Image-Generative AI: Has Technology Evolved Beyond Modern-Day Fair Use?
Image-Generative AI: Has Technology Evolved Beyond Modern-Day Fair Use?
5.24.2024
Introduction
Copyright is uniquely situated between the force of law and the pressure of an ever-evolving society. Since its inception, courts, legislators, and administrators have had to adjust and recontextualize copyright law to appropriately conform to new technologies and artforms. The latest technology to challenge this pillar of intellectual property law is generative artificial intelligence and most recently, image-generative AI.
Although AI is not an inherently new technology, its use and proliferation have skyrocketed in recent years with the release of various consumer-facing AI offerings. In particular, the introduction of image-generating AI platforms – platforms where users can enter text prompts like “draw me a sunset in the style of Van Gogh,” and the software responds by generating a unique image that users can continue to refine through the same process – have challenged our modern understanding of both what constitutes a copyrightable work and what constitutes infringement of copyrighted materials. This article will focus on the latter of these two queries.
The matter of copyright infringement in the image-generative AI space will be resolved in federal court.[1] AI, in general, is not a novel issue for federal courts, which have heard a wide variety of AI or similar technology-related cases since the dot-com boom. Image-generative AI, however, is a question of first impression for the courts, and the first image-generative AI cases have just begun to pop up on dockets across the country.[2] In these new image-generative AI cases, courts are seeing a variety of intellectual property claims, almost all of which include a claim of copyright infringement.[3] With the influx of novel claims, legal scholars posit that defendant AI companies will need to test the limits of the fair use doctrine as an affirmative defense to infringement claims made by artists, photographers, and other visual creators.[4] Beyond the fact-intensive analysis required for fair use defenses, AI fair use cases are further complicated by the lack of public understanding of exactly how AI works, including how AI platforms process their learning materials and how AI programs generate content. These issues are even more complex in the space of image-generative AI given the way such platforms and their outputs are used.
While each pending case deserves its own day in court and a fact-specific inquiry into the infringement at issue, this article attempts to simplify the applicable fair use analysis by considering it separately in instances of each of: 1) infringement for the purpose of machine learning – AI inputs and 2) infringement in the materials created by such machines – AI outputs. By exploring the fundamental premises of the fair use doctrine in relation to both this new technology and available AI-related caselaw, in this article we suggest that our court system is giving deference to technology defendants that it does not extend to traditional artists or creators. In particular, the first factor of the fair use doctrine heavily favors technological advancement such that it almost guarantees sanctuary to AI defendants, particularly AI input defendants. We do not opine on the inherent fairness of the fair use doctrine, but instead summarize how the doctrine’s structure favors new technology to such an extent that plaintiffs stand little chance when pitted against novel technological advances such as image-generative AI.
Groundwork: AI Inputs, AI Outputs, and Fair Use
In the first image-generative AI cases making their way through our court system, copyright holders have asserted various intellectual property claims against AI companies, including violations of the Digital Millenium Copyright Act, unfair competition, and copyright infringement.[5] When addressing these infringement claims, we are beginning to see the courts fragment them into distinct inquiries surrounding AI inputs and AI outputs.[6] As we explain in more detail below, the categorization as an input or an output has a significant impact on how easily an AI defendant can assert a fair use defense.
AI input claims concern the many works that are loaded onto an AI platform for the purpose of machine learning before the image-generation process has even begun. In order to create a functioning image-generative AI platform, developers must “teach” the machine how to read, analyze, and ultimately create unique images.[7] AI input infringement claims allege that a violation occurs at this early point of the development process when developers feed the machine hundreds of thousands, if not millions, of images, typically using image datasets pulled or scraped from the internet.[8] Here, the inclusion of a copyrighted work in a training dataset is the basis for an AI input infringement claim. The platform copies these images[9] and analyzes them for various probabilities. As the platform reads more and more images, it begins to develop associations to predict how the various elements of a work will lend themselves to the whole. Once trained, these associations allow the AI platform to impute knowledge regarding the various elements of style, subject, and composition from the inputs to the images that the program ultimately creates – the AI outputs.
AI output infringement claims revolve around the final images created by the AI platforms at the direction of end users.[10] These claims are being heavily refuted by AI companies who say that users would have to actively endeavor to create an infringing AI output and, even then, infringements only occur around .0003% of the time.[11]
Legal scholars and commentators in the space have speculated that, to the extent AI companies don’t refute the infringement claims altogether, AI defendants could rely on fair use as a viable affirmative defense.[12] When asserted, courts consider the following factors to determine if an infringement is permissible under the fair use doctrine:
(1) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes; (2) the nature of the copyrighted work; (3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and (4) the effect of the use upon the potential market for or value of the copyrighted work.[13]
As discussed in more detail below, the first factor of the fair use doctrine has historically been weighted heavier in technology inquiries than in cases surrounding more traditional artistic mediums. Specifically with respect to image-generative AI, this disproportionate emphasis on factor one has the ability to skew the entire fair use analysis. This is particularly true of AI input infringement cases, but will extend to AI output cases as well, and may signal a need for a technology-specific doctrine.
Analyzing the First Fair Use Factor
Under factor one of a fair use analysis, courts are asked to consider the purpose and character of the infringing use, including whether such purpose and character is commercial in nature.[14] In practice, courts typically ask if the infringing use is transformative.[15] In other words, does the actual use of the infringing work differ from the intended or normal use of the original work?[16] While this may appear to be a straightforward inquiry, the way courts have analyzed the ultimate purpose of the allegedly infringing works in technology-related infringement cases has yielded surprising results.[17]
For example, in two Second Circuit cases brought by the Authors Guild in 2014 and 2015, the plaintiffs in each case brought a copyright infringement action against a party who scanned over 10 million published books for use in an internet database where users were able to perform full-text searches on titles of their choosing and could see snippets of the searched work in the context of the actual copied material.[18] Even though each book was scanned in its entirety and could, in some instances, be observed in the database in its original form, the Second Circuit found this unlicensed copying to be a transformative use.[19] The court reasoned that the purpose of the copying was to create an online database that allowed users to search within the books, emphasizing that a searchable database had an entirely different function than that of the book itself, and that difference in function weighed in favor of fair use.[20]
Similar logic can be found in Google v. Oracle, in which the Supreme Court held that Google’s copying of 11,500 lines of plaintiff’s copyrighted code was a fair use.[21] In this case, Google had acquired Android, Inc., an early-stage smartphone software company, and through the purchase, sought to create a new smartphone platform written in the Java programming language. To facilitate the code-writing process, Google copied unaltered code from the application programing interface tool within the Java Standard Edition program owned by Oracle. Google used the programming interface code for the same purpose as its original use: to develop new software programs. In finding the work transformative under the first factor of its fair use analysis, the court reasoned that Google’s use did not simply create a new software program, but instead created a new software platform that would spur technological innovation in such a manner that was “consistent with that creative ‘progress’ that is the basic constitutional objective of copyright itself.”[22] Notably, in its factor one analysis, the court emphasized that computer programming differs from traditional literary works in that computer programs “almost always serve functional purposes,” and noted that this has led some courts to lament that the application of copyright law to computer programming matters is “like applying a jigsaw puzzle whose pieces do not quite fit.”[23]
The message from our court system in these technology-related infringement cases is two-fold: first, these cases tell us that the empirical value of a new invention weighs heavily in favor of a finding of transformative use; next, the dicta tell us that transformative use, and the entire fair use doctrine, might not be the appropriate test for this type of defendant.[24]
Application to AI Inputs
The aforementioned data copying cases have created a strong blueprint for AI companies to defend their use of copyrighted items as AI inputs, since the metamorphosis from image to intelligent machine will almost certainly be deemed sufficiently transformative. Though plaintiffs can assert numerous traditional “uses” of the original artwork (e.g., public enjoyment, culture, licensing, etc.), defendants of image-generative AI will likely assert that the original images were used to train technology, or, taken one step further, to train technology to create new and unique artwork. Similar to the Authors Guild cases, AI defendants can argue that the copied images serve an entirely unique function from the original works. If that on its own does not weigh in favor of fair use, companies may look to assert the defense used in Google and claim that the copying was a means to creating an entirely new technology.
Application to AI Outputs
Defendants in an image-generative AI output case will face a more substantial hurdle in proving the first factor of the fair use test because image-generative AI outputs are pictorial. This means they take the same form as the original, allegedly infringed-upon work, so AI output defendants will not be able to avail themselves of the unique “transformative purpose” argument in the same way that AI input defendants may. Instead, the inquiry will focus more on the transformation of the image itself, asking if the new work functions as a substitution for the original or if it uses the original to “serve a distinct end.”[25] Where output defendants can assert that the infringing work serves a traditionally viable purpose, such as to criticize, comment upon, or learn from the original work, courts will have a clear roadmap to a ruling.[26] But what about art for the sake of art? In the event that the infringing work was created for no other reason than to create art, defendants will still have to justify the reason behind their work, and courts will be put in the position of assessing the relative value of the pictorial illustrations in the infringing work. The Supreme Court has warned that this was a “dangerous undertaking” as early as 1903,[27] and has maintained the concern in recent years, emphasizing that judicial interpretation of artwork could usurp the policy choice embedded within the letter of the law.[28]
The first factor analysis is further complicated by its reference to commercial use. Courts have historically varied in their emphasis of this subfactor, but typically state that commercial use is not dispositive in a fair use analysis.[29] Most recently, however, the Supreme Court in Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith ruled that when the original and infringing work share similar purposes, and the infringing use is commercial, the commercial nature of the use is likely to weigh against a finding of fair use.[30] Image-generative AI outputs, having a less transformative purpose than AI inputs, may have to directly address the commercial use analysis in factor one.
Analyzing the Second Fair Use Factor
Factor two simply instructs the court to consider the nature of the original work. Though short and sweet, this rule gives credence to the original intent of the Copyright Act and tells factfinders to assess whether the original work is: 1) more expressive or factual, and 2) published or unpublished.[31] When the original work is closer to the “core of intended copyright protections,” with the pinnacle of intended protection being expressive and unpublished works, the law dictates that fair use defenses should be subject to greater scrutiny.[32]
Application to Both AI Inputs and Outputs
In practice, factor two rarely carries significant weight in the overall determination of a fair use dispute.[33] In analyzing factor two, courts often make reference to their factor one findings, in some instances going so far as to say that when the use is found to be transformative under factor one, factor two is of “limited usefulness.”[34] For AI input cases, which will generally have a strong transformative use argument, the inquiry into factor two is likely to be brief and inconsequential. Still, to the extent that factor two analysis is examined in detail in either input or output cases, the courts will likely determine that the second factor is not dispositive. Caselaw is instructive here. When examining factor two for expressive, published original works, courts have typically found that the second factor is not dispositive to a fair use finding.[35] In both input and output image-generative AI cases, the original use at issue will most likely be expressive (as a work of art) and published (as most inputs are scraped from publicly available websites), and there is nothing to suggest that the nature of AI inputs or outputs, or the nature of the underlying work, will be afforded any unique considerations under this factor.
Analyzing the Third Fair Use Factor
Factor three asks how substantial the infringing use is and looks to both the amount of the original work used and whether the copied content was integral to the expression of the original work.[36] Courts consider if the infringing work has taken more of the original work than is necessary and, similarly, if the infringing work’s use of the original work is excessive.[37]
As a question of scale, this factor may seem like a simpler inquiry than factors one or two, but courts have emphasized that there is no limit to how much of a work may be used and still be considered fair use.[38] In Campbell v. Acuff-Rose Music, the Supreme Court emphasized that the factor three inquiry should focus not on the proportion of the original work used in the infringing work, but instead on if more of the original work was taken than was necessary.[39] While Campbell was an analysis of a parody song, and the court focused on the necessity of copying the entire work for the purposes of parody, courts have taken the concept of what is “necessary” and have applied it to the transformative infringing use asserted under factor one.[40] This application stretches the text of the law, which directs the factfinder to assess the amount of the original work used in relation to the original work as a whole. By contrast, under the “necessary use” doctrine concocted by the court in Campbell, when a use is found to be transformative under factor one, the factfinder instead must consider how much of the original work was necessary to achieve the ultimate purpose of the infringing work.
This dichotomy is demonstrated in Authors Guild v. Hathitrust, in which the Second Circuit found that factor three weighed in favor of fair use even though the defendants had scanned each book at issue in its entirety. The Second Circuit reasoned that Hathitrust’s purpose was to create a searchable book database, and thus copying each entire book was necessary as the entire book must be searchable for the technology to function.[41] The Supreme Court echoed this application of the third factor in Oracle v. Google, when it found that, even though Google had used more of Oracle’s code than was necessary for the immediate purpose of writing in a specific programming language, assessing only Google’s immediate purpose analyzed its “legitimate objectives too narrowly,”[42] and, rather, when considering Google’s larger vision of creating a new smartphone platform, the ends justified the means.[43] These cases demonstrate that when courts engage in a “necessary use” analysis, they effectively place greater emphasis on factor one – which is already favorable to AI platforms thereby decreasing the relative impact of factor three.
Application to Inputs
As applied to AI inputs, AI developers may try to refute assertions of substantial use with claims that the AI software does not store any of the training images and, by its nature, the software is only using the necessary amount of the original work. According to the plaintiffs in Andersen, the image-generative AI platform reads the training data sets and records various mathematical equations and calculations based on each training image, and only these formulas are maintained within the platform.[44] Under such assertions, AI developer defendants could have a strong defense that any taking amounts only to the taking of factual materials (the data and calculations surrounding each training image), and not any of the expressive elements of the original works (which are not otherwise utilized or stored within the platform) and, therefore, the AI is only scraping that which is necessary. Even if AI developers do store the original works in their platforms, they could still seek shelter in the broad view of ultimate use expanded upon by the Oracle court, finding safe haven under factor one and claiming that any use of the training images – indeed, copying them in their entirety – was necessary to create a new, functional technology from which invention and creativity can blossom. Conversely, might artists refrain from publishing their works for fear of their being fed into an image-generative AI product which might erode the artist’s rights in its creation?
Application to Outputs
If an AI defendant is able to successfully allege that it did not store any of the original work files, the affirmative fair use defense should be equally persuasive for outputs as well. If courts opt to compare the data derived from the original work to the image created in the secondary work, plaintiffs may have difficulty demonstrating how this data translates into the expressive factors of the original work, or how it generates an image that is substantially similar to the original work.[45] If courts reject a comparison of the original work data, and instead opt to assess the original work file, both parties may face difficulty in making or defending their position and the factor three determination will hang in the balance. AI output defendants may have to justify the copying of the entire original work without being able to rely on the same necessary transformative use analysis availed by AI input defendants. AI output plaintiffs will also face an uphill battle and will likely need to show an infringing work that maintains a substantial similarity to the original, which image-generative AI plaintiffs have not yet been able to do in their pleadings,[46] and which AI defendants claim is a near impossibility.[47]
Analyzing the Fourth Fair Use Factor
The fourth factor assesses the market of the copyrighted work, and whether the secondary work will detract from that market.[48] The mere existence of the secondary work alone is not sufficient justification for any lost sales in the original work; instead, the doctrine again ties back to factor one, as lost sales can only be considered if the original and secondary uses are of the same nature[49] (i.e., both were created with the purpose of licensing). Further, when considering lost revenue, courts must look to the recipient (if any) of the diverted cash flow. If the copyright holder’s lost revenue was not converted into a revenue stream for the infringer, the lost profits are not considered a cognizable loss under the Copyright Act.[50] Finally, courts have dictated that plaintiffs cannot make broad claims in anticipation of future lost revenue but, instead, future-looking claims must be certain losses.[51] Limiting the calculation of lost profits in this manner makes it very difficult for a plaintiff to collect sizable damages without quantifying definite losses, which is of itself an inherently burdensome task.
Application to Inputs
With respect to inputs, the fact that the original work and the secondary work are fundamentally different will likely weigh in favor of a fair use finding. AI defendants can assert that their use of the original material(s) is for the entirely different purpose of machine learning, thus neutralizing any claims of lost sales or licensing opportunities.
Application to Outputs
With respect to outputs, a finding in favor of fair use will be more difficult, but not impossible, to obtain. Because both the original work and the infringing output are of the same nature and thus avail themselves to the same market, there is likely to be a more in-depth factor four analysis for outputs than for inputs. Copyright holders will have to demonstrate with substantial certainty the impact on the market with respect to both sales of the original and infringing work, which will be difficult. Any profit realized by the AI defendants is likely to come from the licensing of the AI platform rather than from payments tied to the allegedly infringing work itself, which calls into question the applicability of the factor four inquiry as the original and infringing uses are not of the same nature. As of the date of this writing, no image-generative AI input or output infringement claims have been fully adjudicated, but when such claims are analyzed, the outcome of factor four will depend upon such defendant’s decision to monetize the infringing work in the same manner as the plaintiff, which is not guaranteed.
Conclusion
The assertion of a fair use defense by an image-generative AI developer is more viable in relation to AI inputs, and less certain, though not foregone, with respect to AI outputs. But is this truly indicative of a “fair” use? Under the current doctrine, AI companies are poised to profit handsomely from their inventions, while the artists and creators whose work the AI platforms are built upon, are not entitled to any compensation. Some may view this as an unjust use of the original creative labor. Others may find this to be an example of copyright law working in the way that it should: promoting the development of new and important technology. Either way, it is difficult to ignore the caselaw and dicta, which suggest that technology is simply different. If technology truly is different, and that difference eliminates the practical applicability of the current fair use doctrine, it may be time to afford the technology sector its own infringement doctrine and a fair use inquiry that is thoughtfully tailored to this difference. This will be increasingly true as image-generative AI cases come to the forefront.
Wendy Heilbut is the founding partner of Heilbut LLP. Her practice focuses on intellectual property and corporate transactional matters through which she guides high-growth companies, bridging legal services with strategic advice. Danielle Maggiacomo focuses on protecting intellectual property assets, while Maggie Casey advises entrepreneurs and startups on a wide variety of corporate transactional matters.This article appears in a forthcoming issue of Bright Ideas, the publication of the Intellectual Property Law Section. For more information, please visit NYSBA.ORG/IPS.
Endnotes:
[1] See Letter from Reg. Perlmutter to Congress (May 12, 2023) (explaining that the Copyright Office is actively monitoring litigation and has no current plan to address infringement issues directly); see also infra.
[2] See, e.g., Andersen v. Stability AI Ltd., No. 3:23-cv-00201, 2023 WL 7132064 (N.D. Cal. Jan. 13, 2023); Getty Images (US) Inc. v. Stability AI, Inc., No. 1:23-cv-00135 (D. Del. Feb. 3, 2023).
[3] See, e.g., id.
[4] See, e.g., Christopher T. Zirpoli, Congressional Research Service, Generative Artificial Intelligence and Copyright Law (Sept. 29, 2023), https://crsreports.congress.gov/product/pdf/LSB/LSB10922; Gil Appel, Juliana Beelbauer, David A. Schwiedel, Generative AI Has an Intellectual Property Problem, Harvard Bus. Rev. (Apr. 7, 2023), https://hbr.org/2023/04/generative-ai-has-an-intellectual-property-problem; Katherine Klosek and Marjory S. Blumenthal, Training Generative AI Models on Copyrighted Works Is Fair Use (Jan. 23, 2024), https://www.arl.org/blog/training-generative-ai-models-on-copyrighted-works-is-fair-use/.
[5] See, e.g., Andersen, No. 3:23-cv-00201; Getty Images (US) Inc., No. 1:23-cv-00135.
[6] See generally Andersen, No. 3:23-cv-00201.
[7] See, e.g., First Amended Complaint § IX, Andersen, No. 3:23-cv-00201 (N.D. Cal. Nov. 11, 2023); see also Van Lindberg, Building and Using Generative Models Under US Copyright Law, 18 Rutgers Bus. L. Rev. 1 (Spring 2023).
[8] See, e.g., First Amended Complaint § VIII, Andersen, No. 3:23-cv-00201 (N.D. Cal. Nov. 11, 2023).
[9] Note that AI developers in some instances have rejected the contention that the images themselves are copied or saved onto the platform. See Lindberg, supra n.7, at 39 (explaining that during the machine learning process, generative AI platforms do not store the files used to train them within the computer program).
[10] See, e.g., Class Action Complaint § IX, Andersen, No. 3:23-cv-00201 (N.D. Cal. Jan. 13, 2023).
[11] See Lindberg, supra n.7, at 11 (citing Nicholas Carlini et al., Extracting Training Data from Diffusion Models (Jan. 30, 2023), https://arxiv.org/pdf/2301.13188.pdf); see also Motion to Dismiss § 1b, Andersen, No. 3:23-cv-00201.
[12] See Klosek & Blumenthal, supra n.4.
[13] See, e.g., Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508, 523-23 (2023) (explaining the fair use factors and citing 17 U.S.C. § 107).
[14] See, e.g., Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508, 527-28 (2023).
[15] See, e.g., id. at 529.
[16] See, e.g., id. at 569 (citing Harper & Row, Publishers, Inc. v. Nation Enterprises, 471 U.S. 539, 588 (1985)).
[17] At the time of writing, we had yet to hear from any courts on image-generative AI, thus, this article reviews caselaw on non-image related AI and other novel technologies in its findings.
[18] Authors Guild, Inc. v. Hathitrust, 755 F.3d 87 (2d Cir. 2014); Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015).
[19] See Hathitrust, 755 F.3d at 97.
[20] See id.; Google, 804 F.3 at 216.
[21] Google LLC v. Oracle America, Inc., 141 S.Ct 1183 (2021).
[22] See id. at 1203.
[23] See id. at 1198 (internal quotation marks omitted).
[24] See, e.g., id. at 1198, 1208 (explaining that some judges have concluded that “applying copyright law to computer programs is like assembling a jigsaw puzzle whose pieces do not quite fit” at 1198, and “the fact that computer programs are primarily functional makes it difficult to apply traditional copyright concepts in that technological world” at 1208).
[25] See Andy Warhol Foundation, 598 U.S. at 528.
[26] See id. (noting that section 107 of the Copyright Act lists “the sorts of copying that courts and Congress most commonly ha[ve] found to be fair uses”).
[27] See Bleistein v. Donaldson Lithographing Co., 188 U.S. 239, 251 (1903).
[28] See Star Athletica, L.L.C. v. Varsity Brands, Inc., 580 U.S. 405, 423 (2017) (citing Bleistein, 188 U.S. at 251 (1903)).
[29] See Andy Warhol Foundation, 598 U.S. at 525.
[30] See id. at 511.
[31] See Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 11 F.4th 26, 45 (9th Cir. 2021).
[32] See Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569, 586 (1994).
[33] See Google, 804 F.3 at 220.
[34] See Hathitrust, 755 F.3d at 98.
[35] See, e.g., id.
[36] See Oracle, 141 S.Ct at 1205.
[37] See Hathitrust, 755 F.3d at 98.
[38] See id.
[39] See Campbell, 510 U.S. at 589.
[40] See, e.g., Hathitrust, 755 F.3d at 98.
[41] See id.
[42] See Oracle, 141 S.Ct at 1205-06.
[43] Id.
[44] See Order on Motions to Dismiss And Strike, Andersen v. Stability AI Ltd., No. 3:23-cv-00201 (N.D. Cal. Nov. 11, 2023).
[45] See id.
[46] See id.
[47] See Lindberg, supra n.7, at 11; see also Motion to Dismiss § 1b, Andersen, No. 3:23-cv-00201.
[48] See Hathitrust, 755 F.3d at 99.
[49] See id. (“At the outset, it is important to recall that the Factor Four analysis is concerned with only one type of economic injury to a copyright holder: the harm that results because the secondary use serves as a substitute for the original work.”) (citations omitted).
[50] See Oracle, 141 S.Ct. at 1206 (“As we pointed out in Campbell, a ‘lethal parody, like a scathing theatre review,’ may ‘kil[l] demand for the original.’ Yet this kind of harm, even if directly translated into foregone dollars, is not ‘cognizable under the Copyright Act.’”) (citations omitted).
[51] See, e.g., Google, 804 F.3d at 224.
