training, whether or not the contents of that database are copyrighted. There are good policy reasons to do so. First, we need to encourage people to compile new databases and to open them up for public scrutiny or innovation. Broad access to training sets will further these objectives, ultimately making artificial intelligence systems using ML algorithms better, safer, and fairer.3131“Fair” from both a commercial perspective and “fair” as the term is understood in the context of social justice and equity. See, e.g., Benjamin L.W. Sobel, Artificial Intelligence’s Fair Use Crisis, 41 Colum. J.L. & Arts 45, 96 (2017) (arguing that “unauthorized use of copyrighted data for the sole purpose of debiasing an expressive program” should fall under fair use protections). See generally Amanda Levendowski, How Copyright Law Can Fix Artificial Intelligence’s Implicit Bias Problem, 93 Wash. L. Rev. 579 (2018) (exploring how copyright law could improve and exacerbate bias within machine learning systems). Close Second, an ML system’s use of the data often is transformative as that term has come to be understood in copyright law, because even though it doesn’t change the underlying work, it changes the purpose for which the work is used.3232See infra subpart II(A). Close And because training sets are likely to contain millions of different works with thousands of different owners, there is no plausible option simply to license all of the underlying photographs, videos, audio files, or texts for the new use. So allowing a copyright claim is tantamount to saying, not that copyright owners will get paid, but that the use won’t be permitted at all, at least without legislative intervention.3333Cf. Keller v. Elec. Arts, Inc., 724 F.3d 1268, 1269 (9th Cir. 2013) (holding that thousands of former college athletes each had the right to prevent Electronic Arts (EA) from making a college football video game; EA shut down the game altogether rather than try to get permission from all the former players). Close While we share some of the concerns about the uses to which ML systems may be put, copyright is not the right tool to regulate those abuses.
The implications of this debate go beyond machine learning. Understanding why the use of copyrighted works by ML systems should be fair actually reveals a significant issue at the heart of all copyright law. People, like machines, often copy expression when they are only interested in learning the ideas conveyed by that expression. That’s true of most ML training protocols. The ML system wants photos of stop signs so it can learn to recognize stop signs, not because of the artistic choices you made in lighting or composing your photo.3434There is a sense in which the creative choices matter even here. The AI is likely to want to see photos of stop signs in a variety of lights, angles, and conditions to train better. But it doesn’t value the creativity as creativity. Close Similarly, a natural language generation system wants to see what you wrote to learn how words are sequenced in ordinary conversation, not because it finds your prose particularly expressive or because it wants to use your turn of phrase.3535Some systems blur this line. Natural language generators do want to know how words are sequenced together in ordinary human speech, so they may be interested in the way a particular text phrases things. But it is only short phrases that are likely to be relevant to the ML system, and short phrases are supposed to be uncopyrightable. See Material Not Subject to Copyright, 37 C.F.R. § 202.1 (2019) (“(a) Words and short phrases such as names, titles, and slogans [are not subject to copyright]”). But see Hall v. Swift, 786 Fed. Appx. 711 (9th Cir. 2019) (permitting suit against Taylor Swift based on the six-word phrase “players gonna play, haters gonna hate”). Close
ML systems are not alone in wanting works for reasons that have little to do with the authors’ actual expression. The issue arises in lots of other copying contexts. In American Geophysical Union v. Texaco, 3636Am. Geophysical Union v. Texaco, Inc., 60 F.3d 913 (2d Cir. 1994). Close for example, the defendants were interested only in the ideas in scientific journal articles; photocopying the article was simply the most convenient way of gaining access to those ideas.3737Id. at 918–19. Close Other examples include two pending Supreme Court cases, copyright disputes over software interoperability cases like Lotus v. Borland3838Lotus Dev. Corp. v. Borland Int’l, Inc., 516 U.S. 233 (1996) (per curiam). Close and Google v. Oracle,3939Google LLC v. Oracle Am., Inc., No.18-956 (U.S. argued Oct. 7, 2020). Close current disputes over copyright in state statutes and rules adopted into law,4040Georgia v. Public.Resource.Org, Inc., 140 S. Ct. 1498, 1504–06 (2020). Close and perhaps even the tangled morass of cases around copyright protection for the artistic aspects of utilitarian works like clothing, bike racks, and even Bikram yoga.4141Star Athletica, L.L.C. v. Varsity Brands, Inc., 137 S. Ct. 1002, 1004 (2017) (clothing); Brandir Int’l v. Cascade Pac. Co., 834 F.2d 1142 (2d Cir. 1987) (bike racks); Bikram’s Yoga College of India, L.P. v. Evolation Yoga LLC, 803 F.3d 1032, 1034 (9th Cir. 2015) (yoga). Close In all these cases, copyright law is being used to target defendants who actually want access to something the law is not supposed to protect—the underlying ideas, facts, or functions of the work.
Copyright law should permit copying of works for non-expressive purposes—at least in most circumstances. While some courts have considered these issues under merger or protectability principles, occasionally denying protection altogether to functional works, the real issue in the cases we consider here is usually not that the work as a whole is unprotectable. Rather, the issue is the fit between what the law protects and what the defendant wants. When the defendant copies a work for reasons other than to have access to the protectable expression in that work, fair use should consider under both factors one and two whether the purpose of the defendant’s copying was to appropriate the plaintiff’s expression or just the ideas.
That said, the protection afforded by fair learning should not be unlimited. When learning is done to copy expression, for example, by training an ML system to make a song in the style of Ariana Grande, the question of fair use can—and should—become much tougher. But, as importantly, we don’t want to allow the copyright on the creative pieces to end up controlling the unprotectable elements.
In Part I, we discuss how machines learn and how this new technological paradigm of “machine learning” differs from other more doctrinally familiar paradigms. In Part II, we explain why ML systems might run afoul of copyright when it comes to compiling training sets and discuss how copyright law today treats such uses (and how it should in the future). Finally, in Part III, we discuss the broader implications of our principle of fair learning for a variety of other copyright disputes.
I. The Voracious Learner
The last decade has seen the subfield of computer science known as “machine learning” (ML) take the world by storm.4242For a short overview of this history, see, for example, Casey & Lemley, supra note 7, at 301–03 (tracing the rise of ML after AI’s so-called “winter”). Close This Part briefly overviews the technological constituents of ML, distinguishing it from “text data mining” (TDM) technologies that copyright law has dealt with—with varying degrees of success—in the last several decades. It then outlines some of the implications ML technologies pose for copyright law,4343Numerous previous works have focused on the subject of AI authorship. See, e.g., Sixty-Eighth Annual Report of the Register of Copyrights for the Fiscal Year Ending June 30, 1965, at 4–5 (1966) (offering cursory discussion of computer-generated works); Boyden, supra note 8, at 378 (asking “who should be considered the author?” of computer-generated works); Bridy, supra note 8, at 18–20 (same); Samuelson, supra note 8, at 1192–94 (same); Ralston, supra note 8, at 300 (same); Grimmelmann, supra note 8, at 403–04 (same). Comparatively little attention, however, has been paid to copyright protections afforded to the collections of training data necessary to power these systems. But see Matthew Sag, The New Legal Landscape for Text Mining and Machine Learning, 6 J. Copyright Soc’y USA 291, 292–93 (2019) (briefly discussing ML technologies but focusing primarily on expert-based TDM systems); Grimmelmann, supra note 18, at 669–70 (focusing primarily on expert-based TDM systems). Close laying the foundation for a more substantive discussion of ML and copyright’s “fair use”4444For a fully-fledged definition of “fair use,” see infra Part II. Close doctrine in the parts that follow.
A. From Readers to Learners
It’s no exaggeration to describe the last three decades as dominated by robot, not human, readership. Beginning with the advent of internet technologies in the 1980s and 1990s,4545See Grimmelmann, supra note 18, at 659. As Grimmelmann wrote: In a world of books and other pre-digital technologies, “copyright . . . left reading, listening, and viewing unconstrained.” Ordinary acts of reading did not result in any new copies, and hence did not trigger any of the copyright owner’s exclusive rights; nor did readers have access to technologies that would have made copying easy. Id. Close the world saw its information go from mostly physical to mostly digital.4646TDM techniques were in use before the advent of the internet. See, e.g., Don R. Swanson, Fish Oil, Raynaud’s Syndrome, and Undiscovered Public Knowledge, 30 Persp. Biology & Med. 7, 8–9 (1986) (using TDM techniques—prior to the advent of the internet—to obtain and determine the connection between academic papers). Close This trend had a certain inexorable logic to it. Instead of locking humanity’s collective knowledge on library shelves or magnetic film rolls, digitization allowed us to interact with information without being throttled by the processing speeds of the physical world. Once digitized, information could be replicated, transmitted, altered, searched, and analyzed at the click of a button. And that, in turn, gave rise to a powerful class of technologies specializing in doing exactly that,4747See, e.g., Jerome H. Reichman & Ruth L. Okediji, When Copyright Law and Science Collide: Empowering Digitally Integrated Research Methods on a Global Scale, 96 Minn. L. Rev. 1362, 1366 (2012) (noting that “[t]he combination of massive storage capacity, powerful data manipulation techniques, and graphical capabilities has revolutionized both how basic research is conducted and how the resulting knowledge is preserved and disseminated in nearly all fields of science”). Close known in some circles by the umbrella term “text data mining” (TDM) tools.4848According to Matthew Sag, “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.” Sag, supra note 44, at 294–95 (citing Eleanor Dickson, Megan Senseney, Beth Namachchivaya & Bertram Ludäscher, Synthesis of Cross-Stakeholder Perspectives on Text Data Mining with Use-Limited Data: Setting the Stage for an IMLS National Forum 5 (2018)). Sag further clarifies: “The term ‘text’ in this context is broad enough to include fixed images, sound recordings, and audio-visual works.” Id. at 295. Close Want, for example, to know when the English-speaking world began using the term “robot” in place of the word “automaton” to describe humanoid machines? Doing so requires mining as many written works produced during the relevant time period as possible and then constructing search queries to isolate and identify that data. And that process begins by instructing a robot (or perhaps automaton?) to read all of them.
Each day, with most of us scarcely noticing,4949See Grimmelmann, supra note 18 (noting that this process occurs almost invisibly). Close TDM technologies of this variety read and index untold terabytes of data. Search engines read—or in software parlance “crawl”—all of the internet’s text, images, audio, and video data in hopes of organizing it; social media platforms read and categorize all of our cat photos; the list goes on. So, while the notion of a typical “reader” might invoke an image of a student flipping through the pages of an overgrown textbook, the reality is that modern reading is an overwhelmingly robotic affair. As James Grimmelmann observed, “if you count by the total number of words read, robotic reading is now overwhelmingly more common than human.”5050Id. at 681. “Words” is perhaps too narrow a term here and also includes content such as video, audio, and images. Close
Today, some thirty years after the rise of the robotic reader, all signs suggest another paradigm shift in readership is underway—this time with a new class of robotics technologies that is less focused on passively reading information than on actively learning from it. The technology comes from a subfield of computer science, known as machine learning (ML), that exploded onto the scene in the last few decades.5151Tom M. Mitchell, Machine Learning 1–2 (1997). Close We have elsewhere provided an overview of ML that we need not revisit in detail here.5252See, e.g., Casey & Lemley, supra note 7, at 303–07 (overviewing the technology’s fundamentals). Close But to understand precisely why this new type of “robotic learner” is in ascent, it’s worth briefly describing how it differs from the mere “robotic readers”5353See generally Grimmelmann, supra note 18 (referring to expert-based systems by this coinage). Close that came before it.
Though sometimes conflated with the TDM tools of yesteryear,5454While the Venn diagrams of TDM tools and ML systems do overlap if one relies on a broad enough definition of TDM, we’re referring to TDM tools that rely on explicit rules-based approaches to mining their insights. Close ML technologies differ in at least one fundamental regard. Unlike “expert”5555The term “expert,” here, refers to a software system programmed to follow an explicit set of hard-coded rules without undergoing a process of learning an implicit set of rules through iteration. Close TDM systems that actually “slavishly”5656L. Batlin & Son, Inc. v. Snyder, 536 F.2d 486, 490 (2d Cir. 1976) (quoting 1 Melville B. Nimmer, The Law of Copyright § 6, at 10.2 (1975)). Close follow a set of rules engineers have hard coded for them, ML systems approach problem-solving tasks in much the same way humans do: by trying to learn. Rather than attempting to imagine and encode every step involved in distinguishing a cat photo from a dog photo, engineers instead create an environment where robots can develop their own rules for making the distinction through trial and error. An ML system trying to recognize cats, for instance, would be exposed to a series of examples. And with each attempt, the system would improve its chances on the next one by learning from its successes and failures.
Freed from the limitations of an expert-based approach, ML systems have proven capable of reaching heights of achievement far beyond TDM. Whereas TDM systems might effectively catalogue the world’s radiology images and perhaps make them searchable by humans, ML systems can now outperform human radiologists at diagnosing medical ailments from them.5757Taylor Kubota, Algorithms Better at Diagnosing Pneumonia than Radiologists, Stanford Med. News Ctr. (Nov. 15, 2017), https://med.stanford.edu/news/all-news/2017/11/algorithm-can-diagnose-pneumonia-better-than-radiologists.html [https://perma.cc/A5GQ-FLS4]. Close And it’s this ability to learn instead of simply following explicit instructions that has proven central to unlocking vast machine potential.
Yet, while ML approaches resemble human learning in many ways, there’s at least one aspect in which humans and machines radically diverge: Machines, it turns out, are remarkably slow learners. To make the judgments we want them to make, ML systems must first be exposed to thousands, millions, or even billions of examples—all of which are collected and stored in a database known as a “training set.” And unlike some TDM systems, which may make only transitory copies,5858Michael W. Carroll, Copyright and the Progress of Science: Why Text and Data Mining Is Lawful, 53 U.C. Davis L. Rev. 893, 959 (2019). As Carroll discusses, the legal status of temporary copies is uncertain under copyright law. Id. Close ML systems generally require a more permanent training data set to test successive iterations of the software against. With a large enough training set,5959And enough computing resources, of course. Close virtually no problem-solving task is insurmountable. As we saw in the introduction, machines can compose music,6060See supra notes 2–5 and accompanying text; see also Lori Dorn, You Can’t Take My Door, A Country Song Created by a Neural Network That Studied a Catalog of Country Hits, Laughing Squid (Apr. 15, 2019), https://laughingsquid.com/country-song-created-by-neural-network/ [https://perma.cc/HED7-FT8E] (describing a “predictive AI country song” created by “training a neural network to learn country music hits and then produce one of its own”). Close generate artwork,6161See Mike Murphy, Computers Can Now Paint Like Van Gogh and Picasso,
Quartz (Sept. 6, 2015), https://qz.com/495614/computers-can-now-paint-like-van-gogh-and-picasso/ [https://perma.cc/ADB2-KMFN] (discussing a machine learning system that can “interpret the styles of famous painters and turn a photograph into a digital painting in those styles”); see also Gabe Cohn, AI Art at Christie’s Sells for $432,500, N.Y. Times (Oct. 25, 2018), https://www.nytimes.com/2018/10/25/arts/design/ai-art-sold-christies.html [https://perma.cc/Q2PF-F3SN] (documenting the sale of a portrait “produced by artificial intelligence”). Close play complex strategy games,6262Cade Metz, In a Huge Breakthrough, Google’s AI Beats a Top Player at the Game of Go, WIRED (Jan. 27, 2016), https://www.wired.com/2016/01/in-a-huge-breakthrough-googles-ai-beats-a-top-player-at-the-game-of-go/ [https://perma.cc/S27A-E3KY]. “Go” is an ancient Eastern strategy game that is comparable to chess, though far more computationally complex. Id.; Tom Simonite, Can Bots Outwit Humans in One of the Biggest Esports Games?, WIRED (June 25, 2018, 10:00 AM), https://www.wired.com/story/can-bots-outwit-humans-in-one-of-the-biggest-esports-games/ [https://perma.cc/7MX5-J6WS]. DotA is one of the internet’s most popular real time strategy games and is more difficult for AI systems than Go or chess. Id. Close and outperform human professionals at the pinnacles of their careers.6363FDA Approves AI-Powered Diagnostic That Doesn’t Need a Doctor’s Help, MIT Tech. Rev. (Apr. 11, 2018), https://www.technologyreview.com/2018/04/11/3052/fda-approves-first-ai-powered-diagnostic-that-doesnt-need-a-doctors-help/ [https://perma.cc/9UTL-Y5V4]. Close And all signs suggest this is just the beginning of the technology’s long arc of progress.
There is at least one obstacle standing in the way of ML’s seemingly inexorable learning curve. Virtually all the data used to compile training sets is protected by copyright. And just as was true of TDM readers in the ’90s, ’00s, and ’10s, this new breed of robotic readers appears destined to give rise to a host of doctrinal and policy challenges in the years ahead. Indeed, it already has.
B. Copyrights, Copyrights Everywhere
It’s shockingly easy to create a copyrighted work. Copyrights cover a broad swath of creations, from the written word to art of all types to software, dance, and even architecture.646417 U.S.C. § 102(a) (2018). The outer limits of copyright law involve things like gardens that evolve on their own. Kelley v Chi. Park Dist., 635 F.3d 290, 292 (7th Cir. 2011) (holding that a park garden was not “fixed” and therefore ineligible for copyright protection). Close The standard for establishing copyright protection is low—you need only have an “original work of authorship” and record it in some more-than-transitory form.656517 U.S.C. § 102(a) (2018). Close And both of those requirements are so trivial as to be almost meaningless. Anything you write longer than a sentence—and art as simple as a red canvas or an accidental stroke of the pen—is likely to meet the originality standard.6666See Bleistein v. Donaldson Lithographing Co., 188 U.S. 239, 250 (1903) (observing that even “a very modest grade of art has in it something irreducible” that may be copyrightable); Alfred Bell & Co. v. Catalda Fine Arts, 191 F.2d 99, 102 (2d Cir. 1951) (noting that “[n]o large measure of novelty is necessary” to meet the originality standard). Close Moreover, a work can be “fixed” in almost any form, including a simultaneous recording.6767See MAI Sys. Corp. v. Peak Comp., Inc., 991 F.2d 511, 517–18 (9th Cir. 1993) (finding that temporary copies loaded in the memory of a computer were “fixed”); see also Stenograph L.L.C. v. Bossard Assocs., Inc., 144 F.3d 96, 102 (D.C. Cir. 1998) (same); NFLC, Inc. v. Devcom Mid-Am, 45 F.3d 231, 235 (7th Cir. 1995) (same). Contra Cartoon Network LP v. CSC Holdings, Inc., 536 F.3d 121, 129–30 (2d Cir. 2008) (concluding that temporary copies are unfixed in the context of ISPs); CoStar Grp., Inc. v. LoopNet, Inc., 373 F.3d 544, 550–51 (4th Cir. 2004) (same). Close Once you’ve created the work, no special formalities or applications need be dispatched to protect it. Copyright protections apply immediately.6868See 17 U.S.C. § 302(a) (2018) (establishing that “[c]opyright in a work . . . subsists from its creation”); cf. Chris Sprigman, Reform(aliz)ing Copyright, 57 Stan. L. Rev. 485 (2004) (arguing that we should bring back formalities to avoid accidental infringement). Close
The threshold is so low, in fact, that it is virtually impossible to go a day without creating multiple copyrighted works. You’d have to stay off email and social media entirely, abstain from selfies or videos, write only extremely short texts, and refrain from doodling. Even that might not be enough. If you have Alexa turned on in your house, your conversations may well be copyrightable.
Once you’ve got all these copyrights, they turn out to be surprisingly hard to undo. Copyright lasts for the life of the author plus 70 years,6969See 17 U.S.C. § 302(a) (2018) (“Copyright . . . endures for a term consisting of the life of the author and 70 years after the author’s death.”). Close so even things created in the mid-1920s are still subject to copyright protection. There is no clear means for abandoning a copyright or dedicating it to the public domain.7070See, e.g., Jerry Brito & Bridget Dooling, An Orphan Works Affirmative Defense to Copyright Infringement Actions, 12 Mich. Telecomm. & Tech. L. Rev. 75, 76 (2005) (describing how copyright law “renders untouchable a large swath of existing artistic, literary, and other works because if a work’s copyright owner cannot be found to secure their permission to use the work, then no one will ultimately use the work lest they risk liability for copyright infringement”); Dave Fagundes & Aaron Perzanowski, Abandoning Copyright, 62 Wm. & Mary L. Rev. 487 (2020) (same); Olive Huang, U.S. Copyright Office Orphan Works Inquiry: Finding Homes for the Orphans, 21 Berkeley Tech. L. J. 265, 265 (2006) (same). Close And even if you do manage to license it to the public at large, perhaps via an open source or creative commons license designed to get around this difficulty, you can always change your mind thirty-five years down the line and get your copyright back, and no agreement to the contrary can stop you.7171See generally Peter S. Menell & David Nimmer, Pooh-Poohing Copyright Law’s “Inalienable” Termination Rights, 57 J. Copyright Soc’y USA 799 (2010) (discussing the implication of the termination of transfer right). Close
The combination of all these factors—a broad range of things protected, the very low standard for copyrightability, the long life of copyrights, and the inability to disclaim them—means that literally tens of billions of new copyrighted works are created every day, and an almost uncountable number of things are copyrighted. Most of them, of course, are worthless. But they’re still protected by federal law.
Given this broad definition, virtually all the training sets used by ML systems include copyrighted works. Object recognition tools and optical scanning software need to train on photographs. So do self-driving cars, self-flying planes, warehouse robots, and any other entity that needs to identify and navigate around obstacles. Speech-recognition and speech-generation systems need to train on recorded audio inputs from radio, TV, movies, and everyday conversations, and virtually all of those were created in the last hundred years and are potentially subject to copyright. Text generation and translation software similarly need to train on a corpus of written works, and if you don’t want your texts to be filled with “thous” and “dosts” that means training on works written within the last hundred years. All those things are copyrightable.
Fortunately, there are a number of doctrinal protections afforded to robot readers. The first, and most superficial, is that facts themselves are not subject to copyright protection.7272See, e.g., 17 U.S.C. § 102(b) (2018) (excluding from copyright protection “any idea, procedure, process system, method of operation, concept, principle, or discovery”); Feist Publ’ns, Inc. v. Rural Tel. Serv. Co., 499 U.S. 340, 340 (1991) (“[C]opyright protection extends only to those components of the work that are original to the author, not to the facts themselves.”). Close So an ML system that needs only pure facts (say, a stock trading algorithm that only studies prior stock purchases) might seem to be off the hook. But a compilation of uncopyrightable facts can itself be copyrighted as long as there is even minimal originality in the selection or arrangement of those facts.7373Feist, 499 U.S. at 340. Close Thus, many databases can be copyrighted even though the individual pieces of data within them are not protected.7474See, e.g., CDN Inc. v. Kapes, 197 F.3d 1256, 1262 (9th Cir. 1999) (holding that a list of wholesale prices for rare coins listed by publisher contained sufficient originality to qualify for copyright protection); CCC Info. Servs., Inc. v. Maclean Hunter Mkt. Reps., Inc., 44 F.3d 61, 63 (2d Cir. 1994) (reversing district court decision finding list of estimated used car valuations to be unprotectable facts and holding that the Red Book numbers as well as the selection and arrangement of the Red Book to be protectable expression). Close A database that is comprehensive—including everything in the field—and not creatively organized won’t get protection,7575Though many of those may effectively get copyright-like protection through the enforcement of ubiquitous terms of use. See generally Mark A. Lemley, Terms of Use, 91 Minn. L. Rev. 459 (2006) (describing this phenomenon). The case that started the “no need to agree” revolution in contract law was itself a case in which the defendant copied a database that contained only uncopyrightable facts and no creative selection or arrangement. ProCD, Inc. v. Zeidenberg, 86 F.3d 1447, 1448−49 (7th Cir. 1996) Nonetheless, Zeidenberg was held liable for copying the “unprotectable” database. Id. Close but many databases will. This includes things we don’t normally think of as databases, like medical admissions forms or police booking sheets, each of which selects data about its subjects in an arguably creative way.7676Compare Practice Mgmt. Info. Corp. v. Am. Med. Ass’n, 121 F.3d 516, 518, 520 (9th Cir. 1997), amended by 133 F.3d 1140 (9th Cir. 1998) (holding that a federal agency’s adoption of work as the standard in preparation of Medicare and Medicaid claims did not render copyright invalid), with Southco, Inc. v. Kanebridge Corp., 390 F.3d 276, 286–87 (3d Cir. 2004) (en banc) (holding that part numbers used to identify and distinguish among types of screw fasteners are not protectable); Bibbero Sys., Inc. v. Colwell Sys., Inc., 893 F.2d 1104, 1106–08 (9th Cir. 1990) (holding that a medical billing form was uncopyrightable because it was “simply a blank form which gives doctors a convenient method for recording services performed”). Close
The copyrightability of databases presents a hurdle for ML systems. But this hurdle is not insurmountable. An engineer or company that wants to train its system using a comprehensive existing database will sometimes (but not always) be able to license that database, and it seems reasonable that it should have to. Precisely because the database collects a lot of information, it offers a valuable form of “one-stop shopping” for ML systems.
There are circumstances where that will not be true. Companies might not voluntarily license their databases to competitors. That’s a problem only if the company that owns the database has exclusive access to the type, volume, or quality of the data stored in the database. That will be true sometimes—think Google’s database of stop signs and its self-driving-car business—but not always. And it should worry us only if we think there are policy reasons to make sure that each competitor has the best possible training data. There’s a good argument for that in some sectors that involve public health and safety, such as medical technologies and safety-promoting technologies (including, perhaps, self-driving cars).
There are also policy concerns related to the transparency and accuracy of the algorithms that employ them. We might want to know what’s in our training sets where their use has public policy implications. Database transparency, for example, might allow us to figure out whether a photo-recognition algorithm is bad at identifying minorities,7777See, e.g., Jacob Snow, Amazon’s Face Recognition Falsely Matched 28 Members of Congress with Mugshots, ACLU Blog (July 26, 2018, 8:00 AM), https://www.aclu.org/
blog/privacy-technology/surveillance-technologies/amazons-face-recognition-falsely-matched-28 [https://perma.cc/QQ6V-7YWW] (highlighting this potential, albeit while using a dubious methodology). Close a criminal-sentencing algorithm replicates racial bias,7878See, e.g., Julia Angwin, Jeff Larson, Surya Mattu & Lauren Kirchner, Machine Bias, ProPublica (May 23, 2016), https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing [https://perma.cc/MTV4-DZX8] (noting the problem of bias in sentencing assessments). Close a credit-rating algorithm uses potentially biased data,7979See, e.g., Kaveh Waddell, How Algorithms Can Bring Down Minorities’ Credit Scores, Atlantic (Dec. 2, 2016), https://www.theatlantic.com/technology/archive/2016/12/how-algorithms-can-bring-down-minorities-credit-scores/509333/ [https://perma.cc/W4YK-W32T] (describing the potential of this phenomenon in machine learning). Close or the like. In those cases, we might want to open access to the database to regulators or compel licensing of the databases to competitors on reasonable terms, just as copyright does in myriad other areas.8080See, e.g., 17 U.S.C. §§ 111, 114–115, 119 (2018) (discussing limitations on exclusive rights in various contexts). While property rights advocates worry that these compulsory licensing schemes prevent bargaining and therefore undermine the incentives to create, the evidence doesn’t support that in copyright. See Mark A. Lemley, Contracting around Liability Rules, 100 Calif. L. Rev. 463, 476–77 (2012) (providing evidence that parties bargain around compulsory licenses in copyright law). Close
But copyright in databases is only a small part of the problem ML systems face. A much more significant and less tractable problem is the copyright in the individual components of the database. It is one thing to license a database of photos of faces from the entity that compiled it. It is quite another to try to get the rights to each individual photo from the millions (or hundreds of millions) of individuals who took them in the first place. ML engineers or companies that want to compile a training set of all books, or all music, or all video content face a similar problem. While books and music tend to have more concentrated copyright ownership, there are still millions of authors and artists and thousands of commercial publishers out there.
Nor will it matter that an ML company didn’t get those books or images directly from the copyright owner. Copyright is a strict liability offense.8181See generally R. Anthony Reese, Innocent Infringement in U.S. Copyright Law: A History, 30 Colum. J.L. & Arts 133 (2007) (tracing the history of copyright infringement that lacked intent); Patrick R. Goold, Is Copyright Infringement a Strict Liability Tort?, 30 Berkeley Tech. L.J. 305 (2015) (further exploring this phenomenon in the context of tort law). Close Acting reasonably in getting a license from the database owner won’t help you if the database owner doesn’t have a license for each and every one of the hundreds of millions of works, even if they claim they do.8282See Lipton v. Nature Co., 71 F.3d 464, 471 (2d Cir. 1995) (holding that copying from a third source without knowledge that that source was infringing does not absolve one from infringement). Close And even if they do have a license, that license might not cover all of the rights needed from all the owners,8383To take just one example, streaming a song requires multiple different licenses for
rights usually held by different entities: the public performance right in the sound
recording, the reproduction right in the sound recording, and the public performance right in
the underlying musical composition. See Jason Koransky, Digital Dilemmas: The
Music Industry Confronts Licensing for On-Demand Streaming Services, Am. Bar Ass’n
(Jan. 2016), https://www.americanbar.org/groups/intellectual_property_law/publications/landslide/2015-16/january-february/digital-dilemmas-music-industry-confronts-licensing-on-demand-streaming-services/ [https://perma.cc/D4UM-B735] (describing the complexities of music licensing) See generally Mark A. Lemley, Dealing with Overlapping Copyrights on the Internet, 22 U. Dayton L. Rev. 547 (1997) (detailing the web of overlapping copyrights with regard to internet activities). Close or it might be limited to uses that do not include the previously uncontemplated use by an ML system in training its algorithm.
Nor will it matter that the work is used only inside the ML system and isn’t copied in a final output. Intermediate copying is still copying and can still be infringing.8484Walker v. Univ. Books, 602 F.2d 859, 864 (9th Cir. 1979) (“[T]he fact that an allegedly infringing copy of a protected work may itself be only an inchoate representation of some final product to be marketed commercially does not in itself negate the possibility of infringement.”); see also Sega Enters. Ltd. v. Accolade, Inc., 977 F.2d 1510, 1519 (9th Cir. 1992) (“[I]ntermediate copying of computer object code may infringe the exclusive rights granted to the copyright owner in section 106 of the Copyright Act regardless of whether the end product of the copying also infringes those rights.”). For a discussion of that principle applied to AI-generated art, concluding that creating intermediate reproductions is an act of infringement by an AI, see Jessica L. Gillotte, Note, Copyright Infringement in AI-Generated Artworks, 53 U.C. Davis L. Rev. 2655, 2672–73 (2020). Close It also won’t matter that no one book, song, or image has much value to the ML system in the course of its training. If the company deploying the ML application could identify in advance all the owners of all the works it will use, perhaps it could negotiate licenses with all of them or exclude the works for which it couldn’t obtain a license. But that is impractical. Most photographs, for instance, have no copyright management information attached to them. It is effectively impossible to find that many copyright owners and negotiate that many licenses. That’s why Congress created compulsory licenses for many modern uses, and why private groups like ASCAP and BMI organized to collect and license copyrights in specific sectors. Perhaps eventually similar organizations will arise to license individual works for training datasets. But they don’t exist now, and the fact that the training set needs to be comprehensive means it will be a long time before anyone could effectively aggregate all the rights needed.
Further, the fact that no one work is valuable in the training process won’t prevent copyright lawsuits. Copyright remedies are structured to encourage lawsuits over even small-value infringements. Copyright law awards statutory damages of up to $150,000 per work, plus attorneys’ fees, regardless of the plaintiff’s actual loss or the defendant’s actual gain.858517 U.S.C. § 504(c)(2) (2018). Close True, statutory damages require registration of the work,868617 U.S.C. § 412 (2018) (establishing registration as a prerequisite for such damages). Close and many private citizens won’t have done that. But most copyrighted books and songs and many commercial videos and photographs are registered, so the risk is greater for databases that build on those works. Multiply that by the number of works at issue and the risk to the ML systems of using training datasets featuring potentially infringing works becomes enormous.
Given these incentives, it’s not difficult to foresee some plaintiffs suing opportunistically, just to collect the potential windfall. In addition, we’re increasingly seeing copyright law being floated as a tool to prevent ML access for noneconomic reasons. The continued expansion of ML systems into virtually all walks of social and commercial life has raised a new set of concerns around the protections that copyright affords such technologies.8787See infra Part II. Close With ML systems entering our roadways, courtrooms, and police stations, headlines decrying new applications that push data usage norms to the limits have become regular features of the news.8888See supra notes 10–17 and accompanying text. Close With few clear legal mechanisms available to potential plaintiffs, the question of whether copyright might be a useful tool for curbing the proliferation of such abuses has come to center stage. Today, copyright’s future regarding ML systems remains far from certain. But what is certain is that, as ML’s capabilities grow with time, the temptation to transform copyright into a tool to restrict them will too.
II. Fair Learning
In Part I, we saw that ML requires access to unprecedented amounts of information. Most, if not all, of that information will be copyrighted. And those copyrights are likely to be owned by thousands or even millions of different parties. At least in theory, that creates a problem for training ML systems. We explain why it hasn’t yet served as a roadblock for ML progress in subpart A. In subpart B, we explore some of the new legal challenges that arise when robots attempt to learn from, rather than merely search, copyrighted material. In subpart C, we explore how current fair use doctrine is coming under pressure in the context of ML. And, finally, in subpart D we propose a standard that courts and technologies can use to get out of the hole that recent copyright precedent and increasingly strident public opinion have dug for them.
A. Copyright for Robotic Readers
With all the copyright pitfalls awaiting anyone attempting to create an ML training set, one might wonder why copyright law hasn’t effectively halted the gears of progress in the field. The oversimplified answer needs only two words: “fair use.”898917 U.S.C. § 107 (2018). Close The fair use doctrine permits certain uses of copyrighted works for valuable social purposes,9090See, e.g., Abraham Drassinower, What’s Wrong with Copying? 78 (2015) (“[T]he defense is not about undoing or overlooking a wrong for reasons extraneous to authorship itself. . . . It is as if, upon hearing the plaintiff’s complaint, the defendant were to say: ‘. . . I am equally an author.’”). Close particularly when such uses “transform” the original source material9191Authors Guild v. Google, Inc., 804 F.3d 202, 214–15 (2d. Cir. 2015); see also Asay, supra note 8 (describing the trend). Close and do not threaten the copyright owner’s core market.929217 U.S.C. § 107(4) (2018). Close In Judge Pierre Leval’s famous articulation:
The use must be productive and must employ the quoted matter in a different manner or for a different purpose from the original. . . . If . . . the secondary use adds value to the original – if the quoted matter is used as raw material, transformed in the creation of new information, new aesthetics, new insights and understandings – this is the very type of activity that the fair use doctrine intends to protect for the enrichment of society.9393Pierre N. Leval, Toward a Fair Use Standard, 103 Harv. L. Rev. 1105, 1111 (1990) (emphasis added) (footnote omitted). Fair use, of course, didn’t originate with Judge Leval. See, e.g., Sag, supra note 43, at 307 (noting that “[f]air use and its historical antecedents have been part of copyright law since very shortly after the enactment of the first copyright act, the English Statute of Anne in 1710”). Close
Since Judge Leval penned those words more than three decades ago, his description of transformative fair use has, itself, transformed—growing to encompass not just a host of human-authored changes to the copyrighted works (e.g., parodies, critiques, fan fiction, and the like) but robotic ones as well.9494Clark D. Asay, Arielle Sloan & Dean Sobczak, Is Transformative Use Eating the World?, 61 B.C. L. Rev. 905, 91 1 (2020). Close Among the most significant moments in this trajectory was Sega Enterprises Ltd. v. Accolade, Inc.9595977 F.2d 1510 (9th Cir. 1992). Close The controversy began when the video game publisher, Accolade, decided it wanted its games to run on Sega’s wildly popular Genesis console. To do so, Accolade needed its software to mirror the communication protocols Sega used to interact with the console. And the only way to unlock those protocols was by copying large parts of Sega’s software verbatim to reverse engineer it. When Sega caught wind of Accolade’s efforts, it brought suit—arguing that the conduct violated its copyright in the video game software.9696Because the question of whether such protocols were protected by trade secrets had already been resolved, e.g., Acuson Corp. v. Aloka Co., Ltd., 257 Cal. Rptr. 368 (Cal. Ct. App. 6th) (depublished), the case instead centered on whether copyright’s fair use doctrine protected it. Sega, 977 F.2d at 1513–14. Close Accolade countered that its direct copying was necessary as an intermediate step toward accessing the unprotectable “ideas and functional elements” hidden in Sega’s object code.9797Sega, 977 F.2d at 1527–28. Close
Given that Accolade had copied significant amounts of Sega’s work without commenting on, or modifying, the original expression, the case wasn’t exactly clear cut (at least, under Judge Leval’s original articulation of transformative use).9898For a more straightforward case under Judge Leval’s articulation, see, for example, Campbell v. Acuff–Rose Music, Inc., 510 U.S. 569, 583 (1994) (holding that a rap version of a popular song was a parody and potentially transformative under fair use analysis). Close But the Ninth Circuit found the use fair, drawing a “distinction between the copying of works in order to make independent creative expression possible and the simple exploitation of another’s creative efforts.”9999Sega, 977 F.2d at 1523. Close Accolade had “copied Sega’s code for a legitimate, essentially non-exploitative purpose” because the act of copying was necessary as an intermediate step to access unprotectable elements of the work.100100Id. at 1522–23. Close The court concluded that so long as copying served simply as a means of accessing ideas, facts, or functionality—and not the original expression of the work—the use was fair game.
Since Sega, the legal system has gone on to clarify and expand the protections afforded to this type of copying, now commonly referred to as “non-expressive”101101See, e.g., Sag, supra note 43, at 301–02 (describing the use as “non-expressive”); Grimmelmann, supra note 18, at 662 (same); Sobel, supra note 31, at 52 (same). Close use. Courts have held in favor of video game “emulators” that copied console firmware in order to access its underlying functionality,102102See, e.g., Sony Comp. Entm’t, Inc. v. Connectix Corp., 203 F.3d 596, 601, 609 (9th Cir. 2000). Close search engine providers that bulk-collected and displayed images in order to make them more readily accessible,103103See Kelly v. Arriba Soft Corp., 336 F.3d 811 (9th Cir. 2003); see also Perfect 10, Inc. v. Amazon.com, Inc., 508 F.3d 1146, 1176–77 (9th Cir. 2007) (“Google’s fair use defense is likely to succeed at trial.”). Close and even plagiarism detection providers that consumed copyrighted materials to use and improve their software.104104A.V. ex rel. Vanderhye v. iParadigms, LLC, 562 F.3d 630, 645 (4th Cir. 2009) (holding that copying a student paper to search it for plagiarism had an entirely different function and purpose than the original work). Close Most famously, courts permitted Google to scan all the world’s books into its internal database as an intermediate step toward producing a book search system that, then, delivered verbatim “snippets” of copyright text to the reader.105105Authors Guild v. Google, Inc., 804 F.3d 202, 229 (2d Cir. 2015); Authors Guild v. HathiTrust, 755 F.3d 87, 90, 105 (2d Cir. 2014). Close (More on this below.)
Non-expressive use protections like these (along with the Digital Millennium Copyright Act)10610617 U.S.C. § 512 (2018). That statute allows sites like YouTube to host (and therefore to index and search) video content uploaded by third parties, so long as they take down infringing material when the copyright owner notifies them of infringement. Without that statutory exception, video sites like YouTube would not exist. Close are the reason most automated search and analysis tools exist in the first place. Without such protections, Google and others wouldn’t be able to copy the large bodies of text and images necessary to make them searchable in the first place.107107See Google, 804 F.3d at 215, 225–26 (noting that Google copies the entirety of a book into its database even though it only displays snippets to the public). For an argument that text data mining should be fair use even if the TDM system doesn’t have lawful access to the text, see Carroll, supra note 58, at 895. For an argument that verbatim copying necessary to create a database should be fair use so long as the output of that database uses only a limited amount of the content, see Edward Lee, Technological Fair Use, 83 S. Cal. L. Rev. 797, 846 (2010). Close But even this favorable line of precedent doesn’t mean that Google or any other technology company has free rein with data it has collected. It is still subject to a fact-specific four-factor test.108108See Barton Beebe, An Empirical Study of U.S. Copyright Fair Use Opinions, 1978–2005, 156 U. Pa. L. Rev. 549, 551–52 (2008) (arguing that the fair use factors “form the core of our fair use doctrine and functionally define what fair use is”); David Nimmer, “Fairest of Them All” and Other Fairy Tales of Fair Use, Law & Contemp. Probs., Summer 2003, at 263, 263–87 (evaluating how numerous cases apply the four factors); David Nimmer, Juries and the Development of Fair Use Standards, 31 Harv. J. L. & Tech. 563, 588–89 (2018) (describing the difficulty with juries applying the fair use factors). Close And, like any fact-specific test, it’s unpredictable. In fact, it is so uncertain that Larry Lessig famously described fair use as nothing more than the right to hire a lawyer.109109Lawrence Lessig, Free Culture 187 (2004). Close
The good news is that most companies deploying ML systems can afford to hire lawyers. But while precedent like Kelly v. Arriba Soft Corp. and Authors Guild v. Google may look promising for those who collect and use datasets,110110Grimmelmann goes so far as to say that courts ignore bulk collection by robots, but that’s too strong. See Grimmelmann, supra note 18, at 674 (stating, “Copyright ignores robots”). Rather, courts focus on what is done with the material after it is collected to influence their assessment of the legality of collecting it. Close there is no guarantee that courts will extend this precedent to similar technologies or legal contexts. Indeed, as we’ll see in the subpart that follows, both the courts and the court of public opinion have begun to depart from the precedents established by the Google Books cases. And these departures could have lasting ramifications for the use of copyrighted data to train ML systems.
B. Copyright for Robot Learners
There is reason to worry that courts won’t find AI learning to be fair. Courts applying existing law to ML systems might plausibly conclude that several of the fair use factors weigh against fair use of individual works in training datasets.111111See Sobel, supra note 31, at 48–49, 67 (explaining how “[c]opyright law forces artificial intelligence into a binary: it is either a mystical author or a dumb machine,” but ML is neither, which can present issues for the application of the fair use doctrine). Our focus here is on the copyright claims by individual plaintiffs rather than a claim involving wholesale copying of a copyrighted database itself. We think the latter is (and should be) much less likely to qualify as fair use. Close First, much like TDM systems, ML systems involve copying the entire work without alteration. That directly affects statutory factor number three, which weighs the fact that the entire work is taken against a finding of fair use.112112Indeed, some judges have said that they would declare wholesale copying to be illegal in all circumstances. See Am. Geophysical Union v. Texaco, Inc., 60 F.3d 913, 917 (2d Cir. 1994) (“[I]f the issue were open, we would seriously question whether the fair use analysis that has developed with respect to works of authorship alleged to use portions of copyrighted material is precisely applicable to copies produced by mechanical means.”). Fortunately, that is not the law. Close But this act of direct copying also affects whether the use is transformative. In Kelly, for instance, the court focused not on the reduced resolution of thumbnail images but on the fact that thumbnail images couldn’t substitute for full-size images (at least on 2002-era devices) and served a very different purpose.113113Kelly v. Arriba Soft Corp., 336 F.3d 811 (9th Cir. 2003); Perfect 10, Inc. v. Amazon.com, Inc., 508 F.3d 1146, 1176 (9th Cir. 2007) (“Google’s fair use defense is likely to succeed . . . .”). Gillotte argues that AI training data sets are “highly transformative” because the purpose of the use is different—to train an AI. Gillotte, supra note 84, at 2684. We discuss that different purpose in more detail infra notes 146–75 and accompanying text. Close
The closest analogies to the type of direct copying involved in the creation of a training set may seem to be the intermediate copying software cases running from Sega to Google Books.114114As discussed in a series of cases, the Second and Ninth Circuits have held that it is fair use to copy the entire body of a work that includes copyrightable expression as an intermediate step towards creating an end product that does not include that copyrightable expression. The Ninth Circuit cases involve reverse engineering computer programs to produce a different program that is interoperable with the original. That process involves creating an internal working version of the plaintiff’s code and then writing new code that correctly interfaces with it. And in Authors Guild v. Google, Inc., the Second Circuit held that Google could scan the text of books in order to produce a search engine that could find text in those books and display short snippets of text in response. 804 F.3d 202, 229 (2d Cir. 2015). Close Notably, however, some of these cases have depended heavily on the fact that the defendant’s end product was a transformative new work and the copying was a necessary step to get there. Muddying the fair use question further is the fact that several cases—including Associated Press v. Meltwater U.S. Holdings115115931 F. Supp. 2d 537 (S.D.N.Y. 2013). Close and Fox News v. TVEyes116116883 F.3d 169 (2d Cir. 2018). Close—have rejected fair use arguments in somewhat analogous contexts.117117TVEyes, 883 F.3d at 180–82; Meltwater, 931 F. Supp. 2d at 550, 561. But see Sag, supra note 44, at 335 (arguing that the TVEyes decision does not necessarily suggest that “the tide will turn against TDM and similar non-expressive uses in American copyright law”). We think those cases can be distinguished on their facts—the output of those systems involved sizeable chunks of the original copyrighted work. But it’s not clear that they will be distinguished. Close TVEyes in particular rejected the district court’s finding that a TV news clipping service that analyzed and made TV news searchable was fair use.118118TVEyes, 883 F.3d at 180–82. Close And the Supreme Court is currently considering its first case involving transformative use since it adopted the doctrine in 1994, so the scope and even the continued existence of the doctrine are up in the air.119119Google LLC v. Oracle Am., Inc., No.18-956 (U.S. argued Oct. 7, 2020). Close
TVEyes might be distinguished on its facts,120120Part of what is driving that case is the fact that the defendants ingested the entirety of Fox News in order to create a searchable database of things said there. TVEyes, 883 F.3d at 175. An AI wouldn’t necessarily have to do that in order to learn. But it might well want to if it is to make better decisions and avoid discrimination. Close and it’s possible that precedents like the Google Books cases will still be extended to ML training in straightforward fashion. But enterprising plaintiffs’ lawyers can point to a number of facts that might distinguish those cases. The fact that in most cases the ML systems aren’t producing a new copyrightable work at all, but are just consuming the plaintiff’s work for profit to improve uncopyrighted systems like self-driving cars or speech recognition technology, might persuade a court that those precedents aren’t all that helpful.121121But see Grimmelmann, supra note 18, at 657–58 (arguing, before TVEyes, that courts ignore automatic copying by robots). Grimmelmann’s focus was on older TDM systems, not the modern ML systems. Close Maybe that won’t matter. Like search engines, ML systems use the plaintiff’s work only inside the computer, and it is not shared with the public. But it is also possible that the fact that the work is sometimes not communicated to the outside world in any form may hurt rather than help the claim to transformation. Copyright owners may argue that, unlike the search engine cases, some ML systems lack a creative “end product” provided to consumers that is different in nature or purpose than the original work. Rather, the argument goes, ML is a consumer of the copyrighted work, outputting a profitable technology rather than new creativity, and these for-profit consumers should pay for what they take.
That argument may appeal to the very strong, if often unarticulated, anti-free riding instinct in courts.122122See Mark A. Lemley, Property, Intellectual Property, and Free Riding, 83 Texas L. Rev. 1031, 1031–33 (2005) (discussing how courts’ implicit association of intellectual property with real property leads to a condemnation of free riding); Mark A. Lemley & Mark P. McKenna, Unfair Disruption, 100 B.U. L. Rev. 71, 76–77 (2020) (addressing the “question of when competition by market disruption is ‘unfair’ in a way the law should forbid”); see also Wendy J. Gordon, The Core of Copyright: Authors, Not Publishers, 52 Hous. L. Rev. 613, 624 (2014) (arguing that the demonization of free riding has endangered a community interest in free access, which in turn has negatively impacted our culture and our business). Close And the appeal of that argument is likely to be strengthened by the commercial nature of many ML applications. While commercial uses are not presumptively unfair,123123See Campbell v. Acuff–Rose Music, Inc., 510 U.S. 569, 584, 594 (1994) (reversing precedent to that effect); Authors Guild v. Google, Inc., 804 F.3d 202, 219 (2d Cir. 2015) (ruling there is “no reason . . . why [a defendant’s] overall profit motivation should prevail as a reason for denying fair use over its highly convincing transformative purpose, together with the absence of significant substitutive competition, as reasons for granting fair use”). Close they still tend to weigh against a finding of fair use.124124See, e.g., Sony Corp. of Am. v. Universal City Studios, Inc., 464 U.S. 417, 451 (1984) (stating that “every commercial use of copyrighted material is presumptively an unfair exploitation,” a rule since reversed by Campbell). Close And commerciality often goes hand in hand with a market effect. Here, the system’s use doesn’t cut into the ordinary market for the copyrighted works in question. But ML companies might be natural candidates for a licensing market: large for-profit companies that stand to benefit financially from using copyrighted works (albeit in bulk, rather than this work in particular).125125Sobel, supra note 31, at 49 (discussing this possibility). Close Given the well-known circularity of the claim to a loss of a licensing opportunity—the use is unfair if there is a lost licensing opportunity, but there is only a lost licensing opportunity if the use is unfair126126See Mark A. Lemley, Should a Licensing Market Require Licensing?, 70 L. & Contemp. Probs., Spring 2007, at 185, 190 (noting this problem); see also Christina Bohannan, Reclaiming Copyright, 23 Cardozo Arts & Ent. L.J. 567, 597–98 (2006) (voicing concerns about the burden on a defendant of proving “the non-existence of market harm”); William W. Fisher III, Reconstructing the Fair Use Doctrine, 101 Harv. L. Rev. 1659, 1671 (1988) (“[I]n almost every case in which the fair use doctrine is invoked, there will be some material adverse impact on a ‘potential market.’”); Pierre N. Leval, Toward a Fair Use Standard, 103 Harv. L. Rev. 1105, 1124 (1990) (“By definition, every fair use involves some loss of royalty revenue because the secondary user has not paid royalties.”); Lydia Pallas Loren, Redefining the Market Failure Approach to Fair Use in an Era of Copyright Permission Systems, 5 J. Intell. Prop. L. 1, 38–39 (1997) (criticizing this circularity because it in turn “trivializes the importance of fair use”); Matthew Sag, God in the Machine: A New Structural Analysis of Copyright’s Fair Use Doctrine, 11 Mich. Telecomm. & Tech. L. Rev. 381, 393–94 (2005) (comparing two photocopying cases to illustrate this problem); Sara K. Stadler, Forging a Truly Utilitarian Copyright, 91 Iowa L. Rev. 609, 656–58 (2006) (explaining that problems arise when courts try to determine which markets belong exclusively to copyright owners). Close—courts may well let their view of the equities creep into the analysis of the fourth factor.127127Wendy Gordon famously wrote that fair use worked as a substitute where transactions costs made a market infeasible. Wendy J Gordon, Fair Use as Market Failure: A Structural and Economic Analysis of the Betamax Case and Its Predecessors, 82 Colum. L. Rev. 1600, 1601–02 (1982). While she has later criticized unwarranted extensions of this work to create spurious markets, Wendy J. Gordon, Excuse and Justification in the Law of Fair Use: Transaction Costs Have Always Been Part of the Story, 50 J. Copyright Soc’y USA 149, 150–51 (2003), plaintiffs have seized on the transactions costs story to create markets for licensing uses that would otherwise be fair. See Lemley, supra note 127, at 190–91 (noting this problem). Jane Ginsburg has argued that the statute makes a use unfair if it cuts into the “value” of the work even if it doesn’t interfere with an established market. Jane C. Ginsburg, Fair Use Factor Four Revisited: Valuing the “Value of the Copyrighted Work”, 67 J. Copyright Soc’y USA 19 (2020) (manuscript at 11, 15), https://scholarship.law.columbia.edu/faculty_scholarship/2677. But that too could be circular. A work would be more valuable if ML systems would pay for its use. Close
A second challenge for those seeking fair use protections arises in situations where ML systems do, in fact, produce tangible outputs that could trigger a similar strain of anti-free riding sentiment. Particularly in the last several years, we’ve seen major strides involving ML systems capable of replicating the outputs of creative professionals. MuseNet, discussed in the introduction, is one such example. But ML applications have shown similarly impressive results in fields as diverse as journalism,128128See John Seabrook, Can a Machine Learn to Write for The New Yorker?, New Yorker (Oct. 14, 2019), https://www.newyorker.com/magazine/2019/10/14/can-a-machine-learn-to-write-for-the-new-yorker [https://perma.cc/T7UP-N2SH] (describing how, with further development, automated writers like GPT-2 could take over many people’s writing responsibilities). Close poetry,129129See, e.g., Kelsey Piper, A Poetry-Writing AI Has Just Been Unveiled. It’s . . . Pretty Good, Vox (May 15, 2019, 3:08 PM), https://www.vox.com/2019/5/15/18623134/openai-language-ai-gpt2-poetry-try-it [https://perma.cc/D9KY-D7B4]. Close painting,130130See Murphy, supra note 62 (discussing the machine learning system that paints by mimicking “the way a brain finds patterns in objects”); see also Cohn, supra note 62 (illustrating the success of the “Generative Adversarial Networks, or GANs” technology that produced the portrait). Close and photography.131131See Samantha Cole, This Company Promises to Place Any Face onto Any Body, Using an Algorithm, Vice (Nov. 21, 2019, 10:50 AM), https://www.vice.com/en_us/article/7x5nv4/rosebud-ai-stock-images-using-an-algorithm [https://perma.cc/PJ47-T8Q6] (describing a generative system that produces high quality stock photos that are on par with a professional photographer). Close These ML systems––virtually all of which are trained on copyrighted works––have produced writing that’s difficult to distinguish from real journalists,132132See, e.g., Seabrook, supra note 128 (describing GPT-2’s linguistic prowess). Close painted in the style of celebrated masters,133133See Murphy, supra note 62 (describing technology that mimics techniques of famous painters to turn photographs into paintings); see also Cohn, supra note 62 (discussing the high-price sale of a portrait created by artificial intelligence). Close and even created stock photos comparable to those of professional photographers.134134Cole, supra note 131. Close Many of these efforts have been so convincing that professionals and opinion columnists alike have begun to openly worry about artificial intelligence as a genuine competitive threat. These concerns, in turn, have triggered criticisms from thought leaders, advocates, academics, and professionals worried that technology companies producing these technologies may be free riding on the labor of creative professionals. Critics of such practices have compared leading ML companies to “robber barons” siphoning up valuable IP.135135Andrew Orlowski, Jaron Lanier: Big Tech Is Worse than Big Oil, Register (Apr. 22, 2016), https://www.theregister.co.uk/2016/04/22/jaron_lanier_on_ip/ [https://perma.cc/FQP3-V863]; Solon, supra note 25 (quoting one critic describing large-scale data collection as “the money laundering of facial recognition [where companies are] laundering the IP and privacy rights out of the faces”). Close Others have vocalized concerns that ML “empowers these companies to extract value from authors’ protected expression without authorization, and to use that value for commercial purposes that may someday jeopardize the livelihoods of human creators.”136136See generally Sobel, supra note 31. Sobel argues that “[c]onstruing fair use to protect this activity will place the doctrine at odds with the public interest and potentially exacerbate the social inequalities that AI threatens. . . . [but at] the same time, finding that expressive machine learning is not fair use would frustrate the progress of the promising technology.” Id. at 97. He views this inherent tension as a dilemma for fair use doctrine in the future. See id.; see, e.g., supra notes 30–31 and accompanying text. In a separate paper, Sobel notes that courts are unlikely to find fair use when an AI creates a new work in the style of an existing artist. Benjamin L.W. Sobel, Elements of Style: Emerging Technologies and Copyright’s Fickle Similarity Standards 57–58 (Aug. 2019) (unpublished manuscript) (on file with author). Close It is not at all clear these practices are reflective of the kind of exploitation of original expression that copyright law is meant to guard against.137137See, e.g., Sony Comp. Ent., Inc. v. Connectix Corp., 203 F.3d 596, 602–03 (9th Cir. 2000) (noting there is a distinction between the copying of works in order to make independent creative expression possible and the simple exploitation of another’s creative efforts). If a new artist writes a different and better song than you because she learned from you, the law celebrates that creativity. It is not obvious that it should feel differently because it is a machine rather than a new artist that wrote the better song. Close But what is clear is that this emerging view of the equities, too, could have consequences for how courts consider the competitive and substitutive implications of a permissive fair use doctrine.138138See infra subpart II(D). Close
A third challenge facing those advancing fair use arguments will be the host of other potential negative impacts that the technology can have on downstream users or consumers. As ML systems have been handed greater decision-making authority over our social, economic, and political lives, they’ve also come under increasing fire by critics who fear the prospect of ML systems replacing human decision makers to negative effect. Today, media outlets, thought leaders, and advocacy organizations decrying new ML systems that push data usage norms to the limits have become a regular feature of the news.139139See supra note 24. Close Critics variously worry that ML systems with free rein to consume copyrighted materials could spread propaganda,140140See, e.g., Alex Hern, New AI Fake Text Generator May Be Too Dangerous to Release, Say Creators, Guardian (Feb. 14, 2019), https://www.theguardian.com/technology/2019/feb/14/elon-musk-backed-ai-writes-convincing-news-fiction [https://perma.cc/SW37-ZUHH] (discussing how GPT2 could be used to create and spread spam, fake news, bigoted text, and conspiracy theories); @BuzzFeed, Twitter (Apr. 17, 2018, 10:00 AM), https://twitter.com/BuzzFeed/status/986257991799222272 [https://perma.cc/C38K-B377] (“We’re entering an era in which our enemies can make anyone say anything at any point in time.”); Tim Mak, Technologies to Create Fake Audio and Video Are Quickly Evolving, NPR (Apr. 2, 2018, 4:20 PM), https://www.npr.org/2018/04/02/598916380/technologies-to-create-fakeaudio-and-video-are-quickly-evolving [https://perma.cc/5GJV-HDDW] (discussing ML systems’ ability to generate videos created for misinformation campaigns). Close facilitate dystopian surveillance,141141See, e.g., Josh Kaplan, License Plate Readers Are Creeping into Neighborhoods Across the Country, Slate (July 10, 2019, 7:30 AM), https://slate.com/technology/2019/07/automatic-license-plate-readers-hoa-police-openalpr.html [https://perma.cc/6R84-CA72] (documenting instances of automated license plate readers being used as a surveillance tool by law enforcement and landlords alike); Cade Metz, Facial Recognition Tech Is Growing Stronger, Thanks to Your Face, N.Y. Times (July 13, 2019), https://www.nytimes.com/2019/07/13/technology/databases-faces-facial-recognition-technology.html [https://perma.cc/26PC-M6WV] (citing instances of facial recognition systems tools trained on photos used by Immigration and Customs Enforcement to identify undocumented immigrants and by Chinese government agencies to engage in “ethnic profiling of the country’s minority Uighur Muslims,” among other instances). Close invade personal and sexual privacy,142142See, e.g., David Greene, We Don’t Need New Laws for Faked Videos, We Already Have Them, Elec. Frontier Found. (Feb. 13, 2018), https://www.eff.org/deeplinks/2018/02/we-dont-need-new-laws-faked-videos-we-already-have-them [https://perma.cc/DA2Q-FLTZ] (referencing pornographic videos made by machine learning technology that splices one person’s face onto another person’s body without the consent of the parties). Close perpetuate bias,143143See, e.g., Natasha Singer & Cade Metz, Many Facial-Recognition Systems Are Biased, Says U.S. Study, N.Y. Times (Dec. 19, 2019), https://www.nytimes.com/2019/12/19/technology/facial-recognition-bias.html [https://perma.cc/WKM2-WMDP] (describing studies showing racial and other bias in facial recognition tools); Snow, supra note 77 (experimenting with facial recognition technology on headshots of Congressmembers and finding that “false matches were disproportionately of people of color”). Close and even threaten to take over the world.144144See, e.g., Grimmelmann, supra note 18, at 676–78 (concluding that a permissive fair use standard “arguably increases the chances that humanity will meet a sudden, violent, and extremely unpleasant end” at the hands of super intelligent machines). Close As these systems grow to play even greater roles in the most intimate aspects of our lives, these concerns will almost certainly grow too. These concerns don’t relate directly to whether a use is fair (though the public interest purpose of a use is one factor the courts consider), but they may incline parties and courts not to give ML systems the benefit of the doubt. That is particularly true when the ML systems are in the hands of established tech giants like Google or Amazon that are already the target of enormous public ire.145145For discussion of the backlash against big tech companies, see, for example, Mark A. Lemley & Andrew McCreary, Exit Strategy, 101 B.U. L. Rev. (forthcoming 2021) (manuscript at 1, 51–55), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3506919. Close
Finally, the sheer number of works ML systems must copy means that taking a chance on fair use is a risky move for an ML company. Copyright statutory damages systematically overcompensate plaintiffs with small-value works, offering them up to $150,000 per work regardless of the value the user places on using that particular work.14614617 U.S.C. § 504(c) (2018). Close An ML that copies millions of works could potentially face hundreds of billions of dollars in statutory damages. And with thousands or even hundreds of thousands of different copyright owners, the risk of multiple opportunistic suits is high. Many ML companies will not share Google’s willingness to bet the company on the legal principle of fair use.
We want to be clear: We aren’t saying courts, or the court of public opinion, will definitely reject the fair use defense as it is currently understood, only that there is a risk they will do so.147147Outside the United States things are much more varied. Consistent with their more high-protectionist view of copyright, especially as applied to technology, European Union courts are much more likely to hold international companies deploying ML systems liable for copyright infringement. See, e.g., Case C-5/08, Infopaq Int’l A/S v. Danske Dagblades Forening, 2009 E.C.R. I-6569 (holding that an eleven-word snippet displayed in a context similar to Google’s search might infringe copyright). In Asia, by contrast, countries are friendlier towards ML. Japan recently adopted a law permitting ML training as an exception to copyright. See, e.g., Japan Amends Its Copyright Legislation to Meet Future Demands in AI and Big Data, Eur. All. for Res. Excellence (Sept. 3, 2018), http://eare.eu/japan-amends-tdm-exception-copyright/ [https://perma.cc/SMQ7-RFTE] (describing Japan’s adoption of legislation that explicitly allows copying by ML engineers collecting data for training sets). The European Union permits TDM in some circumstances. Carroll, supra note 58, at 895–96. And Singapore is considering a similar law. See, e.g., Christian Troncoso, Copyright Proposal Threatens to Undermine Europe’s AI Ambitions, BSA TechPost (Sept. 5, 2018), https://techpost.bsa.org/2018/09/05/copyright-proposal-threatens-to-undermine-europes-ai-ambitions/ [https://perma.cc/87HX-G4YT] (noting that Singapore courts and legislatures are grappling with the issue of fair use by ML systems). Close And we certainly aren’t saying they should reject the defense. To the contrary, there are very strong policy reasons to allow ML systems to copy individual works in the course of training algorithms, as we explain in the following section.
C. Just the Facts, Ma’am
There are several reasons why the fair use doctrine should permit ML systems to train on data sets that include copyrighted works.
First, society benefits from allowing ML systems to compile the best possible databases and to open them for public scrutiny and for open AI. Broad access to training datasets will make AI better, safer, and fairer.148148See supra note 31 and accompanying text. Close Smaller, proprietary datasets—particularly those with large and nonrandom gaps due to failures of copyright licensing—will lead to worse decisions by ML systems. And those worse decisions have real-world consequences. They may mean the difference between a self-driving vehicle that stops at a stop sign at night in the rain and one that doesn’t. Or the difference between a biometric scanner at airports that accurately identifies women of color and one that doesn’t. Or between dictation software that faithfully transcribes what you said and dictation software that lands you in the hall of fame at Autocorrect Fail.149149E.g., Auto Correct Fail, www.autocorrectfail.org [https://perma.cc/H39F-ZCBT]. The original, damnyouautocorrect.com, appears to be defunct, alas. Close
Second, given the large number of works an AI training data set needs to use and the fact that thousands, if not millions, of different people own those works, AI companies can’t simply license all the underlying photographs or texts for the new use. So allowing a copyright claim is tantamount to saying, not that copyright owners will get paid, but that no one will get the benefit of this new use because it will be impractical to make that use at all.150150Cf. Keller v. Elec. Arts, Inc., 724 F.3d 1268, 1269 (9th Cir. 2013). Close That is particularly ironic if the copyright claim is justified by the supposed existence of a licensing market, since the very theory of the licensing market will have blocked the creation of such a market.151151For an argument that any plaintiff who relies on a licensing market argument to defeat a fair use claim should have to forego any claim to injunctive relief, see Lemley, supra note 127. Cf. Alex Kozinski & Christopher Newman, What’s So Fair About Fair Use?, 46 J. Copyright Soc’y USA 513, 525 (1999) (arguing that injunctions should be denied unless “there is strong reason that damages will be inadequate”). But eliminating injunctive relief would still leave a punitive statutory damages regime in place. Close Or (and this may be just as bad), only companies like Google or Facebook that happened to collect the data for other, permissible purposes will be able to compete in AI space.152152We are skeptical that browsewrap terms of service giving companies like Google or Facebook plenary authority over what they do with your data are enforceable. Lemley, supra note 75, at 464. But if they are, they would create a blanket license for those companies to make different uses of the copyrighted material posted to their sites. Close
A third, and less obvious reason, is that providing ML systems with broader access to data actually helps to mitigate some of the very negative outcomes that critics of ML systems fear. As the Obama White House recently identified, “AI needs good data. If the data is incomplete or biased, AI can exacerbate problems of bias.”153153Exec. Office of the President, Preparing for the Future of Artificial Intelligence 30 (2016), https://obamawhitehouse.archives.gov/sites/default/files/whitehouse
_files/microsites/ostp/NSTC/preparing_for_the_future_of_ai.pdf [https://perma.cc/2X74-QBZF]. Close Facial recognition provides a good example. Facial recognition software performs worse at distinguishing individuals in small racial groups because since those groups are small, it has fewer unique data points allowing it to draw fine distinctions between faces in those groups. The solution is to build bigger databases overall or to “oversample” members of smaller groups. Ironically, this was exactly the motivation at play when two facial recognition tools––one by IBM and the other by MegaFace––came under fire in media outlets across the globe.154154See, e.g., Hill & Krolik, supra note 22 (noting, “Millions of Flickr images were sucked into a database called MegaFace. Now some of those faces may have the ability to sue.”); Madhumita Murgia, Who’s Using Your Face? The Ugly Truth About Facial Recognition, Fin. Times (Sept. 18, 2019), https://www.ft.com/content/cf19b956-60a2-11e9-b285-3acd5d43599e [https://perma.cc/9CM8-9MF9] (articulating various concerns with facial recognition technology); Solon, supra note 25 (describing how facial recognition technology could target minority groups). Close On both such occasions, the engineers involved collected millions of images from users that had released their photos under the creative commons license, which allowed for their bulk collection.155155Ryan Merkley, Use and Fair Use: Statement on Shared Images in Facial Recognition AI, Creative Commons (Mar. 13, 2019), https://creativecommons.org/2019/03/13/statement-on-shared-images-in-facial-recognition-ai/ [https://perma.cc/ZJ94-QMLS]; Frequently Asked Questions: Artificial Intelligence and CC Licenses, Creative Commons (2020), https://creativecommons.org/faq/#artificial-intelligence-and-cc-licenses [https://perma.cc/A2HS-C95V]. Close Both had resorted to collecting images outside of their internal datasets in hopes that exposure to a more diverse set of photos would help to increase the accuracy and reduce the potential for bias in their systems.156156To draw accurate distinctions among faces, facial recognition software needs to train on lots of similar faces. Because large racial groups (or groups with heavy representation in a photo dataset) have more members than small ones, facial recognition software frequently performs less well in distinguishing members of racial minorities. Levendowski, supra note 31, at 581–85 (noting how incomplete or biased data can exacerbate existing biases). Close Yet, instead of being lauded, their efforts actually drew a swift rebuke from the media.157157The bias objection and the privacy objection are at odds here. Many people worry about facial recognition because it will make mistakes. Correcting bias and getting better facial recognition helps alleviate that worry. But others may worry about facial recognition precisely because it doesn’t make mistakes. Close
Many such articles deriding IBM and MegaFace also discussed IP rights as a potential mechanism to rein in these uses.158158See supra note 154 and accompanying text. Close And though we’re sympathetic to those fearful of the potential of ML systems to perpetuate bias, enable both private and public sector overreach, and run roughshod over some of society’s most vulnerable citizens, trying to solve these problems by simply restricting access to more data is not a viable solution. As Amanda Levendowski convincingly argues, a permissive interpretation of fair use is “quite literally, [necessary for] promoting fairer AI.”159159Levendowski, supra note 31, at 590. Close In her telling, “The normative values embedded in the tradition of fair use align ultimately with the goal of mitigating bias. Fair use can, quite literally, promote creation of fairer AI systems.”160160Id. at 630. Close
Finally, perhaps the strongest argument for fair use is one that lies at the heart of copyright theory but doesn’t actually show up explicitly in the case law: Like TDM technologies, ML systems generally don’t want to copy the copyrighted work for any copyright-related reason. ML systems generally copy works, not to get access to their creative expression (the part of the work the law protects), but to get access to the uncopyrightable parts of the work—the ideas, facts, and linguistic structure of the works. A self-driving car, for instance, doesn’t care about the composition or lighting of your photograph, or indeed about what you were likely actually intending to depict in your photo. It cares about the fact that there’s a stop sign in it. Mapping software doesn’t care what color you chose for your roads and political subdivisions; it wants to know where the roads are, what they are called, and which ones are one-way. Facial recognition software doesn’t care about the composition choices (if any) you made in taking that selfie. It wants to know what you look like in a variety of lights, with and without a hat or a beard, and with different facial expressions. Search engines want to know what’s in your work so they can help people find it, not because they care how you express yourself. Even ML systems that parse text or music usually don’t care about the things that make those works copyrightable. They may be interested in corpus linguistics—how words are used in relationship to each other.161161For discussion of the role of corpus linguistics in legal interpretation, see, for example, Lawrence M. Solan, The New Textualists’ New Text, 38 Loy. L.A. L. Rev. 2027 (2005); Stephen C. Mouritsen, The Dictionary Is Not a Fortress: Definitional Fallacies and a Corpus-Based Approach to Plain Meaning, 2010 BYU L. Rev. 1915 (2010). Close They may be training to understand or create natural language sentences by seeing how grammar is employed in practice. While there are some ML systems that train on art or writing in order to be able to create their own works of art, as discussed above, most are interested in copyrighted works for reasons that have nothing to do with the things that make those works copyrightable.
Ideas, facts, functions, methods, and stock literary and plot devices (aka scènes à faire) are not protectable by copyright law.16216217 U.S.C. § 102(b) (2018); Baker v. Selden, 101 U.S. 99, 105 (1879) (establishing this precedent); Pamela Samuelson, Why Copyright Law Excludes Systems and Processes from the Scope of Its Protection, 85 Texas L. Rev. 1921, 1922 (2007) (discussing the various things § 102(b) refuses to protect). Close The “idea/expression dichotomy” (along with its cousins the fact–expression dichotomy and the process–expression dichotomy)163163See Samuelson, supra note 162 (describing these dichotomies). Close is perhaps the central doctrine in all of copyright law. Indeed, copyright courts have made it clear that copyright law wouldn’t be constitutional if it gave control over ideas and facts.164164See Eldred v. Ashcroft, 537 U.S. 186, 197–98 (2003) (asserting that copyright is constitutional in part because it does not extend to ideas); Feist Publ’ns, Inc. v. Rural Tel. Serv. Co, 499 U.S. 340, 350 (1991) (holding that the constitution prohibits extending copyright to facts); Harper & Row, Publishers, Inc. v. Nation Enters., 471 U.S. 539, 555 (1985) (acknowledging that facts and ideas cannot be copyrighted). Close ML systems aren’t interested in expression (at least not for expression-related reasons); they just want the facts.
For humans, getting and using the unprotectable parts of a copyrighted work is normally not a problem. Read the book, or watch the movie, and you are free to take the ideas—or its standard plot elements—from that work for your own use. If it’s a factual work, you can use and indeed copy the facts in it without infringing the copyright. If it’s a computer program, you can take the functional aspects of that program, even if it means copying some of the code directly.165165Well, except in the Federal Circuit. See Comp. Assocs. Int’l, Inc. v. Altai, Inc., 982 F.2d 693, 703 (2d Cir. 1992); Lotus Dev. Corp. v. Borland Int’l, Inc., 49 F.3d 807, 815 (1st Cir. 1995); Mark A. Lemley, Convergence in the Law of Software Copyright?, 10 Berkeley Tech. L.J. 1, 6 (1995) (discussing the consensus in the circuits on this point); Peter S. Menell, Rise of the Copyright API Dead? An Updated Epitaph for Copyright Protection of Network and Functional Features of Computer Software, 32 Harv. J.L. & Tech. 305, 329 (2018); Pamela Samuelson & Clark D. Asay, Saving Software’s Fair Use Future, 31 Harv. J.L. & Tech. 536, 548–49 (2018). Contra Oracle Am., Inc. v. Google Inc., 750 F.3d 1339, 1359 (Fed. Cir. 2014) (holding such copying illegal), cert. granted, Google LLC v. Oracle Am., Inc., No.18-956 (U.S. argued Oct. 7, 2020). Close You can’t copy the expression, but you don’t need to in order to get access to the unprotectable elements. And once you have that access, you can reuse those elements without fear of liability.
That reflects an important, but rarely articulated, limit on the scope of copyright law. Unlike a patent, which gives its owner control over any “use” of the patented invention, a copyright only controls certain uses: copying, distributing, publicly performing, and the like. Notably absent from that list are certain activities fundamental to learning, such as watching, reading, and discussing a work and communicating its unprotectable elements to others.166166See Jessica Litman, The Exclusive Right to Read, 13 Cardozo Arts & Ent. L.J. 29, 43 (1994) (arguing in favor of permitting the public to read, listen, and view copyrighted works). Close Further, the first sale doctrine permits reselling, loaning, and otherwise sharing a single copy of a work.16716717 U.S.C. § 109(a) (2018). Close These are all “uses” of the copyrighted work, and important ones at that. But they aren’t uses the law forbids, because the point of copyright is not just to give incentives to create but to “promote the progress of science”168168U.S. Const. art. I, § 8, cl. 8. Close by ensuring those creations are shared by others. The freedoms to read, to learn, and to communicate what you have learned are critical to making the idea–expression dichotomy work in practice, because it helps ensure people can find the ideas in a copyrighted work in order to use them.
True, there are some circumstances in the pre-ML world in which giving copyright owners control over expression may risk allowing them to lock up the unprotectable ideas and facts as well. That is particularly likely when the work is predominantly composed of uncopyrightable elements. In those circumstances, courts and Congress have created a variety of doctrines to help ensure the user’s access to the unprotectable parts of the copyrighted work. We impose a higher standard for proof of infringement of factual works whose copyright is “thin,” requiring “virtual identity” of the works rather than merely “substantial similarity.”169169See, e.g., Hoehling v. Universal City Studios, Inc., 618 F.2d 972, 977, 980 (2d Cir. 1980) (using the “virtual identity” test). The Ninth Circuit questioned whether this was a different standard in its en banc decision in Skidmore v. Led Zeppelin, but it upheld the virtual identity requirement. 952 F.3d 1051, 1080 (9th Cir. 2020) (en banc). Close Doing so reduces the risk that a defendant will wrongly be held liable if their work is too similar to the plaintiff’s because it shares the uncopyrightable facts and ideas of the plaintiff’s work. We theoretically give a broader scope of fair use to defendants copying factual works, though some courts deny that doctrine has much if any force.170170Compare Oracle Am., Inc., v. Google Inc., 886 F.3d 1179, 1204–05 (Fed. Cir. 2018), cert. granted, Google LLC v. Oracle Am., Inc., No.18-956 (U.S. argued Oct. 7, 2020), and Fox News Network, LLC v. TVEyes, Inc., 883 F. 3d 169, 178 (2d Cir. 2018) (both treating this factor as insignificant), with Am. Soc’y for Testing & Materials v. Public.Resource.Org, Inc., 896 F.3d 437, 451 (D.C. Cir. 2018) (treating it as central). Close We deny protection to copyrightable expression altogether if the work is so short171171The Copyright Office’s circular, for example, notes that “slogans, and other short phrases or expressions cannot be copyrighted.” Though not having the force of a statute, it is considered to be “a fair summary of the law.” Kitchens of Sara Lee, Inc. v. Nifty Foods Corp., 266 F.2d 541, 544 (2d Cir. 1959); Ets-Hokin v. Skyy Spirits, Inc., 225 F.3d 1068, 1082 (9th Cir. 2000). But see CRA Mktg., Inc. v. Brandow’s Fairway Chrysler-Plymouth-Jeep-Eagle, Inc., No. 98-CV-6485, 1999 U.S. Dist. LEXIS 11889, at *10–12 (E.D. Pa. July 27, 1999) (holding that fifty-four words was enough to constitute copyright infringement); Hall v. Swift, 786 Fed. App’x 711, 711–12 (9th Cir. 2019) (holding that Taylor Swift could infringe copyright in a song with only six similar words: “players gonna play, haters gonna hate”). Close or so bound up with the ideas that there are only a limited number of ways of expressing those ideas. In that case we say the idea and expression have “merged” and refuse to protect the expression for fear of also locking up the ideas.172172See, e.g., Morrissey v. Proctor & Gamble Co., 379 F.2d 675, 678–79 (1st Cir. 1967) (“[W]hen the uncopyrightable subject matter is very narrow . . . to permit copyrighting would mean that a party . . . could exhaust all possibilities of future use of the substance”); Sampson & Murdock Co. v. Seaver-Radford Co., 140 F. 539, 541 (1st Cir. 1905) (holding that a publisher of a directory cannot be prevented from publishing facts that were published by a prior directory); cf. Kaplan, An Unhurried View of Copyright 64–65 (1967). Close We are supposed to do the same for “useful articles” where we can’t separate the creative from the functional aspects,173173See 17 U.S.C. § 101 (2018) (defining “useful article” as “an article having an intrinsic utilitarian function that is not merely to portray the appearance of the article or to convey information”). Close though the Supreme Court recently muddied the waters around that doctrine to such an extent that it’s no longer clear we deny protection in such a case.174174See Star Athletica, L.L.C. v. Varsity Brands, Inc., 137 S. Ct. 1002, 1016 (2017) (holding that an artistic feature of a useful article is copyrightable if the feature “can be perceived as a . . . work of art separate from the useful article” and “would qualify as a protectable pictorial, graphic, or sculptural work . . . if imagined separately from the useful article”). For a small sample of the numerous trenchant criticisms of that case, see Barton Beebe, Star Athletica and the Problem of Panaestheticism, 9 U.C. Irvine L. Rev. 276 (2019); Christopher Buccafusco, Mark A. Lemley
& Jonathan S. Masur, Intelligent Design, 68 Duke L.J. 75, 121–23 (2018); Mark P. McKenna, Knowing Separability When We See It, 166 U. PA. L. Rev. Online 127 (2017); Jennifer Yamin, Interview: Professor Fromer and the Star Athletica Case, N.Y.U. J. Intell. Prop. & Ent. L. Blog (Apr. 5, 2018), https://blog.jipel.law.nyu.edu/2018/04/interview-professor-fromer-and-the-star-athletica-case/ [https://perma.cc/XQD6-4L2G]. Close
All these doctrines give less, or in some cases no, protection to a copyrighted work because the ideas and facts are so closely bound up with the expression in that work that it is hard if not impossible for users to take the unprotectable bits without treading on the protectable ones. They focus, in other words, on the centrality of ideas to the plaintiff’s work and the difficulty in separating the protectable from the unprotectable.175175See Mark A. Lemley & Philip J. Weiser, Should Property or Liability Rules Govern Information?, 85 Texas L. Rev. 783, 794 (2007) (noting the difficulty in determining “whether a particular aspect of the work is an expression entitled to protection or part of the unprotectable idea”). Close And they limit protection or deny it entirely to works that are primarily composed of uncopyrightable matter.
But that won’t help our robotic learners. They aren’t taking works that are particularly factual or functional in nature. Some might be—protection for photographs is or ought to be thin, for instance176176See generally Jessica Silbey, Justifying Copyright in the Age of Digital Reproduction: The Case of Photographers, 9 U.C. Irvine L. Rev. 405 (2019) (exploring the justification for copyrighting photographs); Shyamkrishna Balganesh, Causing Copyright, 117 Colum. L. Rev. 1 (2017) (arguing that “authorial causation” should be required for copyright). Close—but some will be ordinary copyrighted works. The problem ML systems face is the inability to capture the unprotectable parts to use for training without making a rote copy of the protectable ones. Systems want access to the unprotectable bits of creative works, but the way they get that access is necessarily copying the whole thing.177177See Grimmelmann, supra note 18, 661–68 (describing the technical architecture that makes copying mandatory). Close Unlike humans, they can’t read to learn or observe the idea in a painting or song without making a copy of the whole thing in their training data set.
D. Fair Machine Learning
Above, we saw a host of potential legal challenges await those seeking to build training sets for ML systems and that there are compelling reasons for why policymakers should be concerned. In this subpart, we suggest that the analysis of fair use for AI training data should incorporate a principle we call “fair learning.” If the purpose of the AI’s use is not to obtain or incorporate the copyrightable elements of a work but to access, learn, and use the unprotectable parts of the work, that use should be presumptively fair under the first fair use factor (the purpose of the use). Notably, fair learning by ML systems should be fair even if fair use factors two and three (the nature of the work and the amount taken) would otherwise weigh against fair use. Systems should be able to learn from fictional as well as factual works,178178The second factor also gives special protection to unpublished works in order to preserve the plaintiff’s right of first publication. Harper & Row, Publishers, Inc. v. Nation Enters., 471 U.S. 539, 561–64 (1985). But that shouldn’t apply here because the AI isn’t publishing the plaintiff’s work at all and therefore isn’t preempting the right of first publication. Some plaintiffs use copyright to hide facts from the world. Salinger v. Random House, Inc., 811 F.2d 90, 100 (2d Cir. 1987) (enjoining publication of J.D. Salinger’s letters); Houghton Mifflin Co. v. Stackpole Sons, Inc., 104 F.2d 306, 307, 312 (2d Cir. 1939) (enforcing Adolf Hitler’s copyright claim against Alan Cranston, who sought to translate the full version of Mein Kampf to show Hitler’s true beliefs); Shloss v. Sweeney, 515 F. Supp. 2d 1068, 1080–81 (N.D. Cal. 2007) (refusing the effort by James Joyce’s heirs to stop publication of his letters); Religious Tech. Ctr. v. Lerma, 908 F. Supp. 1362, 1364–67 (E.D. Va. 1995) (holding that the Church of Scientology suing to prevent publication of its scriptures was proper). But those are not socially desirable uses of copyright. After Salinger, Congress amended § 107 to prevent the unpublished nature of the work from being conclusive against a finding of fair use. See Daniel E. Wanat, Fair Use and the 1992 Amendment to Section 107 of the 1976 Copyright Act: Its History and an Analysis of Its Effect, 1 Vill. Sports & Ent. L.F. 47, 47 n.1 (1994) (noting that the legislative history of the amendment “reflects an intention to remedy the perceived chilling effect of Salinger”). And we think the argument for fair use is stronger, not weaker, if the plaintiff uses copyright not to control the timing and profit of first publication but to prevent publication of facts altogether. Close and ML systems naturally learn by reviewing (and therefore “taking”) the entire work. The fourth factor (market effect) should normally not prevent fair learning use of individual copyrighted works. The copyright owner of a book or photograph doesn’t create that work in hopes of selling it to AIs. It is possible that they might make some additional money from licensing the work to AIs. But the mere existence of a licensing market directed to such uses shouldn’t make it unfair, just as a “licensing market” for radically transformative uses doesn’t defeat the value of transformation.179179See Bill Graham Archives v. Dorling Kindersley Ltd., 448 F.3d 605, 614–15 (2d Cir. 2006) (“[A] copyright holder cannot prevent others from entering fair use markets merely ‘by developing or licensing a market for . . . transformative uses of its own creative work.’”). Close Only if the use directly interferes with the plaintiff’s core market should the fourth factor outweigh a finding of fair learning under the first factor.
That doesn’t mean that it should always be fair use for ML systems to copy data for use in a training set. The purpose to which the ML system ultimately puts the information may matter to several of the fair use factors. Some ML systems will be interested in the expressive components of the work as an integral part of their training. That is, the goal will be to teach the system using the creative aspects of the work that copyright values, not just using the facts or the semantic connections the law is not supposed to protect. That is particularly likely of those systems like MuseNet that are training in order to generate their own expressive works. Those ML systems both copy expression for expression’s sake and pose a threat of “significant substitutive competition” to the work originally copied.180180See Sag, supra note 43, at 322–23 (arguing that AI copying to create new works is a tougher case for fair use); Sobel, Style, supra note 137 (arguing that AI that produces new art in the style of an existing artist likely “would be unable to avail itself of the fair use defense”). Close
Learning by such systems might still be fair. First, we might distinguish between the input and the output, and say that the act of learning itself should be protected even if you learn from copyrighted work, just as humans can learn music by singing songs or learn literature by reciting poems. So perhaps an AI that learns what makes an Ariana Grande song an Ariana Grande song should be free to do so even though it cares about the expression, not just the facts, just as humans could.181181We are indebted to James Grimmelmann for this point. Close Certainly an AI that creates parodies of a song should be entitled to the same fair use protection a human would.182182Campbell v. Acuff–Rose Music, Inc., 510 U.S. 569, 579 (1994) (finding that a parody qualified as fair use). In fact, an AI named “Weird A.I. Yancovic” that writes song parodies has had its songs taken down by the recording industry. Katie Canales, A Researcher Created a ‘Weird A.I. Yancovic’ Algorithm That Generates Parodies of Existing Songs, and Now the Record Industry Is Accusing Him of Copyright Violations, Bus. Insider (July 24, 2020, 3:19 PM), https://www.businessinsider.com/weird-ai-yancovic-algorithm-parody-song-fair-use-2020-7 [https://perma.cc/GEN5-7CZD]. Close
The problem comes when we ask what we want such an AI to do with that information. What is the output of that AI? Some answers won’t be worrisome from a copyright perspective. We might train an AI to recognize an Ariana Grande-like song in order to try to catch infringers of her songs, for instance. More likely, the AI will produce creative works as its output. Even that isn’t necessarily unfair. Many of the works created by systems like MuseNet will be transformative uses that society values. But it makes the fair use case closer, because the output of the ML’s learning competes with the plaintiff’s core market. And some purposes—say, a system designed to write a new pop song in the style of Taylor Swift or a translation program that produces a translation of an entire copyrighted work—seem more substitutive than transformative,183183For a discussion of “style appropriation” by AI and its copyright implications, see Sobel, Style, supra note 137. Close so that if they run afoul of the ever-broadening definition of similarity in music,184184See, e.g., Williams v. Gaye, 895 F.3d 1106, 1138 (9th Cir. 2018) (affirming the trial court’s judgment that “Blurred Lines” was substantially similar to “Got To Give It Up” and therefore infringed on the copyright); Hall v. Swift, 786 Fed. App’x 711, 712 (9th Cir. 2019) (permitting suit against Taylor Swift based on six-word phrase “players gonna play, haters gonna hate”); Sobel, Style, supra note 137 (discussing the “fickle” treatment of similarity across copyright’s different domains). But see Skidmore v. Led Zeppelin, 952 F.3d 1051, 1056 (9th Cir. 2020) (en banc) (rejecting copyright claim based on similarity of simple musical phrases). Close fair use is unlikely to save them.185185Gillotte argues that there is unlikely to be market harm because the risk of market substitution is small for those customers who want a known artist’s work. Gillotte, supra note 84, at 2688. While that is true, AI-generated works may displace sales by other, lesser-known artists who are interested in the appearance or sound of the work but not in the brand name. Close
Fair learning will also properly distinguish between the use of individual copyrighted works in a training dataset—the issue we are primarily concerned with here—and the wholesale copying of a competitor’s training dataset. Copying a copyrighted database generally involves copying the things copyright does protect: the selection and arrangement of the data. So the purpose of the use won’t normally favor the copier in such a case.186186There may, however, be cases in which the defendant is uninterested in the things that make the database copyrightable and interested only in the unprotectable facts. For instance, if I don’t want your curated database of representative faces, but simply want to collect as many faces as possible for my dataset, I am arguably not copying your database for the purpose of taking the selection and arrangement that make your database copyrightable in the first place. Close Even if it did, taking an entire training database is likely to have a direct market effect, since the value of that database, unlike the value of any individual copyrighted work, is in its use for ML training. Systems are using copyrighted works, but, generally, they’re not using them in ways inimical to copyright’s purposes. Fair learning properly takes that fact into account.187187For a parallel argument under European law, see Mauritz Kop, The Right to Process Data for Machine Learning Purposes in the EU, Harv. J.L. & Tech. Online Digest (forthcoming 2021), https://ssrn.com/abstract=3653537. Close
Other scholars have focused on the nature of the entity doing the copying188188See, e.g., Grimmelmann, supra note 18 (analyzing the difference between human reading and robot reading). Close or whether the use being made by the defendant is itself communicating to the outside world rather than an internal use.189189Sag, supra note 43, at 320 (“Non-expressive use is also justified in terms of expressive substitution, but even more emphatically so. By definition, a non-expressive use does not usurp the copyright owner’s communication of her original expression to the public because the expression is not communicated.”). Close Both of those approaches overlap with ours, but they aren’t focused on the key question. Fair learning isn’t fair because it is a machine doing it, or because it happens outside the public view. It’s fair because the value the ML system gets from the copyrighted work stems from the part of the work the copyright law has decided belongs to the public, not to the copyright owner.190190Grimmelmann doesn’t take this position directly, but he might well agree with us. Close
III. Fair Learning Beyond AIs
A. Copyright for Literate Humans191191With apologies to James Grimmelmann. Grimmelmann, supra note 18. Close
It might seem unfair that humans have to pay for copyrighted works when AIs (often owned by giants like Google) don’t.192192The AI will likely have to get lawful access to the work somehow, so they will often have to pay for works that the copyright owner didn’t make freely available online. Fair learning protects the act of copying; it may not convey a right to access a work not otherwise publicly available. Close But a fair learning doctrine won’t just benefit machines. Humans want to learn too. As we saw in subpart II(C), the law generally lets them learn without fear of copyright liability. And when that learning is at risk, we’ve created several doctrines to make sure that humans have access to the unprotectable aspects of a copyrighted work. But those doctrines don’t always work. Fair learning won’t just protect robots. It might apply to human learning too in a variety of situations where that learning is at risk. Understanding how fair use protects ML may also help courts do a better job of identifying and protecting fair learning by humans too.
First, the right to read a book or watch a show isn’t guaranteed in the statute. It has historically been a function of the physical way in which content was embodied. You don’t need to make a copy of a book in order to read it, to sell it at a used bookstore, or to lend it out at a library. So none of those things traditionally implicated copyright law. But as consumption of content has moved from physical media to computers, that has changed. Essentially everything anyone can do with a copyrighted work on a computer involves making one (and usually many) copies.193193See Lemley, supra note 84, at 554–55 (discussing the multiple copies of a copyrighted work that computers generate when viewing the work). Close That means that while you might have the right to read or resell your physical book, you don’t have the same rights with your e-book.194194See Chris Jay Hoofnagle, Aniket Kesari & Aaron Perzanowski, The Tethered Economy, 87 Geo. Wash. L. Rev. 783, 829 (2019) (discussing how sellers of tethered digital goods are able to restrict the reselling or transferring of their products by consumers); Litman, supra note 166, at 39–41 (noting that technology has expanded the copyright owner’s exclusive right to include an exclusive reading right). Close Indeed, some courts have held that even turning on your computer is illegal because it copies the operating system software into short-term memory.195195MAI Sys. Corp. v. Peak Comput., Inc., 991 F.2d 511, 518–19 (9th Cir. 1993). Congress overruled this case, but only narrowly, providing a specific carveout allowing people to turn on computers in order to diagnose and repair them. 17 U.S.C. § 117(c) (2018). Close Companies that have tried to create a virtual equivalent of reselling used CDs have been sued for copyright infringement and lost.196196Capitol Records, LLC v. ReDigi Inc., 910 F.3d 649, 652 (2d Cir. 2018), cert. denied, 139 S. Ct. 2760 (2019). Close Humans today, in other words, face a similar problem to ML technologies—they have to copy the whole digital work even if they only want to learn the facts or ideas in it. Fair learning can ensure that humans too don’t lose the right to read simply because the book they are reading has moved online.
Computers aren’t the only circumstance in which copyright law might punish humans who just want the facts from a work. In American Geophysical Union v. Texaco Inc.,19719760 F.3d 913 (2d Cir. 1994). Close the Second Circuit held that research scientists did not engage in fair use when they made photocopies of scientific journal articles circulated by the in-house library for a research file or to take into the lab with them.198198Id. at 931. Close The case has occasioned much commentary and criticism, most of it focused on the court’s expansive use of a licensing market to sway the fourth factor against Texaco.199199See, e.g., Lemley, supra note 127, at 189–90 (criticizing the lost licensing revenue theory used by the court in Texaco); see also Ann Bartow, Educational Fair Use in Copyright: Reclaiming the Right to Photocopy Freely, 60 U. Pitt. L. Rev. 149, 165–66 (1998) (discussing the shrinking fair use doctrine, which considers photocopying to be an economic market substitute for the original copyrighted work); Pamela Samuelson, Unbundling Fair Uses, 77 Fordham L. Rev. 2537, 2585–86 (2009) (highlighting that Texaco has “caused a good deal of agitation and anxiety in educational, library, and research communities”). Close But we can also see Texaco as a fair learning case. The scientists who made the photocopies didn’t want the articles in the Journal of Catalysis for their beautiful prose style or even the clever structure, sequence, and organization of the paragraphs. They wanted access to the facts in the article: the graphs and data results. Photocopying the article was a convenient alternative to writing the results down (something that clearly would have been legal). But photocopying, like reading on a computer, makes a copy of the whole work, not just the facts, so the court held it illegal. Because the scientists weren’t commercializing the journal articles and weren’t interested in the things that made those articles copyrightable, Texaco seems a plausible candidate for fair learning in the context of automated reproduction—less automated than ML training, to be sure, but automated still.200200Texaco, 60 F.3d at 921. Close
Fair learning may also help resolve the current dispute over the copyright status of the law itself. This may seem an odd topic to be controversial. Surely the law itself is free for the public to use? Federal law is.20120117 U.S.C. § 105 (2018). Close But the status of official court reporters, official annotations to state codes, local ordinances like planning maps, and private industry codes and standards that are adopted into law is less clear. Copyright owners have filed suit over each category of work. While courts have so far mostly rejected those suits,202202Georgia v. Public.Resource.Org, Inc., 140 S.Ct. 1498, 1509 (2020); Am. Soc’y for Testing & Materials v. Public.Resource.Org, Inc., 896 F.3d 437, 453 (D.C. Cir. 2018); Veeck v. S. Bldg. Code Cong. Int’l, Inc., 293 F.3d 791, 800 (5th Cir. 2002) (en banc). But see West Publ’g Co. v. Mead Data Cent., Inc., 616 F. Supp. 1571, 1579 (D. Minn. 1985) (holding that West Publishing’s court reporters are copyrightable works). Full disclosure: one of us (Lemley) has consulted for Public Resource in the first two cited cases. Close they have struggled to articulate a theory for why a new standard or ordinance couldn’t be copyrighted and, if it is, why wholesale copying should be permitted. Some courts have held the ordinances unprotectable under merger,203203Georgia, 140 S. Ct. at 1507–08; Veeck, 293 F.3d at 801. Close while others have held the copying of those ordinances to be fair use.204204Am. Soc’y for Testing & Materials, 896 F.3d at 453. Close In 2020, the Supreme Court held that the annotations to state statutes were not protected by copyright because they were “government edicts” given legal significance even though drafted by private parties.205205Georgia, 140 S. Ct. at 1509. Close But it left open the possibility that other documents adopted into law, such as private standards required by regulation, might retain copyright protection. If the Court does find them copyrightable, we think even wholesale copying of industry standards mandated by regulation for any purpose other than commercial sale is likely to be protected by fair learning. People don’t read statutes or regulations for entertainment; they read them for the legal mandates they contain. Learning what law governs a person’s behavior seems a particularly important form of learning in a democratic society.
We might also apply fair learning to excuse copying of newsworthy material for the purpose of reporting the news. That would require changing some case law, particularly the cases involving videos of famous historical events. The person who happens to take a video of the Kennedy assassination or the Reginald Denny beating has created a copyrighted work,206206See Robert A. Gorman, Copyright Protection for the Collection and Representation of Facts, 76 Harv. L. Rev. 1569, 1597–98 (1963) (discussing the argument for granting photographs copyright protection). Close but viewers and the news media aren’t interested in the copyrighted bits. They want to see what actually happened, not the accidents of the plaintiff’s angles and lighting. Because often only one person filmed the event, anyone who wants to watch that event needs to watch the copyright owner’s version. Giving control to the copyright owner locks up the unprotectable as well as the protectable parts, and here those are the parts that people care about. Nonetheless, most, though not all, cases involving media reporting of news videos have rejected fair use claims.207207Monge v. Maya Magazines, Inc., 688 F.3d 1164, 1176 (9th Cir. 2012) (holding that the photos of celebrity couple’s secret wedding “did not transform the photos into a new work . . . or incorporate the photos as part of a broader work”); L.A. News Serv. v. Reuters Television Int’l, Ltd., 149 F.3d 987, 994 (9th Cir. 1998) (holding that a rebroadcast of a video clip of beating of Reginald Denny during the 1992 Los Angeles riots was not fair use); L.A. News Serv. v. KCAL–TV Channel 9, 108 F.3d 1119, 1123 (9th Cir. 1997) (same). But see L.A. News Serv. v. CBS Broad., Inc., 305 F.3d 924, 940–42 (9th Cir. 2002) (holding that a shorter use of the same clip was fair use); Núñez v. Caribbean Int’l News Corp., 235 F.3d 18, 25 (1st Cir. 2000) (finding fair use of a newsworthy photograph); Time Inc. v. Bernard Geis Assocs., 293 F. Supp. 130, 146 (S.D.N.Y. 1968) (finding the use of frames from the Zapruder film of the Kennedy assassination to be fair use). Close We think that is wrong. Selling the video for profit might not be fair use, but using it to learn what happened should be.
We might extend the concept of fair learning even further to the analogous idea of “fair functioning.” Many copyrighted things serve a purpose beyond just communication.208208Cf. Michael J. Madison, Law as Design: Objects, Concepts, and Digital Things, 56 Case W. Res. L. Rev. 381, 389 (2005) (differentiating the “material” and “communicative function[s]” of “things” in the law). Close That list includes software, which is a literary work (code is written) but is for all practical purposes a functioning machine;209209Dan L. Burk, Method and Madness in Copyright Law, 2007 Utah L. Rev. 587, 613–15; Samuelson, supra note 199, at 2607–08. Close a sequence of yoga poses, which can be creative and expressive but which are clearly designed to train the body and not just as a form of dance;210210Bikram’s Yoga Coll. of India, L.P. v. Evolation Yoga LLC, 803 F.3d 1032, 1036 (9th Cir. 2015) (rejecting copyright claim to a sequence of yoga poses where the sequence itself served a functional purpose). Close and clothing design, which combines both artistry and function.211211Star Athletica, L.L.C. v. Varsity Brands, Inc., 137 S. Ct. 1002, 1012–13 (2017). For a host of critiques of that decision, see also supra note 174. Close Just as the desire to access the facts and ideas of a written work should influence the analysis of fair use, arguably the defendant’s desire to make use of the functional aspects of a useful work rather than its artistic ones—to practice Bikram-style yoga, to make a computer program work with another, and so on—should be a factor that favors a finding of fair use.
Fair learning, in short, isn’t just for machines. Humans too may need a practical, fair use-based right to learn and to do the things that copyright law nominally says they can learn and do.
B. Toward a Pluralist Theory of Fair Use
Treating fair learning as a lawful purpose under the first factor also offers some broader lessons for fair use doctrine more generally. It provides a desirable counterbalance to the recent emphasis on transformative use, opening the way to a more pluralistic vision of fair use. Transformative use has arguably swallowed fair use doctrine in the past twenty-five years.212212See supra notes 93–95 and accompanying text. Close Transformation is important. Transformative works are themselves creative works that copyright law should encourage, not discourage. But the rush to make transformative use the centerpiece of fair use doctrine has obscured the fact that uses need not be transformative to be fair.213213Tony Reese emphasizes that use of a work might have a transformative purpose even if it doesn’t transform the content of the work. R. Anthony Reese, Transformativeness and the Derivative Work Right, 31 Colum. J.L. & Arts 467, 485 (2008). Michael Carroll argues that compiling a database for research purposes is a transformative purpose, though not a transformation of the content. Carroll, supra note 58, at 941–44. Close Some of the classic examples of fair use, such as recording a song or television show at home, are fair not because the defendant did anything new or creative but because they aren’t commercial and don’t have any likely market effect.214214E.g., Sony Ent. Corp. of Am. v. Universal City Studios, 464 U.S. 417, 451 (1984). Close Other examples, like copies for classroom use21521517 U.S.C. § 107 (2018). Close or the reproduction of images from the Zapruder film of the Kennedy assassination,216216Time Inc. v. Bernard Geis Assocs., 293 F. Supp. 130, 146 (S.D.N.Y. 1968). Close are fair not because they are transformative or because they have no market consequence but because they serve valuable social purposes, educating students or permitting informed discussion of political and social issues.217217Adolf Hitler’s suit against Alan Cranston for translating the full version of Mein Kampf to show how the official translation had been sanitized was held not to be a fair use, Houghton Mifflin Co. v. Stackpole Sons, Inc., 104 F.2d 306, 307, 312 (2d Cir. 1939), but it seems a quintessential example of a public benefit that should have been held fair. A number of scholars have recently challenged the primacy of transformative use. Haochen Sun argues that the public interest factor should play a larger role in fair use regardless of whether the use was transformative. Haochen Sun, Copyright Law as an Engine of Public Interest Protection, 16 Nw. J. Tech. & Intell. Prop. 123, 141 (2019); see also Asay et al., supra note 94 (criticizing the undue emphasis on transformative use); Alexander McMullan, Returning to the Fair Use Standard, 63 N.Y.L. Sch. L. Rev. 359, 370 (2018–2019) (calling for courts to return to a balancing of the transformativeness of a use of a copyrighted material with the other fair use factors—including the promotion of useful works for the “public good”). We agree with those critiques. Close Fair use has long been about more than transformation. It is important to recall that more pluralistic vision of fair use.218218See Samuelson, supra note 199, at 2618 (describing the interest of the public encompassed within the fair use doctrine as being necessary to “what constitutes a good society”). Close
Fair learning adds two important policy rationales to this pluralistic vision of fair use. First, it addresses the problem of overinclusiveness in copyright enforcement. Copyright law gives owners control over some parts of their work but not others. When users can’t separate the protectable from the unprotectable parts, however, control over part can easily become control over all. That can happen as an accident, but it can also happen deliberately. Copyright owners regularly use the law as a tool to prevent disruptive competition that threatens their incumbent markets.219219Lemley & McKenna, supra note 122, at 120. Close That’s why the law denies protection altogether in the merger and inseparable useful article cases. But denying all protection is unfair to copyright owners in many cases because they have in fact contributed substantial expression that the law wants to encourage. Another approach is to limit remedies so that they are consonant with the scope of what is actually protected.220220Mark A. Lemley & Mark P. McKenna, Scope, 57 Wm. & Mary L. Rev. 2197, 2235, 2239 (2016). Close That may mean denying injunctive relief and reforming damages in automated copying cases, as Lemley and Weiser have proposed.221221Lemley & Weiser, supra note 175, at 803. Close But while that is achievable when it comes to injunctions,222222See eBay Inc. v. MercExchange, L.L.C., 547 U.S. 388, 392–93 (2006) (holding that courts should apply the traditional four-factor test in deciding whether a permanent injunction should issue in a patent case and rejecting a rule that an injunction should automatically issue when a patent is infringed); Salinger v. Colting, 607 F.3d 68, 79–80 (2d Cir. 2010) (holding that the Supreme Court’s eBay decision applies with equal force in the copyright context and thus rejecting a general rule that an injunction should issue when a party infringes another’s copyright). Close it would require changes to the damages statute. Treating fair learning as fair use can help to calibrate the scope of copyrights, ensuring that copyright owners get control over expressive elements and uses of their work but can’t leverage that right to effectively control the unprotectable elements of their work.223223Samuelson, supra note 199, at 2587; Lemley & McKenna, supra note 220, at 2201, 2210. Close
Fair learning offers a second theoretical lesson, one that goes to the heart of the purpose behind copyright law. A central problem with allowing copyright suits against ML is that the value and benefit of the system’s use is generally unrelated to the purpose of copyright. That is true not only because the ML system wants the facts, ideas, and other unprotectable elements of the work. Arguably it’s true even if the technology gains a non-expressive benefit from the expressive parts of the work too, perhaps by learning how to recognize a particular artist’s song or painting. That is use of the protectable expressive parts of the plaintiff’s work. But the ML system doesn’t care whether the work is expressive or not and which aspects are protected. It just wants to learn from the work in order to put that knowledge to a different instrumental use.224224Yeah, we know, we’re anthropomorphizing AIs. They don’t really “want” anything. Deal with it. As we’ve noted elsewhere, everyone does it. See Casey & Lemley, supra note 7, at 353–55 (describing the tendency to anthropomorphize robots and citing examples in various fields). Close Perhaps this should be a broader principle of fair use, one not limited to fair learning: If the defendant has no interest in the work because of the thing that makes that work copyrightable, the use is presumptively not one that interferes with the purpose of copyright law, and so ought to be considered fair.
Whether or not you agree with us that fair learning by ML should be fair use, the concepts underlying fair learning are concepts fair use doctrine should take into account. ML systems therefore have much to teach us about copyright law for humans too. As James Grimmelmann notes, “paying attention to robotic readership refocuses our attention on the really fundamental questions: what is copyright, and what is it for? To say that human readers count and robots don’t is to say something deep about the nature of reading as a social practice, and about what we want robots—and humans—to be.”225225Grimmelmann, supra note 18, at 681. Close We don’t think the law should treat robots and humans differently. On the contrary, each should be entitled to learn from a copyrighted work in the way they naturally learn.
Conclusion
Machine learning requires the copying of extraordinary amounts of copyrighted material. That copying should generally be permitted. Most ML systems copy works not to consume the expression copyright law protects, but to get access to the facts or structures copyright law dedicates to the public. Understanding this as fair learning can help ensure we can train ML systems without interference from the law. But the idea of fair learning doesn’t just matter for robots. It can help us resolve a number of troubling copyright cases involving humans too. And it reminds us that fair use is about more than just transforming copyrighted works into new works. It’s about preserving our ability to create, share, and build upon new ideas. In other words, it’s about preserving the ability to learn—whether the entity doing the learning is a person or a robot.