# An Empirical Study of U.S. Copyright Fair Use Opinions Updated, 1978-2019

**Authors:** Barton Beebe
**Citation:** "An Empirical Study of U.S. Copyright Fair Use Opinions Updated, 1978-2019," 10 *NYU Journal of Intellectual Property and Entertainment* 1 (2020)
**Source:** https://jipel.law.nyu.edu/wp-content/uploads/2020/12/JIPEL-Beebe-Fall-2020.pdf

## INTRODUCTION

*p. 5*
Section 107 of the Copyright Act of 1976 sets forth four factors that courts "shall" consider in determining whether an unauthorized use of a copyrighted work qualifies as a non-infringing "fair use."! The four-factor test is central to the operation of American copyright law and the system of commercial and artistic speech that it regulates. Since the January 1, 1978 effective date of the Copyright Act, the test has been the subject of four Supreme Court cases? and perhaps a fifth one this term in Google LLC v. Oracle America, Inc.* Over the decades, the four-necessarily reflect or influence everyday fair use adjudication, particularly in the district courts.'

*p. 7*
The updated data show that much has remained the same in our fair use case law since 2005. The same two factors that drove the test through 2005 have continued to do so: the first factor, going to the "purpose and character" of the defendant's use,!° including whether it qualifies as "transformative,"'' and the fourth factor, going to "the effect of the [defendant's] use upon the potential market for or value of the copyrighted work.''!? As between these two factors, the updated data indicate that at least in the general population of our fair use case law, factor four has remained the single dominant factor in courts' adjudication of the fair use defense-economic analysis continues ultimately to define fair use in the American copyright system.'? Meanwhile, appeal, reversal, and dissent rates remain unexceptional as compared to other areas of private law.'* Courts continue to apply the four-factor test mechanically and rarely consider additional factors beyond the four that are prescribed by the statute.'> The lower courts continue mistakenly to cite old, obsolete dicta from the Supreme Court case law, dicta that the Court itself has repeatedly tried to overwrite.'° Admirably, however, judges continue to resist stampeding all the factor outcomes to conform with the overall test outcome; instead, " It cannot be stressed strongly enough, however, that the statistics set forth in this updateand indeed in the original article-cannot present a complete picture of our fair use case law, let alone of the larger operation of copyright fair use outside of the case law. The data are a necessary supplement to, but they cannot substitute for, traditional systematic doctrinal analysis. At best, each statistic is an additional form of circumstantial evidence, a description of a different part of the elephant, that may be adduced to develop a better understanding of our overall fair use case law. See Neil Weinstock Netanel, Making Sense of Fair Use, 15 LEWIS & CLARK L. REV. 715, 732-34 (2011) (discussing the problem of selection bias and defending the utility of systematic content analysis of reported federal opinions); Matthew Sag, Predicting Fair Use, 73 OHIO ST. L.J. 47, 83 (2012) ("The potential for selection bias notwithstanding, the fact remains that litigated cases are important and they are constantly subject to ad hoc empirical assessments. . . . The unrepresentative nature of written opinions does not imply that scholars should abandon the field of empirical legal studies, but it does call for some caution in interpreting results."); see also Beebe, supra note 6, at 565-66 (discussing selection bias). 1017 U.S.C. § 107(1) (2018). 'l See Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569, 578-85 (1994) they generally continue to admit in their written opinions when a factor does not support their overall fair use finding.'' But the updated data also show some notable developments in the case law since 2005. A substantially higher proportion of our fair use case law now takes the form of motion to dismiss and summary judgment opinions.'® There has also been a significant rise in the proportion of opinions addressing unauthorized uses of photographs.'? Among the lower courts, the courts of the Second Circuit still exert the most influence on the fair use case law, but that influence 1s declining.*? More substantively, while the transformativeness test appeared to be waning in influence by 2005, it has since recovered its previous level of influence, even in the lowerprofile, workaday fair use opinions that make up the majority of the data."! However, the data indicate that in these same opinions, while courts now routinely consider transformativeness, a finding that the defendant's use is transformative is still far from necessary to trigger a finding of fair use. Notwithstanding the provisions of sections 106 and 106A, the fair use of a copyrighted work, including such use by reproduction in copies or phonorecords or by any other means specified by that section, for purposes such as criticism, comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research, is not an infringement of copyright. In determining whether the use made of a work in any particular case is a fair use the factors to be considered shall include-(1) the purpose and character of the use, including whether such use 1s of a commercial nature or is for nonprofit educational purposes;

*p. 9*
(3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and *4 Interestingly, however, 3.3% of the opinions (or 19 of the 579 opinions) made an adverse inference that to the extent the defendant's use did not take the form of "'criticism, comment, news reporting, teaching .. . , scholarship, or research," this supported a determination that the use did not qualify as a fair use. See, e.g., Peteski Prods. v. Rothman, 264 F. Supp. 3d 731, 739-40 (E.D. Tex. 2017) ("Here, [defendant] made no alteration to the work nor did she use it as part of a commentary or criticism."). > See, e.g., Swatch Grp. Mgmt. Servs. v. Bloomberg L.P., 742 F.3d 17, 31 (2d Cir. 2014), amended and superseded by 756 F.3d 73 (2d Cir.). Three of these eight opinions nevertheless found no fair use. See, e.g., Monge v. Maya Magazines, Inc., 688 F.3d 1164, 1168(9th Cir. 2012). *6 Harper & Row Publishers, Inc. v. Nation Enters., 471 U.S. 539, 560 (1985). reported that 17.0% of the 306 opinions through 2005 considered other factors;7' since 2006, that proportion has dropped to 4.7%.*8 The original drafters of section 107 were concerned that the section would freeze what had up to then been an evolving judge-made doctrine."? The data show this concern to have been valid.

*p. 10*
As for the manner in which courts have used the factors, courts continue to apply them "mechanistically,'*° perhaps even more so than they did in the past. The original study reported that 59.5% of the 306 opinions through 2005 adopted the practice of explicitly stating which party, if either, each factor favored;+' since 2006, 77.0% of fair use opinions have done so. Overall, of the 579 opinions in the updated dataset, 87.5% considered all four factors, and 85.9% did so in order. (As in the past, this all greatly facilitated reliable coding of the opinions.) *9 See Beebe, supra note 6, at 559. 30 Harper & Row, 471 U.S. at 588 (Brennan, J., dissenting). 31 Beebe, supra note 6, at 562. 3? Of the 435 cases that produced these opinions, 336 produced only one opinion. By contrast, five cases produced five opinions each. For the final majority opinion in each of these cases, see Harper & Row, 471 U.S. 539;New Era Publications International, APS v. Henry Holt, Co., 884 F.2d 659 (2d Cir. 1989);Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994);Princeton University Press v. Michigan Document Services, Inc., 99 F.3d 1381(6th Cir. 1996); Bouchat v. Baltimore Ravens Ltd. Partnership, 737 F.3d 932 (4th Cir. 2013). This study coded the United States Court of Federal Claims as a district court.

*p. 10*
33 The rapid increase in district court opinions has meant that for the first time since 1992, there were more federal court opinions on copyright fair use in both 2018 and 2019 than there were law review articles focused on the issue in each of those years. The ratio of court opinions to law review articles in 2018 was 1.11 and in 2019, 1.03. A second explanation for the increase in district court opinions since 2010 may emerge out of the data represented in Table 1 andFigure There has also been a significant rise in opinions addressing motions for summary judgment on the fair use defense, suggesting that litigants and courts have become more comfortable in the past decade with addressing the defense in the consisted of opinions addressing motions for summary judgment on the fair use issue; more specifically, 52.5% of fair use circuit court opinions did so. of the opinions sampled (regardless of posture and disposition on appeal) found fair use. Because 99 of the 435 cases included in the dataset were responsible for more than one opinion, a better statistic may be the fair use win rate for the last-in-time, non-concurrence, non-dissent opinion in each particular case, which I will refer to as a case's final opinion. Of these 435 final opinions, 38.2% found fair use. Table 3 below details the distribution and fair use win rates of these opinions by posture. As they did in the original study, the fair use win rates for preliminary injunction and bench trial opinions in Table 3 fall significantly below 50%, with, for example, only one in four final bench trial opinions in the dataset finding fair use. I had suggested in the original study that the low fair use win rates in these postures may be the result of cases in which the defendant pled a relatively weak or even frivolous fair use defense because it 1s inexpensive to do so.°? Good judges will dutifully work through the four-factor analysis even for the most ridiculous claims of fair use, and so such opinions are for better and worse included in the dataset.

*p. 17*
The original study introduced one viable (though hardly flawless) method of filtering out opinions addressing a frivolous fair use defense, which is to filter out those opinions that devoted a small proportion of their overall word count to the fair use analysis.* For the 435 final opinions in the updated dataset, there is a moderate >? See Beebe, supra note 6, at 575-80. fair use win rate requires filtering only for opinions that devoted at least 40% of their word count to the fair use analysis.*° This is all to say that even more so than in the original dataset, the updated dataset shows that a significant majority of our fair use case law-and precedent-consists of analyses that found no fair use.

*p. 18*
Overall, the updated data show no statistically significant variations among the circuits in fair use win rates. °° Specifically, the fair use win rate was .450 for the 60 opinions whose fair use analysis consisted in word count of at least 40% of the opinion's overall word count.

*p. 18*
>" See Beebe, supra note 6, at 594 n.142; Sag, supra note 9, at 81-82 (discussing intercircuit differences in fair use win rates).

*p. 19*
the opinions are filtered. For example, Figure 4 reports five-year moving averages of the proportion of opinions by year that found fair use for all 579 opinions in the dataset and for the subset of the 314 final opinions in the dataset that dispositively found fair use or no fair use (rather than outstanding fact issues). The former trend is basically flat. The latter suggests some uptick in fair use outcomes since 2000 (starting notably before the Supreme Court case Campbell v. Acuff-Rose in 1994).

*p. 19*
But overall, the data are inconclusive. judges tended to call the factors as they saw them and openly admitted when a factor did not support their overall determination."

*p. 24*
The data continue to support this view of the case law. Figure 6 A. An Overview of the Significance of the Subfactor Findings

*p. 25*
To help in understanding what role each of the various subfactor considerations play in the overall fair use analysis, Table 5 presents the results of various logistic regression models estimated for all opinions that were dispositive (i.e., either found fair use or no fair use), were unreversed, and in the case of appellate opinions were neither dissents nor concurrences. I will refer to such opinions, for lack of a better term, as core opinions.' The models regress a binary dependent variable representing whether the opinion found fair use on various binary independent variables representing underlying factual findings, whether the opinion came out of the courts of the Second or the Ninth Circuits, and, following Sag,' whether the plaintiff or defendant was a natural person. Table 5 reports odds ratios, which are not particularly easy to understand. For each independent variable, the odds ratio reports the ratio of the odds of a finding of fair use to the odds of a finding of no fair use when the condition the independent variable represents is satisfied and all other variables are held constant.''? Though the regression models are admittedly rough, the odds ratios can at the very least provide a sense of the relative impact of various factual findings and other variables on the overall fair use determination. 16 See Sag, supra note 9, at 74-78. 7" For example, according to model | in Table 5 (which omits judge gender, circuit, and natural person status of each party), when a court finds that the defendant's use is transformative, the odds of a finding of fair use are 91.3 times as large as the odds of a finding of no fair use. Expressed as predicted probabilities under model 1, the predicted probability that a court will find fair use when the use does not qualify as transformative is .42 (with all other variables set to 0) as against a predicted probability of .99 when the use qualifies as transformative. [Vol. 10:1 finding. Additionally, once we control for the factual characteristics of the cases, the impact of gender is only marginally statistically significant. Finally, consistent with Sag's work, the fact that the owner of the copyrighted work 1s a natural person (rather than, for example, a corporation) appears to be associated with a significant increase in the likelihood that a court will find fair use.' transformative, the ratio of the odds a defendant will prevail in its fair use defense to the odds it will fail is anywhere from 86 to 91 times greater. By this measure, a finding of transformativeness exerts by far the greatest impact of any finding on a court's likelihood of making an overall determination of fair use. More simply, as Consistent with the original study, a finding of transformativeness still does not appear to stampede the factors. The mean stampede score for the 78 core opinions since Campbell that found transformativeness was 1.99, with the mode stampede score of 2 reported by 38.5% of the opinions. As the original study explained, this makes sense because transformative uses most commonly target works classified as creative under factor two and often involve quite substantial uses of those works under factor three.®' At the same time, however, a finding of transformativeness correlates very strongly with a finding that the defendant's use will have no substantial adverse effect on the potential market for the plaintiff's work under factor four. In the 78 core opinions since Campbell that found that the defendant's use was transformative, only six (7.7%) found that the defendant's effect on the plaintiff's mark disfavored fair use under factor four."

## I will discuss the results reported in

*p. 31*
A subset of transformativeness cases consists of cases in which the defendant made a parody of the plaintiff's work. In nearly all of these, the defendant's conduct was deemed to be transformative and a fair use. Overall, the updated dataset consists of 35 opinions from 26 cases in which the court explicitly found that the defendant's work constituted a parody. In all but three of these cases, the defendant's use was ultimately found to be a fair use,°! and the last time a court found a parody not to be a fair use was 1988, six years before Campbell." Of the 23 cases in which the courts ultimately found the parody to be a fair use, 20 involved parodies deemed to be commercial in nature. Even more so than generally transformative works, the species of such works that qualify as parodic are especially privileged under factor one and the overall four-factor fair use analysis.

## Commerciality

*p. 31*
Though transformativeness continues to attract the bulk of scholarly attention, the commerciality of the defendant's use remains the subfactor that courts most consistently invoke in their factor one analysis, as the black line in Figure 7 shows. °2 See New Line Cinema Corp., 693 F. Supp. 1517. See also Samuelson, supra note 28, at 2550

*p. 31*
(noting that after Campbell, "every subsequent parody case has been adjudged a fair use').

*p. 32*
Of the 545 fair use opinions following the 1984 Supreme Court majority opinion in Sony, which put substantial weight on the commerciality or noncommerciality of the defendant's use,"? 85.3% have explicitly addressed the commerciality subfactor. The regression results further show the substantial effect that the commercial or noncommercial status of the defendant's use has on the overall outcome of the test, though the effect is not nearly on the order of that exerted by a finding of transformativeness. While transformative uses nearly always qualify as fair uses, noncommercial uses may fail to so qualify. Of the 47 core opinions since Sony that found that the defendant's use was noncommercial in nature, a respectable minority of 11 opinions (23.4%) found overall that it was nevertheless not a fair use.

*p. 32*
Meanwhile, commercial uses may occasionally qualify as fair uses. Of the 233 core opinions that found that the defendant's use was commercial in nature, 36.9% found the use to be a fair use.

*p. 32*
Though there is thus some flexibility in the relation between the commerciality or noncommerciality of the use and a court's determination of whether it qualifies as a fair use, there remains a continuing problem in the case law under the factor one commerciality analysis, which is shown by the dashed line in Figure 7. As the original study detailed, the Sony majority opinion in 1984 quite explicitly established a presumption under factor one that a use "to make copies for a commercial or profit-making purpose . .. would presumptively be unfair," while the "contrary presumption" would apply to noncommercial uses. Fair use opinions continue routinely to denigrate factor two as unimportant to the overall fair use analysis,'" and the updated data support the view that the factor typically has a relatively minimal impact on the outcome of the overall four-factor test. In general, courts rarely find that the factor supports a finding of fair use.!TM Nevertheless, the data suggest that certain findings under both of factor two's subfactors-whether the work is creative or factual and whether the work is published or unpublished-continue to have an at least statistically significant effect on a court's overall determination. The regression results reported in Table 5 indicate, as expected, that the odds a court will find fair use increase when the plaintiff's work is factual in nature and decrease when it 1s creative in nature. More interestingly, and consistent with the original study and the irony it noted,!® while the unpublished nature of the plaintiff's work has no apparent impact on a court's overall fair use determination, the fact that a work is published appears to increase the odds that a court will find fair use. Specifically, the ratio of the odds of a finding of fair use to a finding of no fair use more than doubles when the court finds that the plaintiff's work was published.' D. Factor Three: Amount and Substantiality of the Use Overall, the status of factor three has not changed appreciably since the original study.!°' Its outcome continues to correlate very strongly with the overall test outcome, particularly when factor three is found to favor the defendant. In the 101 core opinions in which the court found that factor three favored a finding of fair use, all but two found in favor of fair use overall.!° However, in contrast to the findings of the original study, Table 5 suggests that the fact that the defendant copied the entirety of the plaintiff's work no longer appears to significantly impact a court's overall fair use determination. In the updated data, of the 148 core opinions 1n which 105 See Beebe, supra note 6, at 614 ("[T]he Supreme Court sought to establish that a certain finding (here, that the work is unpublished) disfavors fair use. Lower courts appear not to have acted on that dictum, however, other than to invert it to conclude that the opposite of that finding (here, that the work is published) favors fair use.").

*p. 34*
106 Stated in terms of predicted probabilities, the regression results under model | suggest that the predicted probability that a court will find fair use when the work does not qualify as published is .42 (with all other variables set to 0) as against a predicted probability of .65 when the work qualifies as published.

*p. 34*
'07 For a thorough quantitative analysis of the role of the third factor in the overall fair use the court found that the defendant's use copied the entirety of the plaintiff's work, a reasonably balanced 58.8% found no fair use and 41.2% found fair use.!

*p. 35*
One trend that has become clearer with the updated data is the increasing degree to which courts explicitly assess under factor three whether the defendant's use of the plaintiff's work was necessary to the defendant's purpose, and even if it was, whether the defendant used more than was necessary to accomplish that purpose.'!® Figure 8 estimates a five-year moving average of the proportion of opinions by filing year approaching factor three in this manner. Though this gradual shift toward a necessity requirement could be understood to benefit plaintiffs, the data show no strong correlation between a court's invocation of necessity and either the outcome of factor three or the overall outcome of the fair use test.

## E. Factor Four: Effect on the Market

*p. 36*
The scholarly literature continues to debate whether the leading factor in courts' fair use analysis is currently factor one, with its focus on the "purpose and character" of the defendant's use, or factor four, with its focus on the effect of the defendant's use on the "potential market for or value of" the plaintiff's work.

*p. 36*
Looming in the background of this debate is the larger question of whether courts are generally pursuing a more defendant-friendly, purpose-centered approach to fair use or a more plaintiff-friendly, economic approach to fair use. Most scholars endorse the former approach, believing it to be more supportive of the public domain. This debate has largely relied on the latest headline-making cases. The updated data suggest that with respect to our fair use case law as a whole, beyond the headline cases, the status of factor four has not appreciably changed over the past three or four decades.'!® It continues to be the factor whose outcome correlates most strongly with the overall test outcome, as Table 4 indicates. Of the 169 core opinions that found that factor four disfavored fair use from 1978 through 2019, all but three ultimately found no fair use, and none of the three outlying opinions that found fair use offers particularly compelling analysis to explain its divergence between factor four and the overall outcome.''' Meanwhile, of the 154 opinions that found that factor four favored fair use, all but nine found fair use. A majority of these nine outlying opinions ruled that factor four favored the defendant because there was no market for the plaintiff's work As with factor one, the very strong correlation between the outcome of factor four and the overall test outcome prompts the question of which subfactor factual findings drive the outcome of factor four itself. The problem with factor four, however, 1s that historically courts have not developed any subfactor factual findings under it. Instead, as I argued in the original study, courts typically treat factor four as essentially a "metafactor" in which they integrate their analyses of the preceding three factors.''? In doing so, they balance the justification for the defendant's use of a work against its effect on the plaintiff's economic incentives to create and further exploit that work. Crucially, however, when courts engage in this balancing test in the analytical space provided by factor four, courts do so in economic terms, within the wheelhouse of law and economic analysis. By contrast, had the four-factor test been designed so that the analysis of the justification for the defendant's use came fourth, in the cleanup position, one imagines that outcomes might be different in close cases.

*p. 37*
Recent scholarship has brought to light one important, more specific function that factor four plays-or should play.!"° In instructing courts to assess "the effect of the use upon the potential market for or value of the copyrighted work," factor four requires courts to define the limits of the "potential market" and "value" that the copyright owner should have the exclusive right to exploit. Thus, for example, courts have established under factor four that the owner should not have the exclusive right to exploit the market for harsh reviews of its work or for parodies that ridicule the work.!'?! But here too factor four plays a largely synthetic role. Under it, the first three factors aid the court in determining whether the defendant's use falls within the category of uses that should be reserved exclusively to the copyright owner as a matter of copyright policy or simply of industry custom. !" '19 See Beebe, supra note 6, at 617-18. Indeed, regressing the outcome of factor four (rather than the overall test outcome) on the various subfactor findings and other objective factors listed in Table 5 results in largely the same odds ratios as those reported in 1899, 1930-45 (2007). !24 Harper & Row Publishers, Inc. v. Nation Enters., 471 U.S. 539, 566 (1985). '25 Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569, 578 (1994) ("All are to be explored, and the results weighed together, in light of the purposes of copyright."').

*p. 40*
In the spirit of the original study, this brief update has focused on the mass of everyday copyright fair use cases. On their own, most individual fair use cases may not seem to be particularly important except to the parties involved, if even to them.

*p. 40*
However, taken together, they form a body of case law whose patterns and characteristics reveal the daily life of our fair use case law, a daily life that is in some ways surprisingly different from the life of the "leading cases." The lower courts sometimes ignore higher court cases or are simply unaware of them. Old, ill-advised dicta can keep cropping up like perennial weeds even decades after efforts to eradicate them. But perhaps most surprisingly, when viewed as a whole, the fair use case law presents itself as far more stable and predictable-and unchanging-than the headline-making cases might suggest.

*p. 42*
Finally, JIPEL is providing the underlying data and data key used by Professor

*p. 42*
Beebe in his analysis. The professor encourages feedback and collaboration and has agreed to this data sharing full-heartedly.

*p. 42*
(1) Professor Beebe's underlying data coding (excel)

*p. 42*
We acknowledge that this method is not suitable for every empiricallyfocused article. Nevertheless, we believe that making it available may help other journals move forward along the path toward adopting more rigorous and standardized review for the underlying data and assumptions in empirical legal works.?

## Footnotes

> district court opinions in 2018, 18 (or 72.0%) involved the copying of photographs; of the 27 district court opinions in 2019, 14 (or 51.9%) did so.*4

> *Courts with no fair use opinions during this period are excluded 5-YEAR MOVING AVERAGE OF ANNUAL PROPORTION OF DISTRICT AND CIRCUIT COURT FAIR USE OPINIONS COMING FROM THE DISTRICT AND CIRCUIT COURTS OF THE SECOND

> datasets, I coded their characteristics (other than their JCS scores) by hand. The data were combined and analyzed in Stata, a common statistical analysis software program. 63 See Beebe, supra note 6, at 582-85. doing

> 6° See Campbell v. Acuff-Rose Music, Inc, 510 U.S. 569, 578-85 (1994). 10 See Netanel, supra note9, at 736-38. to

> indicates, of the 78 core opinions since Campbell in which a court found transformativeness, in all but three the court went on to find fair use. Tellingly, in each of the three outlying opinions, the court took pains to minimize the significance of its transformativeness finding.*® 1 See generally Jiarui Liu, An Empirical Study of Transformative Use in Copyright Law, 22 STAN. TECH. L. REV. 163 (2019) (reviewing all federal court fair use opinions addressing transformativeness through January 1, 2017); R. Anthony Reese, Transformativeness and the transformative'); Castle Rock Ent. v. Carol Publ'g Grp., 955 F. Supp. 260, 272 (S.D.N.Y. 1997) (characterizing the use as transformative "only by a generous understanding of what it means for a work to be 'transformative''').

> C. Factor Two: Nature of the Copyrighted Work

> Questionable Use of Custom in Intellectual Property, 93 VA. L. REV.