Copyright Office’s DMCA Report Raises an Interesting Question: Does a Computer Know What is Fair?

On May 21, 2020, the Copyright Office issued a Report on the DMCA (“Report”)^[1] expressing the view that case law applying the DMCA’s safe harbor provision have fallen out of balance, tilting too far in favor of online platforms and reducing protection for copyright owners.  As an example of this perceived imbalance, the Report points to the Ninth Circuit’s ruling in Lenz v. Universal Music Corp., aka “the dancing baby case,” which held that before sending a DMCA takedown notice, the copyright owner must make a good faith determination that the use is infringing, rather than a fair use.  This ruling collides with the increasingly common practice of using Artificial Intelligence to monitor websites for potentially infringing content and raises an interesting question: can a computer meaningfully assess whether a use is “fair use” for purposes of the Copyright Act?  And relatedly—if a computer can know what is fair use, doesn’t that also mean the computer can know when its infringing?

The Dancing Baby Case

In Lenz, the Ninth Circuit held that copyright holders are required to consider fair use before sending a takedown notification, and that there can be a triable issue as to whether the copyright holder formed a subjective good faith belief that the use was not authorized by law.^[2] The case became widely known as the “dancing baby case” because it concerns Lenz uploading a home video to YouTube of her two young children dancing to “Let’s Go Crazy” by Prince. Universal was Prince’s publishing administrator at the time and would monitor YouTube on a daily basis for use of Prince’s songs. Universal would evaluate the videos for whether they “embodied a Prince composition” by making “significant use of ... the composition, specifically if the song was recognizable, was in a significant portion of the video or was the focus of the video.” ^[3] According to Universal, the general guidelines were that they “review the video to ensure that the composition was the focus and if it was [] then notify YouTube that the video should be removed.”” ^[4] Universal concluded that in Lenz’s video, the song “was very much the focus of the video.” ^[5] As a result, Universal decided the video should be the subject of a takedown notice. The notice included a “good faith belief” statement as required by 17 U.S.C. § 512(c)(3)(A)(v).^[6]

YouTube removed the video and Lenz countered, saying her video was not a violation. Lenz claimed that Universal’s takedown notice was a misrepresentation in violation of Section 512(f) as her use of the song was not the focus of the video.^[7] In response, Universal argued that copyright owners should not have to consider fair use before sending DMCA takedown notices.^[8] The Ninth Circuit rejected this argument and held that the DMCA requires copyright owners to consider whether works are lawful fair uses before sending takedown notices.^[9] The Ninth Circuit further added that copyright holders should be held to a subjective standard.^[10] A rightsholder that sends a notice that ends up being false due to fair use could still be excused from liability, so long as they subjectively believed that the material they targeted was infringing, even if the court ultimately disagrees with their position.^[11]

The Report’s Discussion of Lenz

The Report notes that several participants at a roundtable addressed this interpretation of the good faith requirement and questioned the practical application of the court’s determination that a copyright owner must evaluate whether a use is permitted by the fair use doctrine and affirmatively decide that it is not before sending a takedown notice.^[12] The result places “potential liability on rightsholders who fail to undertake a fair use inquiry before sending a takedown notice, without regard to whether or not the material is actually infringing.”^[13] The Report argues that instead, based on the language of the statute, section 512(f) should look at whether the rightsholder “knowingly materially misrepresent[]” that “the material or activity is infringing” (or, for counter-notices, that the “material or activity was removed or disabled by mistake or misidentification”).^[14] And therefore, to find a rightsholder liable under section 512(f), a court should first determine whether or not the use is, in fact, infringing, and if not, “whether the copyright owner made the misrepresentation [regarding the infringing nature of the material” ‘knowingly.’”^[15] The Office suggested that Congress monitor how the courts apply Lenz, and that Congress should consider clarifying the statutory language if needed.^[16]

The report adds that a “number of rightsholders were uncertain about implication of Lenz for their ability to use automated processes to identify infringing material and send takedown notices.”^[17] One OSP asserted that, under Lenz, “automated notices should not be considered valid notices, in part because algorithms that generate automated notices are not able to assess whether a particular use is infringing or might be lawful,” since “a conclusion [on fair use is one] that is impossible for an algorithm to draw.”^[18] Several rightsholders disagree, arguing that automated programs, assisted by human review either in design or execution, do provide the appropriate level of review to meet the notice requirement.   

Can Computers Make a Good Faith Determination of Fair Use?

The Ninth Circuit did not resolve the AI issue in Lenz.  The Court’s original opinion initially discussed automated scanning for infringement, but this section was entirely cut out in the Court’s amended opinion.  The discussion of automated infringement notification services that appeared in the opinion pre-amendment is as follows:

We are mindful of the pressing crush of voluminous infringing content that copyright holders face in a digital age. But that does not excuse a failure to comply with the procedures outlined by Congress...

We note, without passing judgment, that the implementation of computer algorithms appears to be a valid and good faith middle ground for processing a plethora of content while still meeting the DMCA’s requirements to somehow consider fair use. Cf. Hotfile, 2013 WL 6336286, at *47 (“The Court . . . is unaware of any decision to date that actually addressed the need for human review, and the statute does not specify how belief of infringement may be formed or what knowledge may be chargeable to the notifying entity.”).

For example, consideration of fair use may be sufficient if copyright holders utilize computer programs that automatically identify for takedown notifications content where: “(1) the video track matches the video track of a copyrighted work submitted by a content owner; (2) the audio track matches the audio track of that same copyrighted work; and (3) nearly the entirety . . . is comprised of a single copyrighted work.”^[19]

The Ninth Circuit initially contemplated the use of automated services for first level identification of potentially infringing material but added that “Copyright holders could then employ individuals … to review the minimal remaining content a computer program does not cull” to perform a fair use analysis.

However, since the Amendment to Lenz struck this section, there is no clarity as to how much consideration must be given to fair use and the propriety of using automated services. This poses an unknown query regarding the use of commercially available monitoring or notice companies, implicates whether Artificial Intelligence can be used both for monitoring or determining which infringements to send notices for, and leaves open the level of review required to comply with the Ninth Circuit’s interpretation of evaluation under the “fair use” requirement.

Post-Lenz cases have not offered sufficient guidance on the issue. Courts citing Lenz hold fast to the fair use standard established in Lenz but note that “[t]he fair use factors in Section 107 are not intended to be applied in an isolated and mechanical way. They should be explored and weighed together in light of copyright’s purpose” and that “[e]very application of fair use is different, and the inquiry must be made on the specific facts before the Court on a case-by-case basis”.^[20] What commonly occurs is a deep dive into the adequacy of the pleading surrounding proof of evaluating “fair use”.^[21]

Proponents of using AI to facilitate monitoring and take downs argue that this imposes a heavy burden on rights holders.^[22] Using third party or automated monitoring services assists locating the myriad of possible infringers, however, adding a second level of human review to evaluate for fair use has additional associated costs. The problem is exacerbated by the fact that often even when a listing is removed by a service provider, the infringing use just pops up again elsewhere – a never ending whack-a-mole problem. On the other hand, opponents argue that the current DMCA take down system works fine, despite the large numbers of requests Online Service Providers receive.^[23] Opponents, however, cite that automated take down notices often increase the number of notices and can inaccurately flag content that isn’t infringing or at times doesn’t depict the copyrighted work at all.^[24] Thus, some of these automated notices result in perfectly legal content being removed.^[25] The Electronic Frontier Foundation is one of the many critics of automated takedown notices and notes that “[t]he use of robots, without any human review, simply cannot satisfy” the good faith belief standard as “whether a use of copyrighted material constitutes a fair use protected by federal copyright law is often a question only a human can answer, after taking into account the context and purpose of the speech in question.”^[26]

It’s worth noting that humans aren’t always good at determining fair use either. Hundreds of cases illustrate that fair use is complex enough to warrant thorough opinions, some later reversed by appellate courts, and even reaching the Supreme Court of the United States.^[27] One of the first examples of fair use in copyright in this country’s courts can be seen in Justice Story’s opinion in Folsom v. Marsh, 9 F.Cas. 342 (No. 4901) (C.C.D.Mass.1841). Justice Story noted that “what constitutes a fair and bona fide abridgement, in the sense of the law, is one of the most difficult points, under particular circumstances, which can well arise for judicial discussion.”^[28] Justice Story characterized copyright cases as approaching “the metaphysics of the law, where the distinctions are, or at least may be, very subtle and refined, and, sometimes, almost evanescent.”^[29] Even the leading Treatise by Nimmer notes that the statute offers “no guidance as to the relative weight to be ascribed to each of the listed factors,” and, in the end, “courts are left with almost complete discretion in determining whether any given factor is present in any particular use.” Nimmer on Copyright § 13.05 (footnotes omitted). These examples illustrate that the evaluation of fair use is a difficult one for humans, let alone the current state of technology and AI.

The future of improvements in Artificial Intelligence and Machine Learning and law surrounding their use is dynamic and unpredictable. It is quite possible that the future will provide a system where AI is able to make a reasonable fair use determination for the purpose of a take down notice and this could provide service providers and rightsholders with a simpler and more cost effective method for reducing infringement. Once technology reaches that point, Congress may have to step in, yet again, to revise the DMCA to stay relevant to the new technology.

Take Aways

At the moment, the law isn’t clear whether a machine learning algorithm that can distinguish certain uses and transformations of copyrighted work to identify whether there is fair use, would be sufficient under the Ninth Circuit’s standard. A safe position to take is that AI could further assist in initial review of potentially infringing use to further cull the potential uses that a human would still review prior to deciding to send a takedown notice. It will be interesting to watch how courts continue to apply the “good faith belief” standard and evaluation of fair use, as well as the potential for case law to develop around the use of machine learning as part of the identification and take down notice process.   

It will also be interesting to see how the law develops with respect to infringing conduct posted and/or hosted by AI systems. Thus far, most cases that do not find Safe Harbor protection for service providers relate to human curation activities. However, as AI and algorithms are trained and improved to make decisions in place of human review, liability may also change. At what point is liability imparted on a company posting potentially infringing material when an algorithm selects the material that would otherwise be done by a human? Machine Learning typically begins with teaching an algorithm, much like another person, to select what is desired. At what point is the machine responsible for making a selection that mimics human behavior? These questions are only beginning to circulate in litigation as this area of law and technology evolves.