Microsoft Word - 19-3-3-perel-elkin-koren-final.docx

473

ACCOUNTABILITY IN ALGORITHMIC COPYRIGHT

ENFORCEMENT*

Maayan Perel** & Niva Elkin-Koren***

CITE AS: 19 STAN. TECH. L. REV. 473 (2016)

ABSTRACT

Recent years demonstrate a growing use of algorithmic law enforcement by online

intermediaries. Facilitating the distribution of online content, online intermediaries offer

a natural point of control for monitoring access to illegitimate content, which makes them

ideal partners for performing civil and criminal enforcement. Copyright law has been at

the forefront of algorithmic law enforcement since the early 1990s when it conferred safe

harbor protection to online intermediaries who remove allegedly infringing content upon

notice under the Digital Millennium Copyright Act (DMCA). Over the past two decades,

the Notice and Takedown (N&TD) regime has become ubiquitous and embedded in the

system design of all major intermediaries: major copyright owners increasingly exploit

robots to send immense volumes of takedown requests

and major online intermediaries, in

response, use algorithms to filter, block, and disable access to allegedly infringing content

automatically, with little or no human intervention.

Algorithmic enforcement by online intermediaries reflects a fundamental shift in our

traditional system of governance. It effectively converges law enforcement and

adjudication powers in the hands of a small number of mega platforms, which are profit-

maximizing, and possibly biased, private entities. Yet notwithstanding their critical role

in shaping access to online content and facilitating public discourse, intermediaries are

* We thank Oren Bracha, Miriam Marcowitz-Bitton, Jane Ginsberg, Ellen Goodman,

Eldar Haber, Lital Helman, Ethan Katsh, Shelly Kreiczer-Levy, Edward Lee, Neil Netanel,

Gideon Pharchomovsky, Orna Rabinovich-Einy, and Tal Zarsky, as well as the participants of

the “Trust and Empirical Evidence in Law Making and Legal Process” conference at the

University of Oxford on June 19-20, 2015, and the participants of the “Openness and

Intellectual Property” conference at the University of Pennsylvania on July 22-24, 2015, for

their insightful comments. We are further grateful to Dalit Kan-Dror, Esq., and Nati Perl for

their academic assistance. This research was supported by I-CORE Program of the Planning

and Budgeting Committee and the Israel Science Foundation.

** Dr. Maayan Perel, Lecturer, Netanya Academic College; Reserach Fellow, Haifa

Center for Law & Technology, University of Haifa Faculty of Law; S.J.D., University of

Pennsylvania Law School.

*** Professor Niva Elkin-Koren, Director, Haifa Center for Law & Technology,

University of Haifa Faculty of Law.

474 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

hardly held accountable for algorithmic enforcement. We simply do not know which

allegedly infringing material triggers the algorithms, how decisions regarding content

restrictions are made, who is making such decisions, and how target users might affect

these decisions. Lessons drawn from algorithmic copyright enforcement by online

intermediaries could offer a valuable case study for addressing these concerns. As we

demonstrate, algorithmic copyright enforcement by online intermediaries lacks sufficient

measures to assure accountability, namely, the extent to which decision makers are

expected to justify their choices, are answerable for their actions, and are held responsible

for their failures and wrongdoings.

This Article proposes a novel framework for analyzing accountability in algorithmic

enforcement that is based on three factors: transparency, due process and public oversight.

It identifies the accountability deficiencies in algorithmic copyright enforcement and

further maps the barriers for enhancing accountability, including technical barriers of

non-transparency and machine learning, legal barriers that disrupt the development of

algorithmic literacy, and practical barriers. Finally, the Article explores current and

possible strategies for enhancing accountability by increasing public scrutiny and

promoting transparency in algorithmic copyright enforcement.

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 475

ABLE OF CONTENTS

I. INTRODUCTION ........................................................................................................... 475

II. ALGORITHMIC ACCOUNTABILITY ............................................................................... 478

A. The Rise of Algorithmic Enforcement by Online Intermediaries ........................................478

B. Accountability of Online Intermediaries Engaging in Algorithmic Law

Enforcement...................................................................................................................................................481

C. Accountability Matters in Algorithmic Copyright Enforcement .........................................484

1. Accountability and the Rule of Law ..........................................................................................486

2. Accountability and the Public Sphere ........................................................................................488

3. Accountability and Copyright Policy ........................................................................................492

D. The Virtues of Accountability—A Three-Factor Framework ................................................493

III. EXPLORING ACCOUNTABILITY IN ALGORITHMIC COPYRIGHT ENFORCEMENT

SYSTEMS ....................................................................................................................... 497

A. The Deficient Accountability Standards of the DMCA ............................................................497

B. Regulated Versus Voluntary Mechanisms of Algorithmic Copyright

Enforcement...................................................................................................................................................502

1. Transparency .......................................................................................................................................505

2. Due Process ............................................................................................................................................506

3. Public Oversight ..................................................................................................................................509

C. Corporate Copyright: YouTube’s Content ID ................................................................................510

1. Transparency .......................................................................................................................................513

2. Due Process ............................................................................................................................................514

3. Public Oversight ..................................................................................................................................516

IV. ENHANCING ACCOUNTABILITY: BARRIERS AND STRATEGIES ................................... 516

A. Mapping the Barriers to Algorithmic Accountability ...............................................................517

1. Technical Barriers .............................................................................................................................517

2. Legal Barriers ......................................................................................................................................520

3. Practical Barriers ...............................................................................................................................524

B. Accountability Enhancing Strategies .................................................................................................525

1. Encouraging Public Participation...............................................................................................525

2. Watchdogs .............................................................................................................................................527

3. Intermediaries ......................................................................................................................................529

4. Regulators ..............................................................................................................................................529

V. CONCLUSION ........................................................................................................... 532

I. INTRODUCTION

[A]lgorithms take over from the messy, human process of democratic decision-

making. Citizens become beholden to them, unsure of how they work, but afraid

to disregard their guidance. This creates a sort of prison of ‘invisible barbed wire’

which constrains our intellectual and moral development, as well as our lives

more generally.

Why is certain content automatically blocked from appearing on certain

1. John Danaher, Rule by Algorithm? Big Data and the Threat of Algocracy, PHIL.

DISQUISITIONS (Jan. 6, 2014, 9:55 AM),

http://philosophicaldisquisitions.blogspot.co.il/2014/01/rule-by-algorithm-big-data-and-

threat.html [https://perma.cc/JL3L-V8T4].

476 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

online platforms? Why is the same content nonetheless approved on other online

platforms? When an online hosting facility receives a complaint of copyright

infringement, how does it evaluate the complaint? When Google announces an

anti-piracy policy that will push copyright infringers down in the rankings, who

is considered an infringer? Do online platforms consider fair uses of copyrighted

material? Detailed doctrines in copyright law are carefully designed to guide

traditional, human law enforcement agents in addressing these questions. But

increasingly, it is mostly algorithms—not humans—that enforce the rights of

with copyright doctrines? Unfortunately, we really do not know.

As evident in anecdotal reports, online algorithmic copyright enforcement is

chaotic.

It has blocked a ten-year-old boy’s self-authored original video starring

his LEGO mini-figures and garbage truck despite the fact that he used royalty-free

music.

It has also facilitated the removal of a home video uploaded by Stephanie

Lenz, which featured her children dancing in the kitchen along to Prince’s “Lets

Go Crazy” song, although the video was obviously protected as fair use.

Algorithmic copyright enforcement has also allowed governmental agents in the

U.S. Department of Homeland Security to remove a conspiracy video about

President Obama.

Why would the U.S. Department of Homeland Security issue a

takedown if it can’t really own a copyright? Content restrictions of this sort may

be censoring legitimate speech. Yet because we do not know what criteria

enforcement algorithms employ to determine online copyright infringement, it is

largely impossible for us to assess their practices. Unable to understand the

practices of law enforcement algorithms, we cannot further challenge and correct

their flaws.

2. In a recent complaint against Google Inc., Viacom Inc., and several others, the

plaintiff, Benjamin Ligeri, alleged that defendants’ use of Content ID unlawfully restricted his

content, which he claimed to be protected as fair use. Among other contentions, he argued that

Content ID is an opaque and proprietary system where the accuser can serve as the judge, jury

and executioner. Content ID allows individuals, including Defendants other than Google, to

steal ad revenue from YouTube video creators en masse, with some companies claiming content

they don’t own deliberately or not. The inability to understand context and parody regularly

leads to fair use videos getting blocked, muted or monetized.

Complaint at 4-5, Ligeri v. Google, Inc., No. 1:15-cv-00188-M-LDA (D.R.I. May 7, 2015).

3. Dineen Wasylik, Take Down Abuse: From Harry Potter to LEGOs, DPW LEGAL: INTELL.

PROP. & APPEALS (Feb. 7, 2014), http://ip-appeals.com/take-down-abuse-from-harry-potter-

to-legos [https://perma.cc/D5QP-8C7G].

4. Lenz v. Universal Music Corp., 572 F. Supp. 2d 1150, 1151-52 (N.D. Cal. 2008). The

uploaded video was a twenty-nine second recording in which Prince’s song was heard playing

in the background. Stephanie Lenz uploaded the video to YouTube to share it with her friends

and family. This use of Prince’s copyrighted material was fair because it was completely non-

commercial, it used only a small portion of the song, and it did not replace the original song in

any manner that could affect the potential market for Prince’s original song.

5. Mike Masnick, Homeland Security Issuing Its Own DMCA Takedowns on YouTube to Stifle

Speech, T

ECHDIRT (Aug. 1, 2012, 10:06 AM),

https://www.techdirt.com/articles/20120720/02530219774/homeland-security-issuing-its-

own-dmca-takedowns-youtube-to-stifle-speech.shtml [https://perma.cc/K8TE-UHXC].

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 477

in the early 1990s, conferring safe harbor protection to online intermediaries who

removed allegedly infringing content upon notice under the Notice and

Takedown (N&TD) procedure designed by the Digital Millennium Copyright

Act

(DMCA). Over the past two decades, N&TD has become ubiquitous and

embedded in the system design of all major intermediaries. To confront the

immense volume of takedown notices sent by copyright owners

—many of which

are sent simultaneously and automatically by robots that scan the web for

allegedly infringing content, major online intermediaries use algorithms to filter,

block, and disable access to allegedly infringing content automatically, with little

or no human intervention.

Major platforms, such as Google, Facebook, and

Twitter, thereby engage in algorithmic copyright enforcement on a daily basis,

applying various algorithms to perform qualitative determinations, including the

discretion-based assessments of copyright infringement and fair use.

Algorithmic copyright enforcement carries some obvious advantages. It is

often more efficient, saving the cost of hiring staff and renting office space.

may further ensure consistency in applying legal doctrines and eliminate the

hassle of human review.

While online intermediaries occasionally use

algorithms to automatically implement the DMCA safe harbor provisions, some

go beyond the statutory requirements, using voluntary measures to further block

the distribution of infringing materials before they become available online.

classic example is YouTube’s Content ID, which can automatically block access to

online content ex ante, and not upon receiving a particular notice of copyright

6. Digital Millennium Copyright Act, 17 U.S.C. § 1201 (2012).

7. In 2012, Google’s 441,370 notices contained over 54 million individual takedown

requests. In 2013, the company processed over 230 million takedown requests. In 2014, it

processed 345 million requests. In 2015, Google received over half a billion removal requests.

See Daniel Seng, The State of the Discordant Union: An Empirical Analysis of DMCA Takedown

Notices, 18 VA. J.L. & TECH. 369, 444, 460-61 (2014); Ernesto, Google Asked to Remove 558 Million

“Pirate” Links in 2015, T

ORRENTFREAK (Dec. 30, 2015) https://torrentfreak.com/google-asked-

remove-558-million-pirate-links-2015 [https://perma.cc/S3C9-HC2Y]; Ernesto, Google

Discarded 21,000,000 Takedown Requests in 2013, TORRENTFREAK (Dec. 27, 2013),

https://torrentfreak.com/google-discarded-21000000-takedown-requests-in-2013-131227

[https://perma.cc/UZ7P-PKQB]; Joe Mullin, Google Handled 345 Million Copyright Takedowns in

2014, A

RS TECHNICA (Jan. 6, 2015, 1:05 PM), http://arstechnica.com/tech-

policy/2015/01/google-handled-345-million-copyright-takedowns-in-2014

[https://perma.cc/Y42Y-ZQVS].

8. JENNIFER M. URBAN ET AL., NOTICE AND TAKEDOWN IN EVERYDAY PRACTICE 27-28

(2016), https://assets.documentcloud.org/documents/2779722/SSRN-id2755628.pdf

[https://perma.cc/V7TP-J9GP].

9. See infra Part II.C.1.

10. See Thomas H. Davenport & Jeanne G. Harris, Automated Decision Making Comes of

Age, 46 MIT

SLOAN MGMT. REV. 83, 84 (2005).

11. Danielle Keats Citron, Technological Due Process, 85 WASH. U. L. REV. 1249, 1252-53

(2008).

12. See infra Part III.B (highlighting the distinction between regulated mechanisms of

algorithmic copyright enforcement and voluntary mechanisms of algorithmic copyright

enforcement). One example of an automated system implementing the DMCA’s N&TD is

DMCANotice.com. See DMCAN

OTICE.COM, http://www.dmcanotice.com

[https://perma.cc/5UM4-GKP6].

478 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

infringement from the copyright owner.

Despite these efficiency-related advantages, algorithmic copyright

enforcement lacks sufficient measures to ensure that online intermediaries are

held accountable for their actions, failures, and wrongdoings.

Our analysis

shows that algorithmic implementations of N&TD, and even more so voluntary

measures applied to detect and prevent copyright infringement, fare poorly in

accountability measures. Algorithmic enforcement mechanisms are non-

transparent in the way they exercise discretion over determining copyright

infringement and fair use; they afford insufficient opportunities to challenge the

decisions they make while failing to adequately secure due process; and they

curtail the possibility of correcting errors in individual determinations of

This Article proceeds in three parts. Part II discusses algorithmic law

enforcement by online intermediaries and explains why copyright enforcement by

online intermediaries makes an interesting case study for exploring algorithmic

accountability. It further establishes a three-factor framework for assessing

accountability in algorithmic enforcement that is based on transparency, due

process, and public oversight. Part III then applies this three-factor framework to

different mechanisms of algorithmic copyright enforcement to analyze their

accountability deficiencies. Part IV maps the barriers for encouraging

accountability in algorithmic copyright enforcement. In particular, Part IV

considers the complicated and non-transparent nature of algorithms; the

unpredictable nature of enforcement by constantly evolving learning machines;

legal barriers that hinder the ability of the public to review and investigate

algorithmic copyright enforcement; and the practical failure of existing

mechanisms of public scrutiny, such as the counter notice procedure under the

DMCA. Finally, Part IV explores existing and possible accountability-enhancing

mechanisms, including watchdog initiatives, reverse engineering experiments,

voluntary transparency reports from intermediaries, and regulatory mechanisms

of mandatory disclosure.

II. A

LGORITHMIC ACCOUNTABILITY

A. The Rise of Algorithmic Enforcement by Online Intermediaries

Recent years have seen a growing use of algorithms in performing law

enforcement tasks.

Algorithmic law enforcement is becoming pervasive: for

13. See infra Part III.C.

14. Joshua A. Croll et. al., Accountable Algorithms, 165 U. PA. L. REV. (forthcoming 2017)

(arguing that traditional accountability mechanisms and legal standards have not kept pace with

technology).

15. Algorithmic law enforcement has been drawing the attention of law and technology

scholars for over a decade. See, e.g., Tarleton Gillespie, The Relevance of Algorithms, in M

EDIA

TECHNOLOGIES: ESSAYS ON COMMUNICATION, MATERIALITY, AND SOCIETY 167 (Tarleton

Gillespie, et al. eds., 2014); Ian Kerr, Digital Locks and the Automation of Virtue, in F

ROM

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 479

example, cameras are issuing speeding tickets

and GPS-enabled bracelets or

anklets fitted to offenders are tracking their locations, notifying their victims and

the police whenever the offenders enter a prohibited area.

Algorithms not only

enforce laws,

but also detect compliance with terms of use of online vendors

and impose social norms on social media. Algorithmic enforcement generally

involves large-scale collection of data by various sensors, data processing by

algorithms, and automatic performance. It can be built to scale easily, offering an

efficient means to manage, organize, and analyze today’s massive amounts of data

with uniformity and particularity, and then to structure decision-making

accordingly.

Algorithms essentially permit more complex analysis of

information through the “extensive use of data, statistical and quantitative

analysis, explanatory and predictive models, and fact-based management to drive

decisions and actions.”

Algorithmic law enforcement is particularly ubiquitous online, where

behavior is inherently mediated by computer code. Indeed, the software and

hardware that cyberspace is built of create limitations on how people can

behave,

for instance whether Internet users must enter a password to gain

access, how active must they be whilst using a specific website to remain signed

in, or the extent to which users can view data about each other as defined by their

privacy preferences. The rise of algorithmic enforcement for online behavior was

“RADICAL EXTREMISM” TO “BALANCED COPYRIGHT”: CANADIAN COPYRIGHT AND THE DIGITAL

AGENDA 247 (Michael Geist ed., 2010); LAWRENCE LESSIG, CODE: VERSION 2.0 (2006); JONATHAN

ZITTRAIN, THE FUTURE OF THE INTERNET AND HOW TO STOP IT (2008); Kenneth A. Bamberger,

Technologies of Compliance: Risk and Regulation in a Digital Age, 88 T

EX. L. REV. 669 (2010);

Citron, supra note 11; Helen Nissenbaum, From Preemption to Circumvention: If Technology

Regulates, Why Do We Need Regulation (and Vice Versa)?, 26 BERKELEY TECH. L.J. 1367 (2011);

Anjanette H. Raymond & Scott J. Shackelford, Technology, Ethics, and Access to Justice: Should an

Algorithm Be Deciding Your Case?, 35 MICH. J. INT’L L. 485 (2014); Michael L. Rich, Should We

Make Crime Impossible?, 36 H

ARV. J.L. & PUB. POL’Y 795 (2013); Danny Rosenthal, Assessing

Digital Preemption (and the Future of Law Enforcement?), 14 N

EW CRIM. L. REV. 576 (2011); Lisa

Shay et al., Confronting Automated Law Enforcement (We Robot Conference Paper, 2013),

http://www.rumint.org/gregconti/publications/201204_Shay_ALE.pdf

[https://perma.cc/JVB6-CZ3A].

16. Florida Statutes section 316.0083, known as the Mark Wandall Traffic Safety

Program, authorizes local governments to use red light cameras to enforce violations of

sections 316.074(1) and 316.075(1)(c), both of which prohibit the running of red lights. See

2013-160 Fla. Laws 9; F

LA. STAT. § 316.008(8)(a) (2011); Chris Matyszczyk, Tickets Issued Due to

Red-Light Cameras Are Illegal, Says Florida Court, CNET (Oct. 21, 2014, 1:26 PM),

http://www.cnet.com/news/tickets-issued-due-to-red-light-cameras-are-illegal-says-florida-

court [https://perma.cc/L7CC-SH65].

17. See Christine Clarridge, How GPS Bracelets Keep Track of Sex Offenders, SEATTLE TIMES

(Apr. 22, 2009, 12:00 AM), http://www.seattletimes.com/seattle-news/how-gps-bracelets-

keep-track-of-sex-offenders [https://perma.cc/7EPT-DQ66].

18. See Citron, supra note 11, at 1263-67 for additional examples.

19. Bamberger, supra note 15, at 687-89.

20. THOMAS H. DAVENPORT & JEANNE G. HARRIS, COMPETING ON ANALYTICS: THE NEW

SCIENCE OF WINNING 7 (2007) (discussing analytics specifically, rather than algorithms more

generally).

21. LESSIG, supra note 15, at 124-25.

480 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

predicted in the late 1990s by information law scholars such as Joel Reidenberg,

who described the Lex Informatica technological standards that offer

technological solutions for information policy rules,

and Lawrence Lessig, who

coined the term “code is law” to describe how algorithms can substitute for law in

regulating certain behaviors.

Yet the comprehensiveness and robustness of

algorithmic law enforcement on the Internet were hardly foreseen.

While technology has been used to aid law enforcement for many years,

from locks and speed bumps to full-body scans at airports,

algorithmic law

enforcement implemented by online intermediaries (e.g., search engines or

hosting websites) makes enforcement more robust.

Online intermediaries have

acquired an important role in managing online behavior and protecting the rights

of Internet users. They offer a natural point of control for monitoring, filtering,

blocking, and disabling access to content, which makes them ideal partners for

performing civil and criminal enforcement.

Intermediaries currently manage

and police the usage of a tremendous stream of online content pursuant to

22. Joel R. Reidenberg, Lex Informatica: The Formulation of Information Policy Rules

Through Technology, 76 T

EX. L. REV. 553, 556-58, 562 (1998) (such as the Platform for Internet

Content Selection (PICS) that was designed to accommodate different standards for content

without compromising free speech, or technological mechanisms that can anonymize

information that would otherwise be associated with specific users).

23. LESSIG, supra note 15.

24. Rosenthal, supra note 15, at 577.

25. See, e.g., X-ray Full-Body Scanners for Airport Security, GREEN FACTS (2016),

http://copublications.greenfacts.org/en/x-ray-full-body-scanners-for-airport-security

[https://perma.cc/WEU3-F5WK].

26. Other copyright enforcement algorithms employing Digital Rights Management

(DRM) or Graduate Response will hence remain outside the scope of this Article. These

enforcement algorithms largely function in a bi-directional way, limiting the target user’s access

to or use of a copyrighted work, or penalizing her by restricting her access to the web. Their

implementation does not have a robust effect on public discourse and the public sphere. For a

comprehensive historical analysis of DRM, see Bill D. Herman, A Political History of DRM and

Related Copyright Debates, 1987-2012, 14 YALE J.L. & TECH. 162 (2012), and Michael S. Sawyer,

Filters, Fair Use & Feedback: User-Generated Content Principles and the DMCA, 24 B

ERKELEY TECH.

L.J. 363, 380-82 (2009).

27. Extensive scholarship has focused on the role of access providers, hosting facilities,

search engines, social networks, and application providers as gatekeepers. See, e.g., J

ACK

GOLDSMITH & TIM WU, WHO CONTROLS THE INTERNET?: ILLUSIONS OF A BORDERLESS WORLD

(2006); Patricia Sánchez Abril, Private Ordering: A Contractual Approach to Online Interpersonal

Privacy, 45 W

AKE FOREST L. REV. 689 (2010); Annemarie Bridy, Graduated Response and the Turn

to Private Ordering in Online Copyright Enforcement, 89 O

R. L. REV. 81 (2010); Stacey L. Dogan,

Trademark Remedies and Online Intermediaries, 14 L

EWIS & CLARK L. REV. 467 (2010); Mark

MacCarthy, What Payment Intermediaries Are Doing About Online Liability and Why It Matters, 25

BERKELEY TECH. L.J. 1037 (2010); Ronald J. Mann & Seth R. Belzley, The Promise of Internet

Intermediary Liability, 47 W

M. & MARY L. REV. 239 (2005); Joel R. Reidenberg, States and Internet

Enforcement, 1 U.

OTTAWA L. & TECH. J. 213 (2004); Jonathan Zittrain, A History of Online

Gatekeeping, 19 H

ARV. J.L. & TECH. 253 (2006); Paul Sholtz, Transaction Costs and the Social Cost of

Online Privacy, F

IRST MONDAY (May 7, 2001),

http://journals.uic.edu/ojs/index.php/fm/article/view/859/768 [https://perma.cc/4FZU-

9T2V].

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 481

different laws, including the laws of security,

privacy,

defamation,

and

intellectual property.

Algorithmic enforcement by online intermediaries effectively converges law

enforcement and adjudication powers, reflecting a profound transformation in

our traditional system of governance by law. While traditional law enforcement

involves detection, prosecution, adjudication, and meting out punishment,

algorithmic enforcement combines all functions, focusing primarily on detection

and prevention.

As we demonstrate in this Article, the convergence of law

enforcement and adjudication powers in the hands of a small number of mega

platforms, and the robustness of algorithmic enforcement by private, online

intermediaries raise critical challenges to the notions of trust and accountability

that are inherent to reliable systems of law enforcement.

B. Accountability of Online Intermediaries Engaging in Algorithmic Law

Enforcement

Accountability refers to the extent to which decision-makers are expected to

justify their choices to those affected by these choices, be held answerable for their

actions, and be held responsible for their failures and wrongdoings.

Basically,

28. USA PATRIOT Act of 2001, Pub. L. No. 107-56, § 215, 115 Stat. 272, 287-88 (2001)

(facilitating government access to customer data held by service providers).

29. Id. §§ 210-212, 115 Stat. at 283-85 (granting service providers immunity from

damages if they, in good faith, produce data for an investigation undertaken to protect against

international terrorism or clandestine intelligence activities).

30. The Communications Decency Act, 47 U.S.C. § 231 (2012) (criminalizing the online

provision of indecent materials to minors, unless the initiator had undertaken a good faith

effort to determine the age of the person on the other end of the network)

31. Digital Millennium Copyright Act, 17 U.S.C. § 1201 (2012).

32. For instance, proposals to promote wearable cameras among policemen in the US

have followed the incident in Ferguson, Missouri, where a police officer shot an unarmed

teenager named Michael Brown. Presumably, body cameras worn by police officers in the

Ferguson aftermath would reduce police violence and diminish the need to detect and

prosecute police abuse of power. See Michael Brown Shooting: Ferguson Police to Get Body Cameras,

CBCNEWS (Aug. 31, 2014), http://www.cbc.ca/news/world/michael-brown-shooting-

ferguson-police-to-get-body-cameras-1.2752146 [https://perma.cc/K6G5-9JKL]; see also Rich,

supra note 15, at 803.

33. See NAT’L ACADEMY OF SCI., TRUST IN CYBERSPACE (Fred B. Schneider ed., 1999)

(discussing trust in the context of IT systems).

34. Michael W. Dowdle, Public Accountability: Conceptual, Historical, and Epistemic

Mappings, in P

UBLIC ACCOUNTABILITY: DESIGN, DILEMMAS AND EXPERIENCES 1, 3 (Michael W.

Dowdle ed., 2006) (“persons with public responsibilities should be answerable to ‘the people’ for

the performance of their duties.”); Adam M. Samaha, Government Secrets, Constitutional Law, and

Platforms for Judicial Intervention, 53 UCLA

L. REV. 909, 916 (2006) (explaining that democracies

should allow citizens to “appreciably influence the direction of government, and . . . have an

opportunity to assess progress and assign blame”); see also Danielle Keats Citron & Frank

Pasquale, Network Accountability for the Domestic Intelligence Apparatus, 62 H

ASTINGS L.J. 1441

(2011) (focusing on accountability as a measure to cure the problems generated by the growing

use of Fusion Centers); Tal Zarsky, Transparent Predictions, 2013 U.

ILL. L. REV. 1530, 1533

(2013).

482 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

accountability ensures that decision-makers exert power in a fair and effective

manner. It can be produced by ex ante mechanisms that limit the power of

decision-makers through structured guidelines and standards, by ex post

mechanisms of transparency that permit review of the actions of decision-makers

and the outcomes of their decisions, or by both.

Furthermore, accountability

can be achieved through formal mandates, such as legal rules or regulations that

restrain decision-makers’ enforcement power,

and/or through informal means,

such as market forces that check decision-makers’ discretion and promote

voluntary disclosure in relation to their choices and related outcomes.

In the context of algorithmic enforcement by online intermediaries,

generating accountability through these different mechanisms is rather

challenging. Even where formal, structured guidelines exist (i.e., the standards set

by the DMCA

), private players (online intermediaries) translate them into non-

transparent algorithms.

Together with their rapid scalability, these algorithms

are essentially a “black box”

—we do not know, and hence cannot foresee, how

exactly they exercise their power to regulate our online behavior. Nor does

producing accountability through formal or informal mechanisms of transparency

currently seem promising. Indeed, as private, profit-maximizing entities, online

intermediaries who apply enforcement algorithms to manage their users’ behavior

are not required to disclose to the public what content they remove, and for what

specific reason.

They may legitimately craft their own terms of use to manage

the content they distribute. Subjecting private policies of content management

may raise objections similar to those asserted against proposals to interfere with

the editorial discretion of publishers of the daily news or the media.

But when online intermediaries engage in fundamental law enforcement,

35. Orna Rabinovitch-Einy, Technology’s Impact: The Quest for a New Paradigm for

Accountability in Mediation, 11 H

ARV. NEGOT. L. REV. 253, 260 (2006).

36. See Susan P. Sturm, Second Generation Employment Discrimination: A Structural

Approach, 101 C

OLUM. L. REV. 458, 475 (1998) (describing the rule enforcement approach as a

“fixed code of specific rules or commands that establishes clear boundaries governing conduct”).

37. Rabinovitch-Einy, supra note 35, at 261.

38. See infra Part III.A.

39. See infra Part IV.A.

40. See infra Part IV.A.

41. However, under the Open Internet Transparency Rule, ISPs are required to disclose

information about “network management practices, performance, and commercial terms of

service.” The Rule applies to service descriptions, including expected and actual broadband

speed and latency. The Rule also applies to pricing, including monthly prices, usage-based fees,

and any other additional fees that consumers may be charged. Additionally, it covers providers’

network management practices, such as congestion management practices and the types of

traffic subject to those practices. This does not seem to apply to a provider’s copyright policy.

See Open Internet Transparency Rule, F

ED. COMMC’NS COMM’N,

https://www.fcc.gov/guides/open-internet-transparency-rule [https://perma.cc/AZB9-3E67].

42. See Assoc. Press v. United States, 326 U.S. 1, 20 n.18 (1945) (holding that antitrust

law cannot “compel [the Associated Press] or its members to permit publication of anything

which their ‘reason’ tells them should not be published”); see also infra Part II.C.

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 483

they effectively act like judges

who must adhere to the provisions set by the

governing law they enforce. When performing public functions meant to serve

the public at large by a formal or informal delegation of power from the

government—and law enforcement is obviously one of these functions—online

intermediaries effectively act like public administrative agencies.

Just as judicial

enforcement procedures are required to ensure due process and facilitate public

scrutiny to strengthen public trust and promote the rule of law, so too should

algorithms be employed by online intermediaries to enforce the rights of Internet

users.

Otherwise, because online intermediaries resemble a “company town”

more than they do a small, hardly influential newspaper, they could exercise

disproportionate power, which power may ultimately shape public discourse and

even violate users’ fundamental rights. Indeed, if YouTube removes a video, it is

unlikely to be seen; if Google blocks a link to an online business website, it may

literally die. Consequently, interested audiences might be deprived of access to

relevant information provided by individuals or businesses.

We argue that accountability is what distinguishes content management

determinations that online intermediaries make in their private capacities from

content adjudication determinations they make in their administrative capacities

as law enforcers. Adjudication of online content without accountability may lead

to manipulation and abuse of power, create new barriers to open competition and

market innovation, and challenge civil rights.

Through the prism of

43. In relation to the implementation of the Right To Be Forgotten, Google European

Communications Director Peter Barron stated that: “[Google] never expected or wanted to

make [these] complicated decisions that would in the past have been extensively examined in

the courts, [but are] now being made by scores of lawyers and paralegal assistants [at Google].”

See Aoife White, Google EU Ruling Response Vetted as Complaints Pile Up, B

LOOMBERG (Sept. 18,

2014, 6:04 AM), http://www.bloomberg.com/news/articles/2014-09-18/google-eu-ruling-

response-vetted-as-complaints-pile-up [https://perma.cc/UQ6F-UHUE].

44. Edward Lee, Recognizing Rights in Real Time: The Role of Google in the EU Right to Be

Forgotten, 49 U.C.

DAVIS L. REV. 1017, 1055-73 (2016).

45. Id. Lee explains that in relation to implementing the EU’s recent decision about the

Right To Be Forgotten in Case C-131/12, Google Spain SL v. Agencia Española de Protección de

Datos (AEPD) (Costeja), 2014 EUR-Lex 62012CJ0131 (May 13, 2014), which acknowledged the

right of individuals in the EU “to request search engines to remove, from the search results for

an individual’s name, links to web content that contains personal information about the

individual that is ‘inadequate, irrelevant or excessive in relation to the purposes of the

processing,’ ‘not kept up to date,’ or ‘kept for longer than is necessary,’” Google functions

similarly to a government agency or administrative body. Id. at 1022 (citing Costeja, ¶¶ 92, 94).

46. MARGARET JANE RADIN, BOILERPLATE: THE FINE PRINT, VANISHING RIGHTS, AND THE

RULE OF LAW 33 (2013) (using the term “democratic degradation” to describe a situation where

firms displace state regulation); see also M. Todd Henderson, The Nanny Corporation, 76 U.

CHI.

L. REV. 1517, 1535-37 (2009); Tal Zarsky, Social Justice, Social Norms and the Governance of Social

Media, 35 P

ACE L. REV. 154, 166 (2014).

47. As explained by Grimmelman, the free speech interest of a business, which is affected

by its placement in search results, is derivative of users’ free speech, and users’ free speech is

harmed when users are deprived of access to the speech offered by the website. See James

Grimmelmann, Some Skepticism About Search Neutrality, in T

HE NEXT DIGITAL DECADE: ESSAYS

ON THE

FUTURE OF THE INTERNET 435, 441-42 (Berin Szoka & Adam Marcus eds., 2010).

48. See infra Part II.C; see also Lee, supra note 44, at 42-47 (counting several

484 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

accountability, the general public can criticize the manner in which online

intermediaries regulate their content and shape their reactions accordingly.

Particularly, the public may choose not to use the services of platforms that

appear to abuse their power against fundamental rights such as equality, free

speech, and freedom to conduct business, namely, the ability to reach potential

customers in free and competitive markets without undue interference.

C. Accountability Matters in Algorithmic Copyright Enforcement

Algorithmic copyright enforcement offers an excellent case study for

studying the challenges involved in governance by algorithms. Copyright law has

been at the forefront of digital law enforcement since the early 1990s. The ease of

digital copying and mass distribution gave rise to digital locks, digital rights

management (DRM) systems, and technological protection measures (TPM),

which enable rights-holders to technically prevent unauthorized access to and use

of their copyrighted works.

Yet, confronted with the threats of dispersed mass

piracy, rights-holders further increased their pressure on online service providers

(OSPs) to actively participate in fighting online infringement.

Rights-holders

pushed towards active involvement of OSPs in copyright enforcement in

exchange for limited immunity from liability for copyright infringement

committed by their users.

These developments eventually shaped the

intermediary safe harbor regime under the DMCA.

Arguably, copyright enforcement by online intermediaries, pursuant to the

DMCA’s N&TD framework, offers an efficient alternative to the cumbersome,

often impracticable traditional enforcement of copyright through the legal

system, which barely keeps up with the accelerated pace of technological change.

Indeed, the legal system is often understaffed, slow to act, and costly for litigants

and for society,

compared to the cheap, instant, scalable, and robust system of

accountability drawbacks arising from giving Google the primary responsibility of deciding the

contours of the recently recognized Right To Be Forgotten: private anonymous employees that

do not reflect users’ diversity; possible bias of employees in favor of access to information;

minimal due process afforded to affected users; and the inevitable result of mistaken legal

determinations); Zarsky, supra note 46, at 156 (arguing that the notion that a small group of

managers unilaterally sets the rules regulating the social discourse is daunting and may impact

users’ core rights, including their ability to engage in free speech or invoke their right to

privacy).

49. See, e.g., Pamela Samuelson, Intellectual Property and the Digital Economy: Why the Anti

Circumvention Regulations Need to Be Revised, 14 B

ERKELEY TECH. L.J. 519, 534-35 (1999).

50. See infra Part III.A.

51. Niva Elkin-Koren, After Twenty Years: Revisiting the Copyright Liability of Online

Intermediaries, in T

HE EVOLUTION AND EQUILIBRIUM OF COPYRIGHT IN THE DIGITAL AGE 29 (Susy

Frankel & Daniel J Gervais eds., 2014) (“Digital networks have led to an ‘enforcement failure’ in

efforts.”).

52. 17 U.S.C. § 512 (2012).

53. For similar arguments in relation to risk management, see Bamberger, supra note 15,

at 685.

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 485

online copyright enforcement. This explains why much of today’s online

using algorithms not only to remove allegedly infringing content upon notice of

content automatically flagged as infringing.

But should online intermediaries be held accountable for algorithmic

choose to filter allegedly infringing materials or to remove some materials upon

notice, they may simply be making private choices regarding content that is made

available on their platforms.

At the same time however, when online

intermediaries monitor, filter, block, and remove allegedly infringing materials

they engage, in de facto copyright enforcement. The ubiquity of algorithmic

even stronger. In many respects, the N&TD regime under the DMCA, and even

more so voluntary mechanisms of algorithmic copyright enforcement,

effectively privatize governmental functions while blurring the public/private

divide.

Private intermediaries act as both a judge and an executioner,

performing functions of great importance to the public which are normally

reserved to authorized governmental bodies.

Indeed, the original purpose of codifying a safe harbor under the N&TD

procedure was to encourage the cooperation of private, online intermediaries in

combatting online copyright infringement.

Today, most online copyright

enforcement practices take place on privately owned platforms, and not in

traditional legal forums, such as courts and law offices.

Thus, there is a growing

concern that the legitimate business interests of online intermediaries would

compromise the duties involved in performing law enforcement tasks, such as

unprejudiced treatment, transparency, and due process.

Holding online

intermediaries accountable for algorithmic copyright enforcement may therefore

introduce the necessary checks and oversight for the use of semi-governmental

power.

54. See infra Part III.A.

55. James Grimmelmann, Speech Engines, 98 MINN. L. REV. 868, 870-71 (2014).

56. See infra Part III.B.

57. Michael D. Birnhack & Niva Elkin-Koren, The Invisible Handshake: The Reemergence of

the State in the Digital Environment, 8 V

A J. L. & TECH. 6, para. 66 (2003) (explaining that the use

of private sector companies for government censorship may allow government officials, with a

light hand on the trigger, to prevent content from being uploaded to the web, thus violating

freedom of expression without any legal scrutiny); see also Jody Freeman, Private Role in Public

Governance, 75 N.Y.U.

L. REV. 543, 547 (2000).

58. See Lee, supra note 44 (describing the significant role that Google is playing in the

development of the European Union’s Right To Be Forgotten, and positing that Google is

functioning like a private administrative agency).

59. See infra Part III.A.

60. Sharon Bar-Ziv & Niva Elkin-Koren, Uncovering the Invisible: Studying Algorithmic

Online Copyright Enforcement (forthcoming) (on file with authors).

61. Freeman, supra note 57, at 574-75.

486 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

Specifically, three explanations support our proposition that accountability

matters in algorithmic copyright enforcement by online intermediaries: the first

concerns the rule of law and the need to ensure fairness and predictability in

exercising enforcement functions; the second relates to freedom of speech and the

free flow of information, and the third is about preserving sound copyright

policy.

1. Accountability and the Rule of Law

Algorithmic enforcement, like any other law enforcement activity, should

comply with the rule of law. This principle requires that any exercise of power is

duly delegated, and that rules are clear so that people can develop reliable

expectations and make autonomous choices accordingly.

The case of online

intermediaries performing copyright enforcement tasks is no exception: unless

operating pursuant to the authorization of the law (such as the DMCA),

intermediaries may lack copyright enforcement authority. Yet, as explained in this

Subpart, determining whether mechanisms of algorithmic copyright enforcement

employed by online intermediaries abide by the rule of law is far from

straightforward.

Unlike automated enforcement mechanisms that essentially detect strictly

defined unlawful activities, such as red light crossing,

algorithmic copyright

enforcement often involves implementation of flexible legal standards. Many of

the most serious issues in copyright law involve discretion, including determining

the degree of “originality” required to establish copyrightability;

deciding what

amounts to “substantial similarity” to establish infringement;

or considering

what constitutes “permissible use” under fair use.

Resolving these flexible issues

largely requires a qualitative process of assessment and balancing that ought to be

decided on a case-by-case basis.

Translating doctrinal law and policy into code may result in significant, albeit

unintentional, alterations of meaning,

partly because the artificial languages

62. Citron, supra note 11, at 1297.

63. When such detectors issue a ticket, clearly they have detected a car crossing in red

light. This sort of decision-making does not involve any qualitative determinations and

therefore, even when implemented mechanically, the output (a ticket is issued) reveals

sufficient information about the decision-making process (the algorithm identified a car driving

in red light; red light driving is forbidden; therefore, a ticket was issued).

64. See generally Feist Publ’ns, Inc. v. Rural Tel. Serv. Co., 499 U.S. 340 (1991)

(explaining that compilations of pre-exiting facts demand some degree of originality in their

selection and arrangement to satisfy the minimum constitutional standards for copyright

protection).

65. See, e.g., Ideal Toy Corp. v. Fab-Lu Ltd., 266 F. Supp. 755, 756 (S.D.N.Y. 1965), aff’d,

360 F.2d 1021 (2d Cir. 1966).

66. Cambridge Univ. Press v. Becker, 769 F.3d 1232 (11th Cir. 2014).

67. Harper & Row Publishers, Inc. v. Nation Enters., 471 U.S. 539, 561, (1985).

(explaining that fair use determinations demand a thorough case-by-case analysis).

68. See AUSTL. ADMIN. REVIEW COUNCIL, AUTOMATED ASSISTANCE IN ADMINISTRATIVE

DECISION MAKING: ISSUES PAPER 35 (2003),

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 487

intelligible to computers have a more limited vocabulary than human languages.

One risk is that the code being used may not fully capture the nuances of a

particular policy,

which, like fair use, may require case-by-case evaluation. The

code may even alter the original law, either by accident or convenience, with no

prior delegation of public power.

Obviously, “encoded rules that change

established policy cannot be understood by affected individuals or reviewed by

more democratically accountable superiors. In that regard, rulemaking by code

writers is ultra vires even as it is inevitable.”

This is not to say, however, that algorithms cannot process reliable decisions

about copyright infringement and fair use.

While some scholars insist that this

is the case,

others contemplate that machine-learning algorithms may actually

http://www.arc.ag.gov.au/Documents/Automated+Assistance.pdf [https://perma.cc/C43P-

4FPR].

69. James Grimmelmann, Regulation by Software, 114 YALE L.J. 1719, 1728 (2005).

70. See Graham Greenleaf et al., Representing and Using Legal Knowledge in Integrated

Decision Support Systems: DataLex Workstations, 3 A

RTIFICIAL INTELLIGENCE & L. 97, 127 (1995).

71. Citron, supra note 11, at 1297.

72. Id.

73. Relying exclusively on algorithms to execute adequate fair use determinations seems,

at times, unrealistic. For instance, Professors Burk and Cohen, in discussing DRM systems,

stress that “building the range of possible uses and outcomes into computer code would require

both a bewildering degree of complexity and an impossible level of prescience. There is

currently no good algorithm that is capable of producing such an analysis.” Dan L. Burk & Julie

E. Cohen, Fair Use Infrastructure for Rights Management Systems, 15 H

ARV. J.L. & TECH. 41, 56

(2001); see also Mark Lemley, Rationalizing Internet Safe Harbors, 6 J. O

N TELECOMM. & HIGH

TECH. L. 101, 110-11 (2007) (“Image-parsing software may someday be able to identify pictures

or videos that are similar to individual copyrighted works, but they will never be able to

determine whether those pictures are fair uses, or whether they are legitimate copies or displays

made under one of the many statutory exceptions, or whether the individual pictured is 16

rather than 18 years of age.”).

Similarly, some scholars argue that filtering algorithms cannot make reliable qualitative

determinations, such as deciding “the purpose and character of the use” or the “nature of the

copyrighted work.” Furthermore, algorithms are unable to consider information external to the

content itself when analyzing fair use. Yet, in order to assess the fourth factor, “the effect of the

use upon the potential market for or value of the copyrighted work,” the technology must

consider external information about the market. Additionally, because filtering algorithms

make mechanical decisions, some contend that they cannot look at the allegedly infringing

material as a whole. Consequently, they may unlawfully ban transformative uses. 17 U.S.C.

§ 107 (2015); Sawyer, supra note 26 Error! Bookmark not defined., at 389; Edward W.

Felten, A Skeptical View of DRM and Fair Use, C

OMM. A.C.M., Apr. 2003, at 58.

These allegations seem to be outdated in light of today’s algorithms’ learning capacities. In

fact, the more pertinent question should be authority-related: does copyright law permit

delegating the substantial discretion originally accorded to courts to a non-transparent

mechanical process? Determining whether algorithms are both authorized and capable of

making reliable fair use determinations is a normative question that we prefer to leave outside

the scope of this Article. We take the positivistic state of affairs as given: intermediaries

effectively employ different copyright enforcement algorithms and these algorithms are

empowered to execute infringement and fair use determinations.

74. Ira S. Nathenson, Civil Procedures for a World of Shared and User-Generated Content, 48

LOUISVILLE L.R. 912, 938-44 (2010) (describing how content ID procedures may compromise

fair use); Sawyer, supra note 26, at 388-90 (2009) (arguing that fair use considerations cannot be

488 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

enable other algorithms to make smarter decisions based on learned patterns of

data.

The problem is that accomplishing algorithmic accountability becomes

very challenging when algorithms execute discretion-based decisions whose

processing is a “black box.” In other words, even if self-learning algorithms can be

created to engage in case-by-case applications of legal standards, the lack of

transparency remains a serious problem. Most consumers simply do not hold the

necessary expertise to understand the complex technical terminology embedded

in algorithmic enforcement mechanisms.

And as long as online users cannot

comprehend how enforcing algorithms effectively detect online copyright

infringement, they cannot determine whether algorithmic copyright enforcement

effectively complies with what it is authorized to do under the law.

2. Accountability and the Public Sphere

Algorithmic enforcement by prominent online intermediaries who play a

central role in shaping public discourse

carries direct implications for the public

sphere, which further reinforce the need to secure algorithmic accountability. By

offering an alternative medium for information dissemination to which people

increasingly resort—especially when more traditional outlets of distribution are

unavailable—online intermediaries are becoming global arbiters of free speech.

Indeed, it was YouTube Movies, Google Play and Microsoft Zbox Video that

offered online streaming of Sony Pictures’ movie, The Interview, after its release in

theaters was suspended following threats of terrorist attacks.

Online

intermediaries also made the caricatures of the magazine Charlie Hebdo available

encoded in filtering technologies, and therefore when the burden to monitor copyright

infringements shifts from copyright owners to OSP, fair use is likely to be disregarded).

75. NICHOLAS DIAKOPOULOS, ALGORITHMIC ACCOUNTABILITY REPORTING: ON THE

INVESTIGATION OF BLACK BOXES 3 (2013), http://www.nickdiakopoulos.com/wp-

content/uploads/2011/07/Algorithmic-Accountability-Reporting_final.pdf

[https://perma.cc/V9B2-WS9A].

76. Julie E. Cohen, DRM and Privacy, 18 BERKELEY TECH. L.J. 575, 615 (2003).

77. YOCHAI BENKLER, THE WEALTH OF NETWORKS: HOW SOCIAL PRODUCTION

TRANSFORMS MARKETS AND FREEDOM 130 (Yale Univ. Press 2006); see also Niva Elkin-Koren,

User-Generated Platforms, in W

ORKING WITHIN THE BOUNDARIES OF INTELLECTUAL PROPERTY

111, 114-15 (2010).

78. The Interview, a 2014 comedy by Sony Pictures mocking North Korean leader Kim

Jong-Un and depicting a plot for his assassination, was scheduled to be released on Christmas

Day 2014. Following a massive cyberattack, and after receiving threatening messages that any

theater screening the film would be physically attacked, Sony had decided to cancel the release

of the movie in theaters. This controversial decision was widely criticized as self-censorship by

free speech advocates. The decision was publicly denounced by President Obama, artists, and

activists that accused Sony of “caving in” to terrorism and sacrificing free speech. Online

intermediaries came to the rescue of free speech, offering The Interview via online streaming on

YouTube Movies, Google Play and Microsoft Xbox Video. See The Interview: A Guide to the Cyber

Attack on Hollywood, BBC

NEWS (Dec. 29, 2014), http://www.bbc.com/news/entertainment-

arts-30512032 [https://perma.cc/RXX4-PQWU]; ‘The Interview’ Made Available Online After

Cyber Attack, I

RISH TIMES (Dec. 24, 2014, 6:05 PM),

http://www.irishtimes.com/news/world/us/the-interview-made-available-online-after-

cyberattack-1.2048469 [https://perma.cc/Z96D-6PGT].

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 489

after some news outlets decided not to re-publish them.

In other cases, however, online intermediaries have played a different role,

enabling speech control and generating de facto censorship. For instance, actress

Cindy Lee Garcia was successful in raising a doubtful copyright claim against

Google, causing the removal from YouTube of a provocative, anti-Islamic film,

The Innocence of Muslims, based on her insignificant, five-second performance in

the film.

It took Google fifteen months to convince the Court of Appeals that

Garcia’s “weak copyright claim cannot justify censorship in the guise of

authorship,”

and to rescind the order requiring it to take down the controversial

video. In a different case, YouTube facilitated the removal of a documentary film,

India’s Daughter, based on the gang rape of a twenty-three-year-old student, the

screening of which was banned in India due to copyright infringement

allegations.

YouTube also allowed the censorship of the satirical show Fitnah

when it complied with DMCA takedown notices sent by the primary, state-

funded Saudi TV channel, “Rotana.”

From a political perspective, Gannett Co., Inc., a massive media corporation

that owns the Courier-Journal in Kentucky, successfully caused the removal from

YouTube of a forty-second interview with the Democratic candidate for the

79. A few weeks later, in January 2015, a horrific massacre occurred at the Paris offices of

Charlie Hebdo, a magazine which has published satires of the prophet Mohammed. Many

journals have decided not to re-publish the latest Charlie Hebdo caricature for fear they will

also be targeted, and in some countries local stores were reluctant to sell the magazine, fearing

violence (Israel) or in compliance with a government ban (Turkey). CNN, along with other

news outlets, has chosen to censor the controversial cartoons that ran in the magazine. See

Barak Ravid et al., Lieberman Tells Party Activists: Distribute Charlie Hebdo, Israel Must Not Turn

Into ISIS, H

AARETZ (Jan 25, 2015, 9:03 AM), http://www.haaretz.com/news/national/1.638836

[https://perma.cc/RN8D-WZEG]; Charlie Hebdo Attack: Three Days of Terror, BBC

NEWS

(Jan. 14, 2015), http://www.bbc.com/news/world-europe-30708237 [https://perma.cc/2C32-

U6M5]; Constanze Letsch, Charlie Hebdo: Turkish Court Orders Ban on Web Pages Featuring Front

Cover, G

UARDIAN (Jan. 14, 2015, 11:13 AM),

http://www.theguardian.com/world/2015/jan/14/charlie-hebdo-turkey-block-web-pages-

front-cover-muhammad [https://perma.cc/UT99-MTPV]; Alex Stedman, CNN Explains

Decision to Censor Charlie Hebdo Muslim Cartoons, V

ARIETY (Jan. 7, 2015, 1:19 PM),

http://variety.com/2015/tv/news/cnn-addresses-censoring-of-charlie-hebdo-cartoons-

1201395044 [https://perma.cc/K6P5-GM6V].

80. Garcia v. Google, Inc., 766 F.3d 929, 940 (9th Cir. 2014), rev’d en banc, 786 F.3d 733

(9th Cir. 2015).

81. Garcia v. Google Inc., 786 F.3d 733, 743 (9th Cir. 2015) (en banc) (reasoning that

“treating every acting performance as an independent work would not only be a logistical and

financial nightmare, it would turn a cast of thousands into a new mantra: copyright of

thousands”).

82. YouTube removed most copies of the film soon after they became available due to

made the original broadcast from which the uploaded copies were taped. See YouTube Removes

India’s Daughter Videos After BBC Copyright Request, T

RADEMARKS & BRANDS ONLINE (Mar. 9,

2015), http://www.trademarksandbrandsonline.com/news/youtube-removes-india-s-

daughter-videos-after-bbc-copyright-request-4289 [https://perma.cc/X3NN-TAQA].

83. Copyright Law as a Tool for State Censorship of the Internet, BEFORE IT’S NEWS, (Dec. 3,

2014, 11:22 AM), http://beforeitsnews.com/libertarian/2014/12/copyright-law-as-a-tool-for-

state-censorship-of-the-internet-2589350.html [https://perma.cc/WY27-WBBV].

490 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

Senate, Alison Lundergan Grimes, in which she desperately tried to avoid

admitting she voted for President Obama, who was unfavorable in Kentucky.

Notwithstanding the video being a non-infringing fair use that transformed the

original factual work into a political one,

YouTube enabled its censorship, at

least temporarily, less than a month before elections, and exactly when its impact

could have been most powerful.

These examples suggest that algorithmic copyright enforcement may be used

to remove content for reasons that presumably have very little to do with

Unfortunately, although copyright law includes built-in mechanisms that

safeguard freedom of speech, such as the idea/expression dichotomy and the fair

use doctrine,

online intermediaries have several incentives to go beyond these

84. Corynne McSherry, For Shame: Gannett Abuses DMCA to Take Down Political Speech,

LEC. FRONTIER FOUND. (Oct. 10, 2014), https://www.eff.org/deeplinks/2014/10/shame-

gannett-abuses-dmca-take-down-political-speech [https://perma.cc/A2NS-6QK8].

85. Under 17 U.S.C. § 107 (2015),

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 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.

Accordingly, the demonstrative case of Courier-Journal involves a restriction on fair use:

(1) the purpose of using a portion of the interview was completely political, not commercial;

(2) the nature of the interview was highly factual, as it included Grimes’s actual answer to a

question she was asked; (3) the poster took the amount needed to accomplish its political

purpose—only 40 seconds of the interview (a shorter clip might not have managed to achieve

the poster’s purpose); (4) the clip did not provide any financial gain to the poster, which could

have otherwise deprived the copyright owner from income. If at all, the clip may only have had

a positive effect upon the interview’s commercial value, in highlighting its “juicy” parts.

86. For other examples, see Ryan Singel, YouTube Flags Democrats’ Convention Video on

IRED (Sept. 5, 2012, 12:10 AM),

http://www.wired.com/2012/09/youtube-flags-democrats-convention-video-on-copyright-

grounds [https://perma.cc/2HK9-TH58] (highlighting the removal of First Lady Michelle

Obama’s speech from YouTube due to erroneously identifying it as an infringing video).

87. Rossi v. Motion Picture Ass’n of Am., 391 F.3d 1000, 1002 (9th Cir. 2004) (describing

the use of the DMCA’s notice and takedown provisions to induce an ISP to take down a website

from which illegal content could not be downloaded); Online Policy Group v. Diebold, Inc., 337

F. Supp. 2d 1195, 1204-05 (N.D. Cal. 2004) (detailing the use of DMCA notices to induce ISPs

to take down websites containing internal memoranda that embarrassed a voting machine

manufacturer, even though the websites were, in fact, protected fair use). The tactic of using

the DMCA for silencing speech was also discussed in a Forbes article, which explained that the

common corporate wisdom for dealing with Internet critics is to: “ATTACK THE HOST. Find

some copyrighted text that a blogger has lifted from your Web site and threaten to sue his

Internet service provider under the Digital Millennium Copyright Act. That may prompt the

ISP to shut him down.” Daniel Lyons, Fighting Back, F

ORBES (Nov. 14, 2005, 12:00 AM),

http://www.forbes.com/forbes/2005/1114/128sidebar.html [https://perma.cc/MQ9H-45NB].

88. Eldred v. Ashcroft, 537 U.S. 186, 221 (2003) (“To the extent such assertions raise First

Amendment concerns, copyright’s built-in free speech safeguards are generally adequate to

address them.”).

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 491

embedded safeguards and over-enforce copyrights. First, the DMCA has provided

Internet Service Providers, search engines and other intermediaries, with strong

incentives to take down or block access to allegedly infringing content (otherwise,

they may face liability for their users’ infringements).

Second, online

intermediaries may enter into licensing arrangements with prominent entities in

making them vulnerable to business-related

pressures to strictly enforce their partners’ copyrights.

Over-enforcement of copyrights by disregarding fair use, or by removing

works that are in the public domain, does not only limit the right of individuals

who seek to share content online to express themselves and enjoy the fruits of

such expressions liberally,

but also deprives the public as a whole of the benefit

of consuming erroneously restricted speech in the marketplace of ideas.

Information posted by an artist,

an activist,

or a politician

could raise public

awareness of an issue and help communities mobilize around it. But once access

to materials posted online is blocked or removed, a story may not unfold.

Furthermore, the right of speakers to participate in the conversation is

compromised. This triggers concerns regarding the appropriate restraints on

freedom of speech and the compelling need to devise appropriate mechanisms for

algorithmic accountability.

Finally, beyond these critical concerns regarding freedom of expression,

unaccountable enforcement of online content raises additional concerns

regarding other civil rights. From an economic perspective, control over what

89. 17 U.S.C. § 512 (2012) (exempting OSPs from liability for mistaken yet good faith

removals of material); see also Seth F. Kreimer, Censorship by Proxy: The First Amendment, Internet

Intermediaries, and the Problem of the Weakest Link, 155 U. PA. L. REV. 11, 23 (2006) (discussing

the dangers of using proxy censors on free speech); Neil Weinstock Netanel, First Amendment

Constraints on Copyright After Golan v. Holder, 60 UCLA L. REV. 1082, 1120-27 (2013).

90. See supra Part IV.C (discussing YouTube’s Content ID).

91. Kreimer, supra note 89, at 29-30 (“Putting the censorship decision in the hands of the

intermediary allows commercially powerful blocs of customers a potential veto on the speech of

others.”).

92. See Elisa Kreisinger, The Impending Death of the YouTube Mashup, DAILY DOT (June 27,

2014, 9:22 AM CT), http://www.dailydot.com/opinion/youtube-mashup-remix-copyright-

universal [https://perma.cc/5XSA-75SN] (describing the limited ability of artists to upload

mashups to YouTube).

93. See, e.g., Parker Higgins, Houston, We Have a Public Domain Problem, MEDIUM (June 24,

2014), https://medium.com/@xor/houston-we-have-a-public-domain-problem-bd971c57dfdc

[https://perma.cc/6P2T-C6KC] (reporting about an individual who had received a bogus

94. For instance, a video in which Tom Cruise proclaims, in part, that Scientologists are

the only experts on the mind, was removed by YouTube at the request of the Church of

Scientology as part of a long-standing effort to keep copyrighted material from appearing on

the Internet. See Robert Vamosi, Anonymous Hackers Take on the Church of Scientology, CNET

(Jan. 24, 2008, 3:20 PM), http://www.cnet.com/news/anonymous-hackers-take-on-the-

church-of-scientology [https://perma.cc/3EJN-SMSC].

95. See Jacqueline Klimas, Online Campaign Ads May Prove Decisive in Midterm Elections,

ASH. TIMES (Sept. 28, 2014), http://www.washingtontimes.com/news/2014/sep/28/online-

campaign-ads-may-prove-decisive-in-midterm-/?page=all [https://perma.cc/273B-5V68].

492 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

information becomes available may shape the consumer preferences, creating

demand for some content while diminishing demand for other types of content.

Google, for instance, penalizes sites repeatedly accused of copyright infringement,

making them appear lower in Google’s search results.

Users may make an

unjustifiable use of Google’s online mechanism of copyright enforcement to file

bogus copyright complaints against their competitors. This may strongly affect

the business opportunities of targeted sites and their respective owners’

occupation rights, without any warrant or any finding by a court of copyright

infringement. As private, profit-maximizing entities, intermediaries may

potentially abuse their enforcement power due to commercial bias: they may

favor their business partners and other powerful repeat players over weak

Internet users.

Manipulations of this sort impose serious threats on open

competition and market innovation,

further fortifying the importance of

holding online intermediaries accountable for algorithmic copyright enforcement.

3. Accountability and Copyright Policy

Another reason to hold algorithmic copyright enforcement by online

intermediaries accountable relates to its robustness, which may effectively alter

settled copyright policy. Since algorithmic copyright enforcement implemented

by online intermediaries affects a considerable number of people and a large

volume of material, a particular implementation of rules by an algorithm (i.e.,

filtering, removal, blocking) can effectively shape copyright policy. For instance,

algorithmic copyright enforcement may circumvent the objectives of copyright

law by changing the default

: although copyright policy assumes that copyrighted

materials are publicly available unless proven infringing, materials detected by

by the copyright owner.

100

Similarly, the purpose of copyright law is to promote

the creation of new works for the public benefit by providing authors with an

economic incentive to create. Yet if algorithmic copyright enforcement fails to

detect copyright infringement, it may lead to under-enforcement

101

(false

negative), consequently depriving rights-holders of sufficient incentives to create.

Similarly, erroneously removing or blocking access to non-infringing materials

96. An Update to Our Search Algorithms, GOOGLE INSIDE SEARCH (Aug. 10, 2012),

https://search.googleblog.com/2012/08/an-update-to-our-search-algorithms.html

[https://perma.cc/9VHH-874W].

97. See infra note 195 and accompanying text.

98. Jody Freeman, Private Parties, Public Functions and the New Administrative Law, 52

DMIN. L. REV. 813, 845-849 (2000) (acknowledging the potential dangers to democratic

accountability that private actors pose in mixed administration).

99. Jennifer Urban & Laura Quilter, Efficient Process or Chilling Effects: Takedown Notices

Under Section 512 of the Digital Millennium Copyright Act, 22 S

ANTA CLARA COMPUTER & HIGH

TECH. L.J. 621, 636 (2006).

100. See infra Part III.A.

101. See Shay, supra note 15, at 30.

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 493

through over-enforcement

102

(false positive) may further compromise copyright

goals in promoting access to knowledge and encouraging creativity.

Of course, there is no error-free enforcement system,

103

and some level of

error is inevitable in any procedure, including legal procedures.

104

Even

automated systems, which are often perceived to be mistake-resistant, may be

biased.

105

Yet the ubiquity of algorithmic copyright enforcement, derived from its

“codish,” automatic implementation, creates pervasive opportunities for error,

which cannot possibly be corrected promptly.

106

Additionally, because

algorithmic copyright enforcement evolves on private grounds, without the

participation of citizens, public officials, or judges, it is “less likely to encounter

repeal or amendment, compared to laws enforced through traditional means.”

107

Indeed, for an opponent of current copyright laws to contest YouTube’s

preemptive enforcement policy, it is necessary to additionally oppose its

effectiveness, “because, as a result of this effectiveness, the law will be less likely to

change.”

108

Hence, if we wish to preserve the traditional objectives of copyright

law and ensure they are not distorted, we must be able to determine how

D. The Virtues of Accountability—A Three-Factor Framework

The ubiquitous system of embedded copyright governance raises many

concerns. It could effectively change the default rule of copyright law, diminish

freedom of speech, and shape power relations. Online intermediaries are using

algorithms to guard against copyright infringement, but are there adequate

mechanisms ready to guard the guardians? To begin answering this question, we

seek to explore how algorithmic mechanisms of copyright enforcement rank in

accountability measures through the prism of public scrutiny. At this preliminary

stage of algorithmic accountability research, we wish to examine the extent to

102. See Rich, supra note 15, at 812 (“[A]ny technology that seeks to prevent criminal

conduct will inevitably also prevent some non-criminal conduct. This over-breadth may occur

either by design or by mistake.”); Rosenthal, supra note 15, at 594 (“Preemption will likely make

many mistakes in enforcing laws that require subjective, case-specific inquiries to determine

liability,” such as “laws that are designed to be enforced only at the discretion of a private party,

like many copyright restrictions.”).

103. Shay, supra note 15, at 30 n.73.

104. Indeed, in In re Verizon Internet Servs., Inc., the trial court downplayed the significance

of such errors: “[S]uch mistakes are possible using evolving technology, but there is nothing to

suggest they will cause substantial chilling of expression on the Internet.” 257 F. Supp. 2d 244,

264 n. 23 (D.D.C. 2003), rev’d sub nom. Recording Indus. Ass’n of Am., Inc. v. Verizon Internet

Servs., Inc., 351 F.3d 1229 (D.C. Cir. 2003).

105. Batya Friedman & Helen Nissenbaum, Bias in Computer Systems, 14 ACM

TRANSACTIONS ON INFO. SYS. 330, 332 (1996).

106. Id. at 331 (“Computer systems, for instance, are comparatively inexpensive to

disseminate, and thus, once developed, a biased system has the potential for widespread impact.

If the system becomes a standard in the field, the bias becomes pervasive.”).

107. Rosenthal, supra note 15, at 597-98.

108. Id. at 598.

494 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

which the public can monitor algorithmic copyright enforcement by online

intermediaries. Indeed, public review has proved its effectiveness in generating

(SOPA)

109

and the Preventing Real Online Threats to Economic Creativity and

Theft of Intellectual Property Act (PIPA)

110

in January 2012, for instance,

demonstrates how public outcry can influence copyright lawmaking.

111

The

speedy success of the SOPA/PIPA contest

112

shows that public outcry may

sometimes be very powerful in its ability to perpetuate reform.

Due to the dual capacity of online intermediaries who, on the one hand, make

private, business-related decisions regarding content management, and on the

other hand, fulfill governmental duties of law enforcement when engaging in

content adjudication,

113

we leave the aspect of legal scrutiny for future research.

In the remaining discussion, we explore whether members of the public

understand online copyright enforcement policies; whether they enjoy sufficient

opportunities to challenge such policies; and whether they have the capacity to

correct erroneous decisions about online content.

To create a useful accountability toolbox in the context of algorithmic

enforcement by private players, we shift away from traditional administrative law

scholarship, whose inquiry into accountability “focuses inordinately on formal

accountability to the three branches of government,”

114

towards an alternative,

decentralized model of decision making.

115

Perceiving private actors as regulatory

resources capable of promoting the efficacy and legitimacy of administration

leaves room for additional, occasionally informal mechanisms of

accountability.

116

For instance, greater specificity as to the terms enforced by

private players, as well as preservation of minimal administrative procedures,

such as notice and hearing requirements, might play an important role in

oversight.

117

Other possible, non-formal mechanisms of accountability include

voluntary disclosures, which enable the public to access information about

enforcement by private actors and generate market pressure for them to improve

109. Stop Online Piracy Act, H.R. 3261, 112th Cong. (2011).

110. Preventing Real Online Threats to Economic Creativity and Theft of Intellectual

Property Act of 2011, S. 968, 112th Cong. (2011) (also known as the Protect IP Act of 2011).

111. See Yafit Lev-Aretz, Copyright Lawmaking and Public Choice: From Legislative Battles to

Private Ordering, 27 H

ARV. J.L. & TECH. 203, 204 (2013) (“Over one hundred thousand websites

took part in the strike, during which some were effectively closed, while others featured

information about the Bills and directed users to action centers to communicate their worries

to Congress. Users zealously responded and fulminated against the Bills through posts on social

networks, online petitions, and e-mails and phone calls to Congress.”).

112. Id. (“[T]he stated positions by members of Congress on SOPA and PIPA shifted

overnight from 80 for and 31 against to 55 for and 205 against.”).

113. See infra Part III.B.

114. Freeman, supra note 57, at 549.

115. Id. at 548.

116. Id.

117. Id. at 608.

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 495

their decision-making process.

118

Accordingly, our accountability toolbox

identifies three proxies for the public’s practical ability to (1) understand the

algorithmic decision-making process; (2) enjoy sufficient opportunities to

challenge such processes; and (3) correct erroneous/improper decisions about

online content.

The first proxy representing the public’s ability to understand the

implementation of algorithmic law enforcement is transparency.

119

Indeed,

public scrutiny depends on public literacy. Without knowing that specific conduct

took place, it is impossible for the public to render judgment on the merits of such

conduct. In other words, transparency creates public literacy, which is necessary

to establish a demand for fairness and efficiency.

120

Transparent decision-making

processes expose decision makers to the risk of shaming.

121

Fearing that the

public will learn about their missteps, decision makers who function in a

transparent environment are discouraged from engaging in problematic

conduct.

122

Furthermore, transparency also ensures that consumers exercise

meaningful choice regarding which intermediary to use, and put market pressure

on intermediaries to accommodate their interests.

A second proxy in our accountability toolbox, which signifies the ability of

online users to challenge and contest algorithmic decisions about their online

content, is due process. Scholars have increasingly acknowledged the important

role of due process in facilitating algorithmic accountability. Professor Citron, for

instance, coined the term “technological due process”

123

to refer to procedures

designed to ensure that predictive algorithms satisfy some standard of review and

revision to confirm their fairness and accuracy. Other scholars have relied on

Citron’s contribution and expanded its application to scoring systems

124

and

predictive systems of privacy harms.

125

Accordingly, we examine formal measures of due process (i.e., algorithmic

implementations of the counter notice procedure set by the DMCA) and informal

measures (i.e., voluntary dispute procedures) that purport to ensure procedural

118. Id. at 614.

119. Mark Fenster, The Opacity of Transparency, 91 IOWA L. REV. 885, 894 (2006)

(explaining that transparency fosters an informed public debate and generates trust in and

legitimacy for government).

120. Zarsky, supra note 34, at 1533-34.

121. Id. at 1534; Lawrence Lessig, Against Transparency, NEW REPUBLIC (Oct. 8, 2009),

http://www.newrepublic.com/article/70097/against-transparency [https://perma.cc/LVT4-

UHB5].

122. Zarsky, supra note 34, at 1534.

123. Citron, supra note 11, at 1301-13.

124. Danielle Keats Citron & Frank Pasquale, The Scored Society: Due Process for Automated

Predictions, 89 W

ASH. L. REV. 1, 20 (2014).

125. Kate Crawford & Jason Schultz, Big Data and Due Process: Toward a Framework to

Redress Predictive Privacy Harms, 55 B.C.

L. REV. 93 (2014) (relying on a “technological due

process” model to address Big Data’s predictive privacy harms); Neil M. Richards & Jonathan H.

King, Three Paradoxes of Big Data, 66

STAN. L. REV. ONLINE 41, 43 (2013) (calling for a

“[t]echnological [d]ue [p]rocess” solution to governmental and corporate decision-making by

Big Data predictions).

496 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

fairness. The fact that copyright enforcement by online intermediaries happens to

flourish on privately owned grounds should not deem procedural due process

safeguards inapplicable.

126

crucial expressive aspects of individual lives, so we must ensure they “giv[e]

subjects basic rights.”

127

Affected individuals should have the rights to inspect,

correct, and dispute what they believe to be inaccurate adjudication decisions

made with respect to their online conduct. Otherwise, “[i]f law and due process

are absent from this field, we are essentially paving the way to a new feudal order

of unaccountable reputational intermediaries.”

128

Finally, a third proxy of accountability is the extent to which public oversight

may effectively result in correcting errors made by algorithms. For appropriate

public review depends not only on adequately explaining and justifying decision-

makers’ activities to the public, but also on making available accompanying

mechanisms for public sanctions and corrections.

129

Indeed, public pressure can

potentially force the reversal of erroneous content restrictions more quickly than

reversal through the DMCA’s counter notice procedure,

130

and certainly more

quickly than reversal through a lawsuit. Of course, as restricted material becomes

more time-sensitive and newsworthy, the role of public outcry in correcting

improper restrictions of free speech becomes more crucial.

131

In the following Part, we demonstrate that algorithmic copyright

enforcement by online intermediaries ranks poorly in the accountability measures

described above. Specifically, current systems employed by online intermediaries

are non-transparent; they afford insufficient opportunities for affected individuals

to challenge enforcement decisions; and they largely evade public oversight.

126. Martin H. Redish & Lawrence C. Marshall, Adjudicatory Independence and the Values of

Procedural Due Process, 95 Y

ALE L.J. 455, 478–89 (1986) (explaining that the underlying values of

due process include transparency, accuracy, participation, and fairness).

127. Citron & Pasquale, supra note 124, at 19. Securing due process vis-á-vis algorithmic

concerned about the extraordinary power of private entities: e.g., L

ORI ANDREWS, I KNOW WHO

YOU ARE AND I SAW WHAT YOU DID: SOCIAL NETWORKS AND THE DEATH OF PRIVACY 189–91

(2012) (concluding with a proposal for a “Social Network Constitution”); R

EBECCA

MACKINNON, CONSENT OF THE NETWORKED: THE WORLDWIDE STRUGGLE FOR INTERNET

FREEDOM 240–41 (2012) (proposing ten principles of network governance); Jeffrey Rosen,

Madison’s Privacy Blind Spot, N.Y.

TIMES (Jan. 18, 2014),

http://www.nytimes.com/2014/01/19/opinion/sunday/madisons-privacy-blind-

spot.html?_r=0 [https://perma.cc/D2HZ-LCAY] (“What Americans may now need is a

constitutional amendment to prohibit unreasonable searches and seizures of our persons and

electronic effects, whether by the government or by private corporations like Google and

AT&T. . . . [O]ur rights to enjoy liberty, and to obtain happiness and safety at the same time,

are threatened as much by corporate as government surveillance.”).

128. Citron & Pasquale, supra note 124, at 19.

129. Jennifer Shkabatur, Transparency With(out) Accountability: Open Government in the

United States, 31 Y

ALE L. & POL’Y REV. 79, 80-81 (2013).

130. 17 U.S.C. § 512(g)(2)(B) (2012).

131. Sawyer, supra note 26Error! Bookmark not defined., at 392.

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 497

III. E

XPLORING ACCOUNTABILITY IN ALGORITHMIC COPYRIGHT ENFORCEMENT

SYSTEMS

Does algorithmic copyright enforcement by online intermediaries contain

adequate measures to ensure its accountability? In this Part, we apply our three-

factor accountability framework to different systems of algorithmic copyright

enforcement employed by online intermediaries to reveal their accountability

rankings. We proceed in three Subparts. Subpart A presents the legal process of

N&TD set by the DMCA, which establishes the baseline for the development of

algorithmic copyright enforcement systems by online intermediaries. Our analysis

shows that the statutory standards form a deficient starting point in terms of

accountability because they fail to ensure adequate transparency, they only partly

secure due process, and they provide insufficient room for public oversight.

Subpart B uses our three-factor accountability framework to analyze the

accountability of two types of algorithmic copyright enforcement systems:

(1) systems that largely implement the standards set by the DMCA’s N&TD

procedure, performing ex post removals of content; and (2) systems that go

beyond N&TD, enabling blocking, filtering and demoting of content based on

secret, undisclosed proprietary codes.

132

Finally, Subpart C analyzes the accountability of YouTube’s monetizing

system of Content ID as an interesting hybrid of algorithmic copyright

enforcement. Combining an ex ante mechanism of algorithmic content blocking

with an ex post mechanism of content removal, we find the analysis of Content

ID’s accountability to be particularly important, considering YouTube’s dominant

market position and robust impact on free speech and public discourse.

A. The Deficient Accountability Standards of the DMCA

Since the early days of the Internet, online intermediaries were perceived as

potential gatekeepers against the distribution of infringing materials.

133

Given

the ever-growing threats of dispersed mass piracy, copyright owners aimed to

shift some of the burden and costs of monitoring, detecting, and enforcing

copyrights to online intermediaries. But the latter, who had facilitated the

exchange and dissemination of user generated content (UGC), sought to avoid

any cost or burden of online enforcement and to minimize potential barriers to

the free flow of information, which was seen as essential to the development of

their business models.

134

This battle between rights-holders and intermediaries

132. For a comprehensive overview of such voluntary mechanisms, see Annemarie Bridy,

ESEARCH

HANDBOOK ON ELECTRONIC COMMERCE LAW 185 (John A. Rothchild ed., 2016).

133. Niva Elkin-Koren, After Twenty Years: Revisiting the Copyright Liability of Online

Intermediaries, in T

HE EVOLUTION AND EQUILIBRIUM OF COPYRIGHT IN THE DIGITAL AGE 29, 29

(Susy Frankel & Daniel J Gervais eds., 2014) (“[D]igital networks have led to an ‘enforcement

failure’ in copyright-related industries, turning online intermediaries into key players in

enforcement efforts.”).

134. Id. at 29-30.

498 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

shaped the intermediary safe harbor regime under the DMCA.

135

Essentially, this

legislation “help[ed] copyright owners ensure rapid removal of allegedly

infringing material from the Internet while guaranteeing compliant OSPs

136

safe harbor from liability for their users’ acts of copyright infringement.”

137

The N&TD procedure established by the DMCA requires OSPs to respond

“expeditiously” to notices of infringement by removing or disabling access to

allegedly infringing material when certain conditions are met.

138

Hosting services

(websites, social networks) are further required to take “reasonable steps

promptly to notify the subscriber that it has removed or disabled access to the

material”

139

and promptly forward any counter notices from alleged infringers

back to the original complainant.

140

If, after ten to fourteen days following receipt

of a counter notice, the complainant does not notify the OSP that she has filed a

lawsuit, then the OSP must reinstate the contested material.

141

To maintain their

immunity under the N&TD regime, OSPs cannot have actual knowledge that

infringing content is on their systems or be “aware of facts or circumstances from

which infringing activity is apparent.”

142

Moreover, they should not receive a

direct financial benefit from any infringing activity which they have the right and

ability to control.

143

Finally, the DMCA further encourages compliance with

N&TD by exempting OSPs from liability for mistaken yet good faith removals of

material.

144

The DMCA sets minimum standards that afford copyright owners a short,

135. 17 U.S.C. § 512 (2012); Jennifer M. Urban & Laura Quilter, Efficient Process or

“Chilling Effects”? Takedown Notices Under Section 512 of the Digital Millennium Copyright Act, 22

SANTA CLARA COMPUTER & HIGH TEC. L. J. 621, 621 (2005).

136. To maintain immunity from monetary liability for material that is transmitted over

networks, cached on a server, linked to, or stored at the direction of a user, OSPs were required

to adopt and implement certain policies. In particular, OSPs must comply with two preliminary

policies. First, they must adopt and reasonably implement a policy to terminate the accounts of

repeat infringers and must notify users of this plan. Second, they must also accommodate

“standard technical measures” used by copyright owners to identify infringing material. See 17

U.S.C. § 512(a), (b), (c), (d), (i).

137. Urban & Quilter, supra note 99, at 622.

138. 17 U.S.C. § 512(b)(2)(E)(i)-(ii), 512(c)(1)(C).

139. Id. § 512(g)(2)(A).

140. Id. § 512(g)(2)(B). A counter-notification must include the following: (A) a physical

or electronic signature; (B) identification of the material removed and its former location;

mistakenly removed; (D) the user’s name, address, and phone number; and (E) consent to the

jurisdiction of Federal District Court. See id. § 512(g)(3).

141. Search engines, on the other hand, are not required to notify the alleged infringer of

removal because they are not expected to have any service relationship with the alleged

infringer. 17 U.S.C. § 512(d); see also Urban & Quilter, supra note 99, at 626.

142. 17 U.S.C. § 512(c)(1)(A). If OSPs later become aware of such content, they must

expeditiously remove it from their systems.

143. Id. § 512(c)(1)(B).

144. Id. § 512(g)(1) (Intermediaries that fail to act in good faith may lose safe harbor and

may be required to pay damages to content providers whose material was unlawfully removed

under the intermediaries’ stated terms of use).

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 499

clear and efficient procedure for removing infringing materials from OSPs’

platforms. Although the drafters of N&TD arguably did not foresee the

robustness of online copyright enforcement, this procedure articulates a de facto

baseline for automatic, algorithmic online copyright enforcement. Unfortunately,

however, the minimum standards set by the DMCA constitute a deficient starting

point in facilitating accountability, irrespective of their algorithmic application,

for three reasons: (1) they are not entirely transparent; (2) they only partly secure

due process; and (3) they are insufficiently exposed to public oversight.

First, the DMCA’s N&TD regime generates transparency primarily by

establishing a duty to notify targeted users about removals of their content.

145

hosting service (e.g., a website, a social network) is thus required to take

“reasonable steps promptly to notify the subscriber that it has removed or disabled

access to the material.”

146

This legal requirement, however, is insufficient to

foster transparency since it fails to apply to major functions of OSPs—linking and

caching—notwithstanding their significant implications for public discourse.

147

When users search the web for content, the search results they retrieve often

shape their expectations and subsequently, their knowledge of the subject of

inquiry. Thus, removal of links from search indexes makes it harder for users to

find and access the allegedly infringing material.

148

Because the DMCA

encourages online intermediaries who receive prompt takedown notices to

remove allegedly infringing content automatically,

149

online copyright

enforcement mechanisms are likely to remove content when removal is

inappropriate and even in cases where the claim is based on erroneous

misidentification.

150

If intermediaries do not promptly notify content providers

about the removal of their links, it may be too late for the providers to find out

about it independently. Content providers are unlikely to check routinely that all

of the content they have posted online appears in major platforms’ search results.

Consequently, by the time providers independently discover that their links no

145. 17 U.S.C. § 512(g)(2)(A).

146. Id.

147. Id. § 512(d); Urban & Quilter, supra note 99, at 681 (the counter-notification

provision under section 512(g)(A) of the DMCA requires notifying a “subscriber of the service

provider” whose “material is residing at the direction of the subscriber on a system or network

controlled or operated by or for the service provider, or to which access is disabled by the

service provider.” It seems inapt to relate to the owner of a site whose content has been linked

or indexed by the service provider as the provider’s “subscriber” for the simple reason that

search and caching services do not have account holders or subscribers); see also Perfect 10,

Inc. v. Google, Inc., No. CV 04–9484 AHM (SHx), 2010 WL 9479059, at * 4 (C.D. Cal. Jul. 26,

2010). Furthermore, the information location links provided by § 512(d) service providers do

not fit appropriately into the definition of “material residing at the direction” of the subscriber

because they are generated by the service provider. Seng, supra note 7, at 428-29.

148. Urban & Quilter, supra note 99, at 682.

149. 17 U.S.C. § 512(c)(1)(C) (2013).

150. This is not to say the humans cannot make mistakes too. The argument is that in

black-or-white cases, where takedown notices are unequivocally erroneous (because they

identify the wrong copyrighted work, for instance), humans are likely to get it right and decline

the claim.

500 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

longer appear in the search results of major platforms, the harm to their

businesses may be irreparable. Furthermore, content providers cannot count on

public outcry to contest the removal of the links to their content from the search

results. Indeed, unless users are looking for specific content, they are unlikely to

realize that such materials are no longer available.

151

Users of search engines may

sometimes lack precise knowledge of which content they are looking for,

otherwise they might have accessed it directly. Hence, by the time content

providers could appreciate that links to their content are no longer available

through major search engines, they could illegitimately lose a lot of traffic.

Furthermore, the DMCA’s provisions not only deprive link providers of

promptly learning that their links have been removed, but they also deprive them

of learning the cause for the removal. Recognizing that a specific link has been

removed and understanding that the cause for removal was copyright-related are

both essential preconditions for targeted users, as well as third parties, to contest

the removal of their links. Links may also be removed due to non-copyright

related reasons, such as child pornography, defamation or simply because a bug

caused them to be inaccessible. Hence, at least with respect to removals that are

based on § 512(d), a genuine transparency problem is embedded in the DMCA’s

framework, making it difficult for targeted information providers to acquire

knowledge about the removal of their links.

The second reason why the DMCA’s framework constitutes a deficient

starting point in facilitating accountability relates to its mistreatment of due

process. To begin with, the DMCA’s counter notice procedure

152

provides

impractical challenging opportunities for alleged infringers operating in a robust

sphere of online copyright enforcement. Consider for example the domain

4shared.com, which received an average of 1166 removal requests a week from

August 2016 to September 2016, amounting to less than 5% of the domain’s total

indexed URLs.

153

It seems like a heavy burden for the domain owner to employ

151. Niva Elkin-Koren, Let the Crawlers Crawl: On Virtual Gatekeepers and the Right to

Exclude Indexing, 26 U.

DAYTON L. REV. 179, 180-81 (2001).

152. 17 U.S.C. § 512(g)(2)(A). A counter-notification must include the following: (A) a

physical or electronic signature; (B) identification of the material removed and its former

location; (C) statement under penalty of perjury that the user has a good faith belief the material

was mistakenly removed; (D) the user’s name, address, and phone number; and (E) consent to

the jurisdiction of Federal District Court. Id. § 512(g)(3). If after ten to fourteen days, the

complainant does not notify the webhost that it has filed a lawsuit, then the webhost must

reinstate the contested material. Otherwise the webhost risks losing its safe harbor and it may

be found liable for the damages suffered by users whose content had been unlawfully restricted.

153. See Google, Specified Domain: 4shared.com, GOOGLE TRANSPARENCY REPORT (Sept.

2016)

https://www.google.com/transparencyreport/removals/copyright/domains/4shared.com

[https://perma.cc/8RUS-NCLC] (this domain seems to be legitimate as more than 95% of its

indexed URLs are not targeted as infringing by reporting organizations. Nevertheless, it still

suffers from inadequate challenging opportunities under which the domain owner must dispute

each takedown request independently); see also Colette Bennett, Nintendo Issues DMCA

Takedown Notice for Hundreds of Fan-Made Games, D

AILY DOT (Sept 6, 2016, 8:06 AM),

http://www.dailydot.com/parsec/nintendo-pulls-fan-games [https://perma.cc/SUZ9-YBKC]

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 501

the counter notice procedure for each of these requests. Accordingly, granting a

statutory counter notice right simply does not seem to fit today’s vigorous world

of massive, predominantly automatic,

154

online copyright enforcement.

But even for a more moderate reality of online copyright enforcement, the

statutory counter notice procedure does not seem to offer adequate challenging

opportunities.

155

Removing material that may qualify as fair use before notifying

the alleged infringer and before giving her the opportunity to contest the removal

in a hearing may result in “an extra-judicial temporary restraining order, based

solely on the copyright holder’s allegation of copyright infringement.”

156

Requiring OSPs to reinstate materials that were subject to a counter notice in

cases where complainants fail to file suit within fourteen days does not seem to

cure this problem, at least not when removals concern time-sensitive expressive

material.

157

Furthermore, the DMCA’s framework does not require OSPs to

include any explanation of the legal ramifications of filing a counter notice (e.g.,

the OSP’s duty to reinstate the removed material, unless the complainant files a

procedure require OSPs to disclose the identity of the copyright owner and the

details of the allegedly infringed copyrighted work. This obviously frustrates the

ability of alleged infringers to contest content removals, which depends largely on

the comprehensiveness of the takedown notice they receive.

158

Moreover, the DMCA’s safe harbor provisions effectively encourage this

denial of users’ “due process”

159

because they exempt OSPs from liability for

mistaken yet good faith removal of material,

160

without affording a parallel

protection for failure to act on a notice in good faith.

161

While copyright owners

should only request that intermediaries remove content when they have a good

(describing how Game Jolt, an open source indie gaming community received hundreds of

DMCA takedown notice for allegedly infringing uses of Nintendo products).

154. Today, instead of human review of content that may be infringing, there are

companies, such as Total Wipes, that employ robots to automatically send takedown requests

to some of the world’s most famous online services, including Skype, Tor, Dropbox,

LibreOffice, Python, and WhatsApp. See, e.g., Jamie Williams, Absurd Automated Notices Illustrate

Abuse of DMCA Takedown Process, E

LEC. FRONTIER FOUND. (Feb. 24, 2015),

https://www.eff.org/deeplinks/2015/02/absurd-automated-notices-illustrate-abuse-dmca-

takedown-process [https://perma.cc/NEK2-QARG].

155. Urban & Quilter, supra note 99, at 639.

156. Id.

157. Id. at 637.

158. For instance, Canada employs a regime of “Notice and Notice” under which ISPs who

receive a notice from a copyright holder that their subscribers may be infringing copyright

must forward the notice to the subscriber. In the 2012 amendments to the Canadian Copyright

Act, this regime has been codified and became mandatory as of January 2nd 2015. See Notice and

Notice Regime, O

FF. OF CONSUMER AFF. (Jan. 20, 2015), http://www.ic.gc.ca/eic/site/oca-

bc.nsf/eng/ca02920.html [https://perma.cc/ZWD5-F3GG].

159. JAY DRATLER, JR., CYBERLAW: INTELLECTUAL PROPERTY IN THE DIGITAL MILLENIUM

§ 6.05(1)(c) (2007); see also discussion supra Part II.C.3 regarding accountability and copyright

policy.

160. 17 U.S.C. § 512(g)(1) (2012).

161. Urban & Quilter, supra note 99, at 638.

502 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

faith belief that a user’s material is infringing their copyright

162

—a requirement

that was recently interpreted by the Ninth Circuit in Lenz v. Universal Music

Corp.

163

as mandating copyright holders to consider the fair use defense before

sending a notice of removal—it would be naïve to expect copyright owners to

make objective, court-like determinations regarding the fair use of their own,

private creations. Moreover, the fear of OSPs that in reviewing content and

questioning frivolous claims they will risk losing the safe harbor because they

“knowingly” host infringing material,

164

makes it more likely for OSPs to err on

the side of unnecessary removals, further diminishing due process. Finally, to

avail themselves of the benefits of judicial review, alleged infringers must first

submit a counter notice, upon which the complainant must then file suit.

165

Unless the complainant files suit, courts do not review DMCA content

takedowns, and even when they do, it may often be too late, especially when the

removal concerns expressive material.

166

The third reason why the DMCA is a deficient starting point in promoting

accountability in online copyright enforcement is its failure to facilitate sufficient

public oversight and proper opportunities for correcting erroneous removals.

167

The DMCA’s framework theoretically enables the correction of erroneous

content removals by framing an ex post regime of N&TD that allows the removal

of content only after it first appears online. This is insufficient to foster public

participation. Upon receiving a notice of infringement from copyright owners,

intermediaries must remove the allegedly infringing material “expeditiously.”

168

Considering the immense volume of takedown requests facilitated by the

DMCA’s speedy enforcement regime,

169

it would be impossible for the public as a

whole to promptly identify and contest inappropriate content removals, especially

when the content removed does not seem to have any groundbreaking

potential.

170

B. Regulated Versus Voluntary Mechanisms of Algorithmic Copyright

Enforcement

The previous Subpart showed that the procedural standards of N&TD set by

the DMCA fail to offer a sufficient framework for accountability in online

162. 17 U.S.C. § 512(g)(1).

163. Lenz v. Universal Music Corp., 801 F.3d 1126, 1135-36 (9th Cir. 2015).

164. See supra note 142 and accompanying text.

165. Urban & Quilter, supra note 99, at 628.

166. Id. at 637.

167. See supra discussion about the virtues of accountability in Part II.D.

168. 17 U.S.C. § 512(c)(1)(C) (2013).

169. See supra note 7.

170. Some private initiatives have attempted to make it easier for target users and third

parties to keep track with DMCA content removals, but they too seem insufficient to enhance

public review over such a ubiquitous system. See infra Part IV.B.

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 503

algorithmic implementation of online copyright enforcement makes the lack of

accountability even more severe. Particularly, the lack of accountability arises in

three aspects: (1) the non-transparent implementation of algorithmic copyright

enforcement imposes serious limitations over public literacy; (2) its automatic

application restrains the ability of users to challenge it, further diminishing due

process; and (3) its robustness further reduces the likelihood of public-driven

corrections.

Regulated mechanisms of algorithmic copyright enforcement refer to

algorithms that largely implement the DMCA’s N&TD regime, and therefore seek

to comply with its settled accountability standards. Dominant intermediaries,

such as Google,

171

Facebook,

172

and Twitter,

173

apply the DMCA’s framework of

N&TD using algorithms. Facebook, for instance, allows rights-holders (and their

legal representatives) to file a notice of copyright infringement using an online

form, containing all of the detailed information required by 17 U.S.C.

§ 512(c)(3)(A).

174

When it does remove content, Facebook sends a warning to the

user who posted the content, notifying her that content posted to Facebook was

removed due to a notice of copyright infringement.

175

Facebook also provides the

alleged infringer with the claimant’s contact information and allows her to submit

a counter-notification, if the content was removed under the notice and counter

notice procedures of the DMCA.

176

Google,

177

Twitter

178

and YouTube

179

apply

171. See Removing Content from Google, GOOGLE SUPPORT (2016),

https://support.google.com/legal/troubleshooter/1114905?rd=2 [https://perma.cc/T5M9-

NTV4].

172. See Reporting Copyright Infringements, FACEBOOK HELP CTR. (2016),

https://www.facebook.com/help/400287850027717 [https://perma.cc/Y6RG-PVC5].

173. See Copyright Policy, TWITTER HELP CTR. (2016)

https://support.twitter.com/articles/15795-copyright-and-dmca-policy

[https://perma.cc/8N82-2VME].

174. Facebook, for instance, requires the following information to be included in a notice

of copyright enforcement:

The fastest and easiest way to submit a claim of copyright infringement to us is to use our

online form. It may be required by law that you include the following information: Your

complete contact information (full name, mailing address and phone number). Note that we

regularly provide your contact information, including your name and email address, the name

of your organization or client who owns the right in question, and/or the contents of your

report to the person who posted the content you are reporting. You may wish to provide a

professional or business email address for contact by users. A description of the copyrighted

work that you claim has been infringed. A description of the content on our site that you claim

infringes your copyright. Information reasonably sufficient to permit us to locate the material

on our site. The easiest way to do this is by providing web addresses (URLs) leading directly to

the allegedly infringing content. A declaration that you have a good faith belief that use of the

copyrighted content described above, in the manner you have complained of, is not authorized

by the copyright owner, its agent, or the law, the information in your notice is accurate, and you

declare, under penalty of perjury, that you are the owner or authorized to act on behalf of the

owner of an exclusive copyright that is allegedly infringed. Your electronic signature or physical

signature.

See What Should I Include when Submitting a Report to Facebook Alleging Infringement of My

ACEBOOK HELP CTR. (2016), https://www.facebook.com/help/400287850027717

[https://perma.cc/74TF-72DK].

175. Id.

176. This suggests that Facebook also removes content by algorithms that do not

504 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

similar algorithmic implementations of the DMCA’s N&TD procedure.

Nevertheless, these prominent platforms occasionally go beyond the

framework of the DMCA, exploiting algorithms to filter, block and demote

content, sometimes before it ever becomes available online. While these

voluntary mechanisms shift the heavy burden of policing online copyright

infringement from copyright owners and courts to algorithms, they create greater

challenges to accountability because they are not subject to any regulation—not

even that of the DMCA.

Among these voluntary algorithms are filters that allow the ex ante blocking

of content. In late 2007, some UGC

180

sites, including MySpace, Veoh,

DailyMotion, and Soapbox (via Microsoft), collaborated with large content

companies, including Disney, CBS, NBC Universal, and Viacom, in

recommending that UGC sites use copyright filtering technologies.

181

These

technologies essentially compare uploaded material against samples of

copyrighted material (Reference Material) provided by copyright owners. When

users attempt to upload content that matches any Reference Material, the content

can be blocked before it ever becomes available online.

182

Additionally, search engines implement algorithms, such as Google’s Pirate

algorithm, to combat online copyright infringement.

183

Updated in 2014,

184

this

algorithm essentially changes the search algorithm so that it downgrades the

implement the DMCA’s framework of ex post content removal.

177. See Submit a Copyright Takedown Notice, YOUTUBE HELP (2016),

https://support.google.com/youtube/answer/2807622 [https://perma.cc/E2ND-5BTM].

178. See How Do I File a Copyright Claim, TWITTER HELP CTR. (2016),

https://support.twitter.com/entries/15795#5 [https://perma.cc/P8AT-PDZB].

179. See supra note 177.

180. For analysis of the legal issues related to user-generated content, see Niva Elkin

Koren, Governing Access to User-Generated-Content: The Changing Nature of Private Ordering in

Digital Networks, in GOVERNANCE, REGULATIONS AND POWERS ON THE INTERNET 318-43 (2012);

Tom W. Bell, The Specter of Copyism v. Blockheaded Authors: How User-Generated Content Affects

AND. J. ENT. & TECH. L. 841 (2008) (providing an economic view of UGC

and predicting that UGC will drive down the costs and increase accessibility for all content);

Greg Lastowka, User-Generated Content and Virtual Worlds, 10 V

AN. J. ENT. & TECH. L. 893

(2008); Edward Lee, Warming Up to User-Generated Content, 5 U.

ILL. L. REV. 1459 (2008).

181. PRINCIPLES FOR USER GENERATED CONTENT SERVICES, http://ugcprinciples.com

[https://perma.cc/Q6Y3-BUFS].

182. Sawyer, supra note 26, at 365. A classic example for a filtering technology is the

algorithm applied by YouTube’s system of Content ID, which automatically scans videos as they

are uploaded for copyrighted material and blocks access to videos that contain material which

183. Glenn Gabe, Google’s Pirate Algorithm and DMCA Takedowns: Exploring the Impact

Threshold, G-S

QUARED INTERACTIVE (Dec. 9, 2013), http://www.hmtweb.com/marketing-

blog/google-pirate-algorithm-dmca [https://perma.cc/EED9-RPWG]; Stuart Long, Google

Rolls Out DMCA Algorithm Update After More than 2 Years, BRANDED3 (Oct. 27, 2014),

http://www.branded3.com/blogs/google-rolls-dmca-algorithm-update-2-years

[https://perma.cc/3ZR5-ZFAW].

184. Continued Progress on Fighting Piracy, GOOGLE PUB. POL’Y BLOG (Oct. 17, 2014),

http://googlepublicpolicy.blogspot.co.il/2014/10/continued-progress-on-fighting-piracy.html

[https://perma.cc/RLX9-QDUG].

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 505

ranking of allegedly pirated websites, placing them at the bottom of the search

results list. Unfortunately, this method leaves the door open for manipulation by

strategic players. For example, competing websites could abuse this mechanism as

a tool for sabotage. By filing sham notices of copyright infringement against their

competitors, they can increase the likelihood that a content-demoting algorithm

will kick in and make their competitors’ websites appear lower in the search

engine’s search results list. Presumably, politicians could also leverage this

mechanism as a reputation management strategy.

185

We now turn to show that algorithmic implementations of the statutory

N&TD procedure do not adequately facilitate the three aspects of accountability,

and that voluntary measures of algorithmic copyright enforcement, which go

beyond N&TD, fare even worse. While advocates of voluntary measures of

algorithmic copyright enforcement maintain that they solve many of the

problems associated with online infringements,

186

we demonstrate that they leave

almost no room for public scrutiny, notwithstanding their robust implications for

public discourse.

187

1. Transparency

Algorithmic copyright enforcement by online intermediaries imposes serious

limitations on transparency because both regulated and voluntary measures fail to

fully disclose how they exercise their power. It is unclear what precise

circumstances trigger the underlying enforcing algorithms. One could expect that

algorithmic N&TD methods would be slightly more predictable than voluntary

methods, as they largely adhere to the DMCA framework. However, these

methods still retain some unrevealed discretion to determine when copyright

infringement has occurred and to decide when to apply the statutory N&TD

takedown procedure automatically.

For instance, although the statutory language uses the imperative language

“acts expeditiously to remove, or disable access”

188

when defining the

responsibilities of an intermediary seeking to enjoy the DMCA safe harbor in

relation to allegedly infringing content on its platform, Facebook admits that after

receiving a claim of copyright infringement, it may remove the reported content

from its platform.

189

When exactly this would happen is largely a mystery.

185. For example, the DMCA has been used to target political content on YouTube. Elliot

Harmon, Once Again, DMCA Abused to Target Political Ads, E

LEC. FRONTIER FOUND. (Nov. 17,

2015), https://www.eff.org/deeplinks/2015/11/once-again-dmca-abused-target-political-ads

[https://perma.cc/FQ5G-U8JL] (reporting on abuse of the DMCA to takedown a television ad

supporting a controversial proposal to regulate short-term property rental services like

Airbnb).

186. Amir Hassanabadi, Viacom v. Youtube: All Eyes Blind: The Limits of the DMCA in a Web

2.0 World, 26 B

ERKELEY TECH. L.J. 405, 438 (2011).

187. See supra Part II.C.2.

188. 17 U.S.C. § 512(b)(2)(E)(i)-(ii), 512(c)(1)(C) (2012).

189. See What Happens After I Submit a Claim of Copyright Infringement to Facebook?,

ACEBOOK HELP CTR. (2016), https://www.facebook.com/help/400287850027717

506 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

Algorithmic implementations of N&TD are not as straightforward as we would

have expected automated systems to be. Occasionally, they disregard takedown

notices and elect not to remove allegedly infringing content.

190

The use of voluntary measures makes the situation described above even less

transparent. Algorithms are employed to filter, block, and demote content based

on an entirely undisclosed, self-determined threshold. Without understanding the

nuances of their decision-making processes, it is impossible to hold such

algorithms accountable.

2. Due Process

Holding algorithmic mechanisms of copyright enforcement accountable also

depends on the availability of adequate channels to contest content restrictions.

Nonetheless, neither regulated mechanisms of algorithmic law enforcement nor

voluntary ones facilitate adequate challenging opportunities, consequently

diminishing due process. In the case of regulated mechanisms of N&TD, this is

because they often fail to comply with the already-deficient standard of counter

notice set by the DMCA.

191

For instance, under YouTube’s algorithmic

implementation of the DMCA N&TD, the counter notice feature only becomes

available if the alleged infringer elects to dispute the infringement claim.

192

But

an alleged infringer who receives a Content ID claim of copyright

infringement

193

may elect not to dispute the claim, and instead either

acknowledge it, remove the allegedly infringing material, swap out the allegedly

infringing audio track with one of YouTube’s free-to-use songs, or share revenue

with the copyright owner.

194

By choosing one of these alternative options, rather

than disputing the claim, the content provider effectively forgoes his ability to file

a counter notice. But under the DMCA, counter notices should be available in

theory and in practice, regardless of the intermediary’s internal business model.

Furthermore, anecdotal evidence proves that sometimes, counter notices fail

to fulfill their statutory role in ensuring intermediaries repost content they had

[https://perma.cc/RW7G-EPQ8].

190. For instance, Twitter was accused of disregarding DMCA takedown notices filed by

an artist named Christopher Boffoli, who created the popular “Disparity Series,” consisting of

photographs featuring miniature figures in funny poses on various types of food. Since his

photographs went viral, Boffoli has been sending takedown requests under the DMCA to

individuals and sites like Facebook, Pinterest, and Google. In his complaint to a U.S. district

court, Boffol accused Twitter of inducing copyright infringement and failing to disable access to

copyrighted material even after being notified about infringing uses. See Jon Brodkin, Twitter

Won’t Take Down “Giant Food” Photos, so Artist Sues, A

RS TECHNICA (Sept. 11, 2012, 2:52 PM),

http://arstechnica.com/tech-policy/2012/09/11/twitter-wont-take-down-tiny-food-photos-

so-artist-sues [https://perma.cc/7RAC-E5DN].

191. See supra Part III.A.

192. As we explain henceforth in Part III.C, YouTube’s Content ID procedure is a hybrid

of regulated and voluntary copyright enforcement.

193. See infra note 213 and accompanying text.

194. What Is a Content ID Claim?, YOUTUBE HELP (2016),

https://support.google.com/youtube/answer/6013276 [https://perma.cc/H6FW-MUMW].

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 507

previously removed due to a DMCA notice of infringement. Consider the

following message that an artist named John McKelvey received from YouTube

as a response to a counter notice he had submitted after his video, which included

short portions of songs taken from Eric B. & Rakim’s record for critical purposes,

was taken down due to a copyright infringement notification filed by Universal:

“Thank you for your counter-notification. The complainant has reaffirmed the

information in its DMCA notification. YouTube has a contractual obligation to

this specific copyright owner that prevents us from reinstating videos in such

circumstances. Therefore, we regretfully cannot honor this counter-

notification.”

195

YouTube states outright that, it will not repost videos that allegedly infringe

content owned by “partner” companies following their removal, even if their

uploaders file a counter notice showing they constitute fair use, due to YouTube’s

contractual obligations to the copyright owners involved.

196

Considering that

their counter notices may be disregarded due to YouTube’s internal business

relations, it should not surprise anyone why potential target users rarely bother to

file counter notices.

Additionally, regulated mechanisms of copyright enforcement often impede

users’ ability to file counter notices.

197

Consider the following example of an

actual notice of removal sent by Facebook’s algorithm to a user:

We have removed your video entitled (no title) uploaded at 9:00am February

3rd, 2014. This video may include copyrighted material (such as a clip or audio)

that you do not have the right to share.

If you think your video should not have been removed because:

(1) you are the copyright owner, or

(2) you have permission from the copyright owner to upload and distribute this

material on Facebook, or

195. Andy, YouTube’s Deal with Universal Blocks DMCA Counter Notices, TORRENTFREAK

(Apr. 5, 2013), http://torrentfreak.com/youtube-deal-with-universal-blocks-dmca-counter-

notices-130405 [https://perma.cc/6UZB-YJ3A]; see also Mike Masnick, YouTube Won’t Put Your

Video Back up, Even if It’s Fair Use, if It Contains Content from Universal Music, T

ECHDIRT (Apr. 5,

2013, 11:52 AM), https://www.techdirt.com/articles/20130405/01191322589/youtube-wont-

put-your-video-back-up-even-if-its-fair-use-if-it-contains-content-universal-music.shtml

[https://perma.cc/3HKH-Y7LT].

196. Andy, supra note 195. YouTube claims the following:

YouTube enters into agreements with certain music copyright owners to allow use of their

sound recordings and musical compositions. In exchange for this, some of these music copyright

owners require us to handle videos containing their sound recordings and/or musical works in

ways that differ from the usual processes on YouTube. Under these contracts, we may be

required to remove specific videos from the site, block specific videos in certain territories, or

prevent specific videos from being reinstated after a counter notification. In some instances, this

may mean the Content ID appeals and/or counter notification processes will not be available.

Your account will not be penalized at this time.

Id.

197. Under YouTube’s application of N&TD, the process of counter notice only becomes

available if the alleged infringer elects to dispute the Content ID claim, and the copyright owner

can respond to a dispute in several different ways: (1) release the claim; (2) uphold the claim;

(3) take down the video by submitting a formal notice. See infra note 216 and accompanying

text.

508 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

(3) you otherwise believe you are legally entitled to upload and distribute this

material on Facebook you may visit the link below to video an appeal requesting

that it be reinstated:

[link redacted]

If you do not want to appeal, there is no need to take any action. Please be careful

about videos you upload in the future. If they are identified as possibly

containing copyright infringing material, they may also be removed. This could

result in us temporarily or permanently blocking your ability to upload videos,

or permanently disabling your account.

198

This notice does not foster due process. To begin with, an alleged infringer

receiving such a notice does not necessarily understand its merit.

199

Even if she

does, it would be hard for her to dispute this notice because it omits crucial

information that an alleged infringer must have in order to properly dispute

content removal; most importantly, the details of the allegedly infringed work and

the identity of the copyright owner requesting the takedown.

With voluntary algorithmic copyright enforcement, due process is

completely non-existent. Since algorithms replace copyright owners in policing

allegedly infringing content, users cannot issue a counter notification and

threaten suit against copyright owners.

200

Indeed, when content is removed due

to a notice of copyright infringement according to the statutory N&TD regime,

content providers are notified about the removal (except link providers) and

based on this notification, they can file a counter notice, dispute the removal and

force reposting of removed content within ten to fourteen business days.

However, when content is filtered automatically by filtering technologies and

without a takedown notice, content providers cannot threaten suit against

201

Nor can

content providers bring suit against the intermediaries for removing the content

because online intermediaries are usually protected from such suits under their

terms-of-use agreements.

202

Going beyond the statutory framework, voluntary

mechanisms of algorithmic copyright enforcement do not afford alleged

infringers with even the minimum due process protections set by the DMCA:

they do not grant alleged infringers the right to contest content restrictions

through a counter notice procedure, and they do very little in terms of validating

203

Moreover, although the statutory N&TD regime requires complainants to

include in their takedown request an affidavit validating their copyright

198. Wasylik, supra note 3.

199. Lemley, supra note 73, at 115.

200. Sawyer, supra note 26, at 385.

201. Id.

202. Id.

203. The DMCA establishes a procedure—filing a takedown notice requires an affidavit.

17 U.S.C. § 512(c)(3) (2012). If there is no procedure of that sort in voluntary measures, some

materials will be removed without any sufficient legal ground, which could make removal

difficult for the user to contest.

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 509

ownership,

204

voluntary mechanisms occasionally restrict content without any

sufficient legal ground. For instance, in a recent complaint filed against YouTube

under its terms of use, the plaintiff alleged that Content ID removed his parody of

the film The Girl With the Dragon Tattoo, which Content ID designated as being

owned by Pirateria, although Pirateria is not the owner of the rights to this

film.

205

The plaintiff further argued that when he posted, under fair use, a

critique of the 2014 remake of Teenage Mutant Ninja Turtles, a Content ID claim

was made with YouTube on behalf of Viacom, although Viacom is not the true

206

Unfortunately, as mentioned above, target users cannot sue copyright owners

for improper content restrictions, as copyright owners play no active role in

detecting copyright infringement under voluntary regimes. Nor can they sue the

intermediaries for restricting access to their content, since intermediaries, as

private, profit-maximizing entities, can easily prevent these sorts of suits under

their terms-of-use. As a result, target users are left without proper legal recourse

against illegitimate content restrictions by voluntary mechanisms of algorithmic

3. Public Oversight

By removing material after the fact, after receiving a notice of copyright

infringement from a copyright owner, regulated mechanisms of algorithmic

enforcement seem to facilitate public oversight better than voluntary

mechanisms, which restrict access to content ex ante. Nevertheless, given the

immense volume of takedown requests facilitated by algorithmic copyright

enforcement, it may be impossible for the public to follow, inspect and contest

inappropriate content removals, even when content is removed ex post, pursuant

to the DMCA’s N&TD procedure. Indeed, according to Google’s Transparency

Report, in June 2015 Google alone received requests from copyright owners to

remove 39,013,486 URLs.

207

That is more than one million removal requests per

day. Reviewing the practices of such a robust system of online copyright

enforcement is simply impractical.

Voluntary mechanisms of algorithmic copyright enforcement employed by

online intermediaries permit almost no public oversight whatsoever as they

operate behind the veil of the intermediaries’ proprietary code. Hosting platforms

occasionally apply ex ante algorithms to block and filter content before it ever

becomes publicly available,

208

whereas search engines employ algorithms that

downgrade sites that link to allegedly infringing content, without signaling to

204. 17 U.S.C. § 512(c)(3)(A)(vi) (2013).

205. See supra note 2.

206. See supra note 2.

207. Google, Requests to Remove Content Due to Copyright, GOOGLE TRANSPARENCY REPORT

(Apr. 2016), https://www.google.com/transparencyreport/removals/copyright (last visited

May 8, 2016).

208. See infra Part III.C.

510 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

users that some materials are no longer available. Under such circumstances,

public corrections are extremely unlikely. How can the public possibly contest the

removal of content it has never been aware of in the first place?

C. Corporate Copyright: YouTube’s Content ID

In the previous Subpart, we explained three different reasons that contribute

to difficulties in holding copyright enforcement algorithms accountable for

making copyright-related determinations. In the following paragraphs, we

explore the operation of YouTube’s monetizing system of Content ID as an

interesting hybrid of algorithmic copyright enforcement, combining an ex ante

mechanism of algorithmic content blocking with features of DMCA-style

mechanisms of ex post content removal. Content ID’s accountability is

particularly important, not only due to YouTube’s robust implications for public

discourse,

209

but also because Content ID has turned algorithmic copyright

enforcement into a private-financial model of copyright enforcement.

210

Content ID codifies an advanced set of copyright policies and content

management tools, which shift the burden of policing copyright infringement

from copyright owners to identification technologies, while protecting copyright

owners’ interests beyond the basic removal process provided by the DMCA. As a

practical matter, Content ID allows copyright owners to identify their works

using a digital identifying code.

211

Basically, Content ID notifies subscribed

own, offering them one of four choices: (1) mute audio that matches their music;

(2) block a whole video from being viewed; (3) monetize the video by running ads

against it; or (4) track the video’s viewership statistics.

212

When Content ID identifies a match between content provided by a target

209. YouTube is the number one national video website and the third most visited

website globally. See Wendy Boswell, Video Websites: The Top Ten, A

BOUT.COM (Mar. 6, 2016),

http://websearch.about.com/od/imagesearch/tp/popularvideosites.htm

[https://perma.cc/QQ2A-52FR]; see also The Top 500 Sites on the Web, ALEXA,

http://www.alexa.com/topsites [https://perma.cc/XC9S-ETAZ] (last visited Apr. 20, 2016).

210. Content ID currently scans over 400 years-worth of video and utilizes more than 25

million references files of more than 5,000 partners, including US network broadcasters, record

labels, and movie studios. See Statistics, Y

OUTUBE,

http://www.youtube.com/yt/press/statistics.html [https://perma.cc/UV9R-HYPZ] (last

visited Apr. 20, 2016).

211. The technology underlying Content ID relies on digital fingerprinting to sample an

uploaded file and compare it against a database of reference files provided by participating

IMES (Feb. 19, 2007), http://www.nytimes.com/2007/02/19/technology/19video.html

[https://perma.cc/3QVD-VA22] (discussing fingerprinting technologies for identifying audio

and video). For a technical discussion of how fingerprinting is used to identify copyrighted

content, see Craig Seidel, Content Fingerprinting from an Industry Perspective, 2009 IEEE

INT’L

CONF. ON MULTIMEDIA AND EXPO 1524.

212. How Content ID Works, YOUTUBE HELP,

https://support.google.com/youtube/answer/2797370?hl=en [https://perma.cc/PUZ5-C6TR]

(last visited Apr. 20, 2016).

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 511

user and a copyrighted work on file, it automatically sends a Content ID claim to

the target user. At this point the user can either: (1) acknowledge the claim; (2) if

the claim is for a piece of music in the video, choose to remove the song without

having to edit and reload; (3) swap out the allegedly infringing song with a free-

to-use song; (4) share revenue with the copyright owner; or (5) dispute the

claim.

213

The system includes a very detailed mechanism of disputing a Content ID

claim. If the target user elects to dispute the claim, the copyright owner gets thirty

days to respond.

214

Failure to respond promptly causes the Content ID claim to

expire.

215

The copyright owner can respond to a dispute in several different ways:

(1) release the claim; (2) uphold the claim; (3) takedown the video. If the

decision.

216

If the user appeals the owner’s decision to uphold the Content ID

claim, the copyright owner gets thirty days to respond. After the user appeals, the

217

If the owner elects to takedown the video, the user receives a copyright strike

in her account.

218

This means that her video has been taken down from YouTube

following the legal request of the copyright owner. Receiving a copyright strike

puts the user’s account in bad standing, which means that she may lose access to

certain YouTube features.

219

A user cannot resolve a copyright strike by deleting

the allegedly infringing video. Instead, she can either: (1) wait six months for the

213. Dispute a Content ID Claim, YOUTUBE HELP,

https://support.google.com/youtube/answer/2797454 [https://perma.cc/8FHL-HFVJ] (last

visited Apr. 20, 2016).

214. Id.

215. Id.

216. Id. This depends on whether the user’s account is in good standing on the date of the

appeal and possibly on other undisclosed factors. An account is in “good standing” when it has

no Community Guidelines strikes (content removals that are not copyright-related, i.e. sexual

content or violence), no copyright strikes, and no more than one video blocked worldwide by

Content ID. Keep Your YouTube Account in Good Standing, Y

OUTUBE HELP,

https://support.google.com/youtube/answer/2797387 [https://perma.cc/4YZA-GLFR] (last

visited Apr. 20, 2016).

217. What Happens After I Dispute, YOUTUBE HELP,

https://support.google.com/youtube/answer/2797454 [https://perma.cc/DRA7-GGKL] (last

visited Apr. 20, 2016).

218. Id.

219. Particularly, she will be restricted from uploading videos as unlisted, uploading

videos that are longer than 15 minutes, uploading videos under a Creative Commons license,

InVideo programming, customizing video thumbnails, live events and hangouts on Air,

appealing rejected Content ID claim disputes, sharing private videos, and using YouTube Video

Editor. In addition, a user receiving a copyright strike will be restricted from using certain

channel features and joining the YouTube Partner Program (allowing creators to monetize

content on YouTube through a variety of ways including advertisements, paid subscriptions,

and merchandise). Keep Your YouTube Account in Good Standing, Y

OUTUBE HELP,

https://support.google.com/youtube/answer/2797387 [https://perma.cc/4YZA-GLFR] (last

visited Apr. 20, 2016).

512 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

additional copyright strikes during this period of time; (2) contact the person who

claimed the video and ask him to retract his claim of copyright infringement;

(3) submit a counter notification, if the user believes her video either qualifies as

fair use or is not infringing.

220

According to YouTube’s website, a counter notification is:

A legal request for YouTube to reinstate a video that has been removed for

alleged copyright infringement. The process may only be pursued in instances

where the upload was removed or disabled as a result of a mistake or

misidentification of the material to be removed or disabled, such as fair use. It

should not be pursued under any other circumstances.

221

Only the original uploader or an agent authorized to act on her behalf may

submit a counter notification.

222

YouTube forwards the counter notification to

the party who submitted the original Content ID claim of copyright

infringement.

223

This process takes ten days to complete.

224

Users whose

accounts have been suspended for multiple copyright violations cannot access the

counter notification webform, and may instead submit a free-form counter

notification.

225

The major advantage offered by Content ID is the monetizing option. Rather

than simply blocking allegedly infringing materials, the monetizing feature of

Content ID promotes copyright licensing agreements, enabling copyright owners

to permit allegedly infringing uses of their works in return for a share in the

economic benefit generated by advertisements displayed with the content.

226

The

vast majority of copyright holders elect not to remove infringing content, but

instead to monetize it,

227

so much so that video monetizing is becoming a

successful business model.

228

Indeed, copyright owners can build a profitable

online licensing business through which they can obtain revenue easily from

YouTube users who make use—any use—of their copyrighted material. It is not

that copyright licensing was non-existent before the emergence of Content ID; it

220. How to Resolve a Copyright Strike, YOUTUBE HELP,

https://support.google.com/youtube/answer/2814000 [https://perma.cc/9NM5-5AMC] (last

visited Apr. 20, 2016).

221. Id.

222. Counter Notification Basics, YOUTUBE HELP,

https://support.google.com/youtube/answer/2807684?hl=en-GB [https://perma.cc/YXL8-

KWS3] (last visited Apr. 20, 2016).

223. Id.

224. Id.

225. Id.

226. Yafit Lev-Aretz, Second Level Agreements, 45 AKRON L. REV. 137, 152 (2012).

227. Id. at 158.

228. To name one example, Rumblefish, Inc., recently acquired by SESAC, is a full service

music micro-licensing platform for rights-holders providing monetization on YouTube. Bruce

Houghton, Online Video Monetization Heats up: Q&A with Rumblefish Founder/CEO on Acquisition

by SESAC, H

YPEBOT.COM (Aug. 13, 2014), http://www.hypebot.com/hypebot/2014/08/online-

video-monetization-heats-up-qa-with-rumblefish-founderceo-on-acquisition-by-sesac.html

[https://perma.cc/6PX4-JG5M].

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 513

is just that the latter made this process shorter, clearer, and more efficient.

229

Further, the automatic identification technology lowers the transaction costs

associated with entering into licensing agreements.

230

It identifies the potential

parties (a copyright owner on the one side, and a user on the other) and facilitates

a contract by a click of button. Moreover, this business model effectively flips the

fundamental default established by copyright law: authorize first, use later. With

Content ID, owners allow users to experiment with their works without giving

them any explicit license, while exercising their rights to get a share in revenue

only after the (unlicensed) derivative uses become commercially viable (use first,

authorize later).

Standing alone, facilitating copyright licensing in an era of mass infringement

seems both welcomed and appreciated.

231

Nevertheless, Content ID may err in

reporting a technological match over fair use,

232

forcing users to acquire a license

where they should in fact be free to use the materials.

233

Furthermore, in the

absence of a reliable procedure for validating copyright ownership, some works

may be erroneously claimed by non-copyright-owners.

234

Considering the crucial

role YouTube plays in shaping public discourse,

235

it is critical to hold Content ID

accountable for its determinations. That said, figuring out how Content ID

exercises its power is extremely challenging because it operates behind the veil of

a proprietary code that primarily adheres to YouTube’s business interests. Against

this backdrop, we next examine the accountability of Content ID.

1. Transparency

Content ID allows content providers to upload digital copies of video or

audio works, which YouTube’s servers use to create a digital reference against

which all other videos on the site are scanned. If even a portion of another video

matches the sample in either its visual or audio content, the video is flagged as

containing that copyrighted content.

236

It is unclear, and hence unpredictable,

229. Lev-Aretz, supra note 226, at 158.

230. Paul J. Heald, How Copyright Makes Books and Music Disappear 30 (Illinois Public

Research Paper No. 13-54, 2013),

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2290181 [https://perma.cc/6X2D-

BUHV].

231. Id.

232. See Maayan Perel & Niva Elkin-Koren, Black Box Tinkering: Beyond Disclosure in

Algorithmic Enforcement, F

LA. L. REV. (forthcoming 2017).

233. Note that it is not unlikely that alleged infringers will accept an unfair deal and agree

to share advertising profit with copyright owners, notwithstanding their non-infringing uses.

Under Content ID’s rapid takedown policy, which allows content to be removed automatically

before, and sometime absent, prior notice to content providers and without adequate legal

scrutiny (both adjudication before removal and review after removal), agreeing to share profit

with copyright owners at least guarantees both the access to the popular distribution channel

and the prospective income, whereas declining such a deal unfortunately risks losing both.

234. See supra note 79 and accompanying text.

235. See supra Part II.C.2

236. Lauren G. Gallo, The (Im)possibility of “Standard Technical Measures” for UGC Websites,

514 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

what exact portion of copyrighted material must be embedded in an upload to

trigger the system.

237

As a result, Content ID may unlawfully flag fair uses or de

minimis uses of content.

238

Because the thresholds of Content ID are not

disclosed,

239

users cannot appreciate how it exercises its power. Indeed, the non-

transparent nature of Content ID’s identification procedure curtails the ability of

alleged infringers to learn whether their videos were flagged due to an alleged

2. Due Process

Content ID implements the DMCA’s counter notice feature in a way that

diminishes due process. As noted earlier,

240

an alleged infringer must dispute a

Content ID claim before availing herself of the counter notice application.

241

Yet,

YouTube implicitly discourages users from filing a dispute, emphasizing that an

invalid dispute may cause the copyright owner to takedown the allegedly

infringing video, and when this happens, the target user’s account could be

subject to restrictions imposed by a copyright strike.

242

Note, that a copyright

strike is against a specific user and not against a specific account. Thus, a user can

open as many accounts as she wishes, so long she does not receive a total of three

accounts are being terminated. Target users may hence be induced not to dispute

a Content ID claim, especially when the contested content is timely.

243

Neglecting

the counter notice process is especially problematic when the blocked content

could have been allowed as fair use: target users are obviously deprived of their

right to share and reap financial benefit from their own creations, but there are

also serious policy concerns resulting from the fact that the public as a whole is

precluded from hearing, viewing and experiencing the disputed content.

244

YouTube further compromises due process when it fails to verify the rights

34 COLUM. J.L. & ARTS 283, 296 (2011).

237. Id.

238. See Perel &Elkin-Koren, supra note 232.

239. They may even vary among copyright owners.

240. See supra note 192 and accompanying text.

241. See supra note 192 and accompanying text.

242. See supra note 218 and accompanying text.

243. Urban & Quilter, supra note 99, at 626. For instance, during the 2008 McCain-Palin

campaign, several political videos were removed from the McCain campaign’s YouTube

channel for copyright infringement allegations. McCain-Palin’s counsel urged YouTube to

make an exception for the videos posted by political candidates and campaigns, suggesting that

YouTube commit to a legal review of these political videos and decline to remove clearly non-

infringing material, rather than taking down and insisting on the DMCA waiting period of ten

to fourteen business days. This was because the political speech that was removed from the

McCain-Palin YouTube channel took place at the height of election season, making it pointless

to wait for the material to be reinstated. See Wendy Seltzer, Free Speech Unmoored in Copyright’s

Safe Harbor: Chilling Effects of the DMCA on the First Amendment, 24 H

ARV. J.L. & TECH 171, 173

(2010).

244. See supra Part II.C.2

Spring 2016] ALGORITHMIC COPYRIGHT ENFORCEMENT 515

claimed by copyright owners.

245

When a claimant creates a reference file, either

by uploading content or by claiming an already uploaded video, he effectively

asserts ownership over the content in the reference file.

246

Apparently, YouTube

does not require sufficient proof of ownership in claimed works.

247

This flaw is

troubling: users who manage to gain admission to the Content ID system can

upload any content into the system, which would later flag videos as belonging to

them. As a result, swindlers may hijack ad revenue of videos from users (and

possibly from the true right-holders), regardless of whether they really own any

248

GoDigital Media Group

249

is a prominent example of this sort of conduct.

Serving copyright owners who wish to identify and monetize their copyrighted

works in UGC, GoDigital is responsible for thousands of illegitimate copyright

claims on YouTube. AdSHARE,

250

one of GoDigital’s five companies, is

responsible for monetizing fan engagement online, by identifying, tracking and

monetizing user-uploaded versions of copyright owners’ content on social media

websites. AdSHARE works on music compositions, sound recordings, and video,

and is able to identify even short snippets of copyrighted content used on social

media platforms, such as YouTube, Google, Facebook and Soundcloud.

Unfortunately, AdSHARE does not always verify whether its clients actually own

the rights to particular works before it begins monetizing them automatically and

therefore, it ends up claiming a wide variety of royalty free, public domain, and

Creative Commons content.

Finally, YouTube’s Content ID arbitrarily favors dominant copyright holders

over small creators, possibly depriving the latter of substantial due process. As

described earlier, YouTube has publicly admitted that, when it comes to videos

containing content from certain partner companies, it will not repost them

following their removal, even if the video uploaders file a counter notice showing

that the videos constitute fair use.

251

As a result, some alleged infringers targeted

by YouTube’s Content ID might be left with no meaningful recourse to confront

unlawful content removals. Even though they can technically dispute content

takedowns, such disputes are effectively pointless.

245. See supra notes 8 and accompanying text.

246. What is a Policy?, YOUTUBE HELP,

https://support.google.com/youtube/answer/107383?hl=en&ref_topic=24332

[https://perma.cc/ZCC4-DHEQ] (last visited Apr. 18, 2016).

247. Patrick McKay, YouTube Copyfraud & Abuse of the Content ID System, FAIRUSETUBE

(Nov. 23, 2011, 3:22 PM), http://fairusetube.org/youtube-copyfraud [https://perma.cc/82TY-

LKM8].

248. Id.

249. GODIGITAL MEDIA GROUP, http://www.godigitalmg.com [https://perma.cc/B36R-

X8Z7] (last visited Apr. 23, 2016).

250. AdSHARE, http://adshare.tv [https://perma.cc/DUC5-FW8Z] (last visited Nov. 15,

2016).

251. See supra notes 195-196 and accompanying text.

516 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

3. Public Oversight

Because YouTube’s Content ID allows the ex-ante blocking of videos, it

curtails the ability of the public to promptly render its own fair use judgment.

252

Without the opportunity to experience the content of blocked videos, almost no

room is left for public pressure to demand the reversal of the system’s

determination. While users could resort to uploading videos protesting automatic

blocking,

253

they could not include the underlying infringing material in their

protest videos, so they are unlikely to generate sufficient public pressure to

demand the repost of blocked videos.

254

Hence, outside the imperfect option of filing a counter notice against

allegedly unjustified takedowns, the filtering algorithm embedded in Content ID

seems to remain wild and free. While it arguably makes the licensing process

shorter, clearer and more efficient,

255

it imposes unwarranted restrictions on

non-infringing materials and fair uses of content. By automatically locating and

monetizing digital uses of copyrighted content, the system makes it unnecessary

for owners to search and notify YouTube of infringing content. Consequently,

Content ID results in many false positives,

256

or instances where Content ID

automatically blocks or monetizes a video that does not actually contain

infringing content.

257

For example, YouTube has recently even flagged a cat

purring as copyright infringing music.

258

Finally, by lowering the transaction

costs associated with entering into licensing agreements,

259

Content ID may

result in unjust enrichment when targeted users are unjustifiably dragged into

paying to share non-infringing content or fair use videos.

IV. E

NHANCING ACCOUNTABILITY: BARRIERS AND STRATEGIES

In the previous Parts, we explained why accountability is important in

algorithmic copyright enforcement, and why existing mechanisms fail to

adequately promote it. In particular, we demonstrated that the non-transparent

252. Sawyer, supra note Error! Bookmark not defined., at 394.

253. See, e.g., Fred von Lohmann, YouTube’s January Fair Use Massacre, ELEC. FRONTIER

FOUND. (Feb. 3, 2009), http://www.eff.org/deeplinks/2009/01/youtubes-january-fair-use-

massacre [https://perma.cc/CC2T-LTEY].

254. Sawyer, supra note Error! Bookmark not defined., at 394.

255. Lev-Aretz, supra note 227, at 158.

256. Sawyer, supra note Error! Bookmark not defined., at 382-83.

257. Ben Depoorter & Robert Kirk Walker, Copyright False Positives, 89 NOTRE DAME L.

REV. 319, 326 (2013); see also Endgame Removed from YouTube by False Copyright Claim,

NFOWARS (Feb. 15, 2015), http://www.infowars.com/endgame-removed-from-youtube-by-

false-copyright-claim [https://perma.cc/9J73-2YZV].

258. Ernesto, YouTube Flags Cat Purring as Copyright Infringing Music, TORRENTFREAK

(Feb. 11, 2015),

http://torrentfreak.com/youtube-flags-cat-purring-as-copyright-infringing-music-150211

[https://perma.cc/P3Z7-9PMG].

259. Heald, supra note 230, at 37.

nature of copyright enforcement algorithms prevents users from understanding

how the algorithms exercise their power; that such algorithms further diminish

users’ due process by affording them inadequate challenging opportunities; and

that their ubiquity and frequent ex ante application curb public oversight by

diminishing the process of error correction. In this final Part, we identify the

barriers for enhancing accountability in algorithmic copyright enforcement by

online intermediaries and discuss various strategies for removing them.

A. Mapping the Barriers to Algorithmic Accountability

This Subpart divides the barriers for enhancing accountability in algorithmic

categories: the first category explores technical barriers, namely the non-

transparent nature of algorithms, as well as the operation of copyright

enforcement algorithms, which often relies on constantly evolving learning

machines. The second category covers legal barriers that impede public literacy in

relation to algorithmic copyright enforcement exploited by online intermediaries.

The third category focuses on practical barriers that effectively negate the

purported objectives of the counter notice procedure under the DMCA, resulting

in insufficient dispute opportunities.

1. Technical Barriers

There are two intertwined technical barriers to adequate public scrutiny of

algorithmic copyright enforcement. The first is the non-transparent nature of

algorithms, which makes it difficult to review their decision-making processes.

260

Algorithmic copyright enforcement embeds a high degree of automation, even if

it may sometimes involve some degree of human analysis.

261

While algorithms

can be built to advance specific values and policies,

262

they are ultimately complex

260. This is not to say, however, that algorithmic copyright enforcement merits a higher

level of transparency than what manual copyright enforcement demands. Transparency “should

be applied to all steps which might compromise rights of individuals and seem arbitrary, be

they automated or manual. The level of automation needs not, on its own, merit a higher level

of transparency.” See Zarsky, supra note 34, at 1552. Both regimes—the automated one and the

human-implemented one—should eventually reach a similar degree of transparency, yet

automated regimes, by their nature, inherently challenge this goal.

261. Steve Lohr, Algorithms Get a Human Hand in Steering Web, N.Y. TIMES (Mar. 10, 2013),

http://www.nytimes.com/2013/03/11/technology/computer-algorithms-rely-increasingly-on-

human-helpers.html?pagewanted=all&_r=1& [https://perma.cc/X8GF-37SA] (demonstrating

how the resources of intermediaries, such as Google and Twitter, are becoming more human

curated).

262. Nissenbaum, supra note 15, at 1373; Bruno Latour, Where Are the Missing Masses? The

Sociology of a Few Mundane Artifacts, in S

HAPING TECHNLOGY/BUILDING SOCIETY 225, (Wiebe

Bijker & John Law eds., 1992). Regarding copyright enforcement specifically, see R. Polk

Wagner, Reconsidering the DMCA, 42 H

OUS. L. REV. 1107 (2005) (suggesting that the total

regulatory effect combines both law and technology and that changes in technology affect the

law and vice versa).

518 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

codes that we (and most program developers

263

) cannot easily deconstruct. The

ability to integrate an almost unlimited number of variables into automated

decision rules means that algorithms “can successfully apply rules whose

complexity would make them collapse under their own weight if humans were

forced to apply them.”

264

As Bamberger explains, “programming and

mathematical idiom can shield layers of embedded assumptions from high-level

firm decision makers charged with meaningful oversight and can mask important

concerns with a veneer of transparency.”

265

Furthermore, enforcement algorithms used by online intermediaries

effectively advance the intermediaries’ own interpretation of legal norms.

266

This

process of translating legal mandates into code inevitably embodies particular

choices as to how the law is interpreted, which may be affected by a variety of

extrajudicial considerations, including the conscious and unconscious professional

assumptions of program developers, as well as various private business

incentives.

267

It may even be affected by automated bias.

268

As Citron stresses,

policy distortions can arise when possibly biased code writers, who lack “policy

knowledge,” translate policy from human language to code.

269

Some disparity

between the algorithmic representation of law and the law as it operates in

practice is hence unavoidable.

270

Comprehending algorithmic systems of law

enforcement is hence a complex task, which demands knowing what cognitive

frames of reference, as well as social, political, economic, and legal motivations

shaped the choices made by those who designed them.

271

The second technical barrier to proper accountability in algorithmic

learning enables the identification of trends, relationships and hidden patterns in

disparate groups of data.

272

Some algorithms can thus adapt their code and shape

263. Frank Pasquale, Restoring Transparency to Automated Authority, 9 J. ON TELECOMM. &

HIGH TECH. L. 235, 246 (2011).

264. Grimmelmann, supra note 69, at 1734.

265. Bamberger, supra note 15, at 727.

266. Id. at 675.

267. Id. at 675-76.

268. Id. at 675 (explaining that automation biases are “decision pathologies that hinder

careful review of automated outcomes, especially by those with financial incentives that

promote risky behavior. These very phenomena contributed to the failure of risk regulation

and risk management to prevent the recent financial meltdown”); Friedman & Nissenbaum,

supra note 105, at 333.

269. Citron, supra note 11, at 1261-62.

270. Harry Surden et al., Representational Complexity in Law, 11 INT’L CONF. ON ARTIFICIAL

INTELLIGENCE & L. 193, 193 (2007).

271. See Jay P. Kesan & Rajiv C. Shah, Deconstructing Code, 6 YALE J.L. & TECH. 277, 283

(2004) (“Science & Technology Studies examines how technology is shaped by societal factors

such as politics, institutions, economics, and social structures.”).

272. Bernhard Anrig et al., The Role of Algorithms in Profiling, in PROFILING THE EUROPEAN

CITIZEN: CROSS-DISCIPLINARY PERSPECTIVES 65 (Mireille Hildebrandt & Serge Gutwirth, eds.,

2008), http://link.springer.com/chapter/10.1007%2F978-1-4020-6914-7_4

[https://perma.cc/JCD7-EYK8].

performance based on experience. For instance, Google and Facebook may run

dozens of different versions of an algorithm to assess their relative merits, with

no guarantee that the version a user interacts with at one moment in time is the

same as five seconds earlier.

273

Indeed, algorithms are often designed to be

reactive and mutable to inputs.

274

The EdgeRank algorithm which Facebook uses

to determine what posts, and in what particular order, to feed into each user’s

timeline works in concert with each individual user, ordering posts in accordance

with how one interacts with “friends.”

275

These learning capacities make algorithms significantly less predictable than

some computer codes that operate by means of on-off rules,

276

further

diminishing their already-deficient accountability. While we can easily judge why

an automated camera issues a speeding a ticket, determining whether removal,

filtering, or blocking algorithms legitimately vindicate the rights of copyright

owners is quite challenging.

277

For these algorithms are not merely tools for

implementing the goals of the intermediaries employing them; they may

practically shape the meaning of those goals themselves. Under Bamberger’s

interpretation, “they create a Gestalt, or world view, that alters the perceptions of

the decision makers they inform.”

278

A different view may suggest that the learning capacities of law enforcement

algorithms should be weighed in favor of public oversight, and not against it.

Particularly, the elasticity of learning machines,

279

or their ability to reshape and

adapt code as circumstances demand,

280

arguably make them more adjustable to

mistake correction.

281

This is especially so due to their central locations, which

enable quick changes.

282

Nevertheless, after the initial introduction of technology,

its flexibility mostly vanishes as it transitions from a device “oriented toward

human needs” to “an important component of the formative context.”

283

Much

273. Rob Kitchin, Thinking Critically About and Researching Algorithms 16 (The

Programmable City Working Paper No. 5, 2014),

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2515786 [https://perma.cc/6YRN-

HNZZ]

274. Id.

275. Taina Bucher, Want to Be on the Top? Algorithmic Power and the Threat of Invisibility on

Facebook, 14 N

EW MEDIA & SOC’Y 1164 (2012).

276. Bamberger, supra note 15, at 676.

277. Pasquale, supra note 263, at 237.

278. Bamberger, supra note 15, at 676.

279. Grimmelmann, supra note 69, at 1730-31.

280. Bamberger, supra note 15, at 710.

281. See Hal R. Varian, Kaizen, that Continuous Improvement Strategy, Finds Its Ideal

Environment, N.Y.

TIMES (Feb. 8, 2007),

http://www.nytimes.com/2007/02/08/business/08scene.html [https://perma.cc/5U3Z-

4MUZ]; see also Sawyer, supra note Error! Bookmark not defined., at 384 (creating a

distinction between Content ID, which is much more adaptable to change because its central

location enables quick changes, and Digital Rights Managements, which are much more

difficult to fix without rolling out a second generation of technology).

282. Varian, supra note 281.

283. Claudio U. Ciborra, De Profundis? Deconstructing the Concept of Strategic Alignment, 9

520 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

like legislative acts, technology, too, establishes a set of guidelines for public order

that may persist for many generations. Unable to understand and deconstruct the

layers of codes and policies embedded in law enforcement algorithms, we expect

the general public to largely work its way outside this complex regime, rather

than contest it.

2. Legal Barriers

Several legal barriers also obstruct the public’s ability to hold algorithmic

mechanisms of copyright enforcement accountable. First, copyright law may

hinder public literacy that could be pursued through independent, reverse

engineering research.

284

Theoretically, users can engage in self-help and reverse

engineer copyright enforcement algorithms to learn how they exercise their

power. This is exactly what Glen Gabe, a digital marketing veteran, tried to do

with Google’s Pirate algorithm.

285

As part of his attempt to figure out the

threshold Google uses when choosing to target a pirated domain by making it

appear lower in its search results, Gabe analyzed the data on Google’s

Transparency Report, including copyright takedown notices, domains being

specified in those takedowns and top copyright owners. After analyzing the

organic search trends of affected sites, Gabe concluded that the Pirate algorithm

probably kicks in when at least half of one’s indexed URLs are subject to

takedown requests.

286

By analyzing the output of this unknown algorithm, i.e.

how many takedown requests were received by websites that were ultimately

targeted by the Pirate algorithm, Gabe effectively enhanced users’ literacy

regarding the metric underlining Google’s search engine mechanism,

consequently enhancing its accountability.

287

Notwithstanding the DMCA’s important benefit in affording the public self-

help measures that are capable of extracting valuable information on algorithmic

governance, the DMCA’s anti-circumvention provisions

288

may suppress this

SCANDINAVIAN J. INFO. SYS. 67, 76 (1997).

284. See Maayan Perel & Niva Elkin-Koren, Black Box Tinkering: Beyond Disclosure in

Algorithmic Enforcement, 69 F

LA. L. REV. (forthcoming 2017).

285. See supra note 183 and accompanying text.

286. See supra note 183 and accompanying text.

287. See Diakopoulos, supra note 75 (discussing similar strategies of reverse engineering).

288. 17 U.S.C. §§ 1201-1202 (2012). The most pertinent of the DMCA’s anti-

circumvention provisions read in part:

(a) Violations Regarding Circumvention of Technological Measures.—

(1)(A) No person shall circumvent a technological measure that effectively controls access to a

work protected under this title.

. . . .

(2) No person shall manufacture, import, offer to the public, provide, or otherwise traffic in any

technology, product, service, device, component, or part thereof,

that—

(A) is primarily designed or produced for the purpose of circumventing .. ;

(B) has only limited commercially significant purpose or use other than to circumvent . . . ; or

knowledge for use in circumventing . . .

sort of reverse engineering research. Generally, the anti-circumvention

provisions prohibit the use, development, or distribution of technologies which

are designed to “circumvent” (e.g., hack, crack, or break) access control

systems,

289

with some narrow exceptions.

290

Defenders of strong copyright called

for this legal intervention, acknowledging that technological protection measures

(TPM) and digital rights management systems (DRM) are not tamper-proof.

291

Nevertheless, during their first decade of existence, the anti-circumvention

provisions were thought to have a very restrictive impact.

292

At that time, the

prevalent technological measures were basically encrypted computer codes that

were incorporated into fixations of copyrighted material, such as CDs or music

files, to prevent illegal copying and public distribution of copyrighted works.

293

Meaning, the anti-circumvention provisions initially applied to technologies that

restricted access to specific content, and not to particular distribution channels.

294

Therefore, the basic assumption was that the “anti-circumvention rules do not

generally affect the users of copyrighted work.”

295

The overwhelming majority of

users—namely lay computer users—were presumed to be in the same position

after the DMCA’s anti-circumvention enactment as before it.

296

This assumption seems to collapse under algorithmic copyright enforcement

by online intermediaries. The application of online technologies that restrict the

distribution of allegedly infringing content effectively pushes online

intermediaries’ negative impact in suppressing the development of innovative

circumvention technologies forward.

297

Originally designed to prohibit users

. . . .

(b) Additional Violations.—

(1) No person shall manufacture, import, offer to the public, provide, or otherwise traffic in any

technology, product, service, device, component, or part thereof,

that—

(A) is primarily designed or produced for the purpose of circumventing protection . . . ;

(B) has only limited commercially significant purpose or use other than to circumvent

protection . . . ; or

knowledge for use in circumventing protection . . .

17 U.S.C. § 1201(a)-(b).

289. See supra note 288 and accompanying text.

290. Pamela Samuelson, Intellectual Property and the Digital Economy: Why the Anti

Circumvention Rules Need to be Revised, 14 B

ERKELEY TECH. L.J. 519 (1999) (explaining that

circumvention is permissible for some limited purposes, such as achieving program-to-

program interoperability or engaging in encryption research and computer security testing).

291. Id.

292. Stephen M. Kramarsky, Copyright Enforcement in the Internet Age: The Law and

Technology of Digital Rights Management, 11 D

EPAUL-LCA J. ART & ENT. L. & POL’Y 1, 10 (2001)

(“[T]he new anti-circumvention laws prevent sophisticated users from bypassing the

technology.”).

293. Id.

294. See supra note Error! Bookmark not defined. and accompanying text.

295. Wagner, supra note 262, at 1124.

296. Id. at 1125.

297. Id. (The DMCA “simultaneously suppresses and encourages technology. That is, on

the one hand it encourages the deployment of ‘access control’ technologies on copyrighted

522 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

from developing ways to deconstruct a single-content access control, the anti-

circumvention provisions could now be interpreted to discourage users from

challenging technological measures that monitor popular channels of content

distribution.

298

In contrast to DRM technologies, which are bi-directional—

defining the direct relationship between a copyright owner and the holder of a

specific copy of the copyrighted content—mechanisms of algorithmic copyright

enforcement have far-reaching implications over public discourse.

299

Having the

ability to circumvent mechanisms of algorithmic copyright enforcement thus

reaches beyond the narrow interests of lawful owners of specific copies of

copyrighted content and curious technologists because it enables users to contest

what some scholars have characterized as the “privication” of information that

would have otherwise been public.

300

Nevertheless, when technologists have attempted to reverse engineer

filtering programs used in public schools, libraries, and similar institutions to

protect minors from exposure to indecent or otherwise harmful material posted

online, the developers of the filtering programs have been successful in arguing

that reverse engineering the encryption to analyze the list of blocked sites violated

the DMCA’s anti-circumvention rules.

301

In Edelman v. N212, the court refused to

declare that reverse engineering the software was lawful, even though that

information was critically important to a back-then public policy debate over

whether legislatures should mandate use of filtering software in public schools

and libraries.

302

By virtue of comparison, there is a genuine risk that the DMCA’s

anti-circumvention provisions would be invoked to ban the deconstruction of

online copyright enforcement technologies employed by online intermediaries,

consequently refuting its important accountability-enhancing potential.

Another legal obstacle to enhancing accountability in algorithmic copyright

enforcement is created by trade secrecy law. In a famous legal battle between

Viacom and YouTube, a judge refused to force YouTube to provide Viacom with

the computer source code which controls both the YouTube.com search function

and Google’s internet search tool “Google.com.”

303

The court explained that “the

search code is the product of over a thousand person-years of work” and that

works . . . On the other hand, it prohibits the use, development, or distribution of

‘circumvention’ technologies”).

298. See supra Part II.C.2.

299. See supra Part II.C.2.

300. Privication describes the possibility of private publication, where content providers

distribute content on a large-scale but at the same time retain control over access. See Jonathan

Zittrain, What the Publisher Can Teach the Patient: Intellectual Property and Privacy in an Era of

Trusted Privication, 52 S

TAN. L. REV. 1201, 1203 (2000).

301. Press Release, ACLU, In Legal First, ACLU Sues over New Copyright Law: Says

Blocking Program Lists Should Be Revealed (July 25, 2002), https://www.aclu.org/news/legal-

first-aclu-sues-over-new-copyright-law-says-blocking-program-lists-should-be-revealed

[https://perma.cc/4GSR-LK8Q].

302. 263 F. Supp. 2d 137, 138 (D. Mass. 2003).

303. Viacom Int’l, Inc. v. YouTube, Inc., No. 1:07-cv-02103-LLS, 2008 U.S. Dist. LEXIS

50614 (S.D.N.Y. Jul. 2, 2008).

“there is no dispute that its secrecy is of enormous commercial value. Someone

with access to it could readily perceive its basic design principles, and cause

catastrophic competitive harm to Google by sharing them with others who might

create their own programs without making the same investment.”

304

Although

the production and examination of the source code was, according to Viacom, the

only way to find out whether YouTube’s search algorithm effectively increases the

rank or visibility of allegedly infringing material over non-infringing material,

305

the court denied Viacom’s motion to compel YouTube to produce its source code

and consequently lose its trade secret.

As part of the growing recognition of the economic benefits streaming from

protecting intangible assets, it would be reasonable to expect firms to claim an

increasingly broad range of non-public information as trade secrets.

306

Trade

secrecy law encourages intermediaries to keep their proprietary algorithms secret

because once they are revealed, they lose trade secret protection as a matter of

law.

307

As demonstrated by Viacom Int’l, Inc. v. YouTube, Inc., protecting the secrecy

of proprietary codes is necessary not only to prevent intentional infringement,

but also to prevent competitors from free riding on an intermediary’s economic

investment in developing its codes.

Nonetheless, secrecy has major negative consequences for society because it

obstructs accountability.

308

In the digital realm, which largely enhances

consumers’ capacity to actively engage in creative processes,

309

trade secrecy can

further undermine access to information.

310

Accordingly, keeping copyright

enforcement algorithms secret is particularly problematic because “they are de

facto sovereigns over important swaths of social life.”

311

Their invisible hand

effectively controls what content is available and further determines its ranking

and accessibility.

312

Yet so long as copyright enforcement algorithms are

potentially trade secrets, investigating their misconduct may never reach a

conclusive end,

313

notwithstanding their bothering impact on public discourse.

304. Id. at *8.

305. Id. at *10.

306. See, e.g., Robert P. Merges, One Hundred Years of Solicitude: Intellectual Property Law,

1900-2000, 88 C

AL. L. REV. 2187, 2233-40 (2000) (discussing expansions in intellectual property

protection during the twentieth century).

307. Pasquale, supra note 263, at 237.

308. David S. Levine, Secrecy and Unaccountability: Trade Secrets in Our Public Infrastructure,

59 F

LA. L. REV. 135, 170-77 (2007); Mary L. Lyndon, Information Economics and Chemical

Toxicity: Designing Laws to Produce and Use Data, 87 M

ICH. L. REV. 1795, 1855-56 (1989).

309. Niva Elkin Koren, Making Room for Consumers Under the DMCA, 22 BERKLEY TECH.

L.J. 1119, 1120 (2007).

310. Pasquale, supra note 263, at 245-46.

311. DAVID STARK, THE SENSE OF DISSONANCE: ACCOUNTS OF WORTH IN ECONOMIC LIFE 1

(2011).

312. See supra Part II.C.2.

313. Pasquale, supra note 263, at 245. Google’s secrecy about its website-ranking algorithm

has provoked unsuccessful investigations in Europe and the U.S., however. See Editorial, The

Google Algorithm, N.Y. TIMES (July 15, 2010),

http://www.nytimes.com/2010/07/15/opinion/15thu3.html (last visited May 8, 2016) (“[T]he

524 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

Finally, beyond the DMCA’s anti-circumvention provisions and trade secrecy

law, general “computer intrusion” and “anti-hacking” laws form another legal

barrier to enhancing accountability in algorithmic copyright enforcement.

314

Indeed, state and federal statutes further protect computer network owners from

hacking and unauthorized intrusions, including the Computer Fraud and Abuse

Act (CFAA),

315

the Wiretap Act,

316

the Electronic Communications Privacy Act

(ECPA),

317

and a variety of state computer intrusion statutes. These laws

criminalize different conducts of computer hacking, generally subject to a

financial damage threshold requiring that a plaintiff prove that the intrusion

caused some harm.

318

This criminalization includes the penetration of computer

systems to gain knowledge about their operation,

319

which could further restrain

users from investigating copyright enforcement algorithms.

3. Practical Barriers

A third barrier to enhancing accountability in algorithmic copyright

enforcement relates to the practical ineffectiveness of the counter notification

procedure under the DMCA. Under § 512(g)(2)(B), hosting services are required

to promptly forward any counter notices from alleged infringers back to the

original complainant.

320

If after ten to fourteen days, the complainant does not

notify the OSP that she had filed a lawsuit, then the OSP must reinstate the

contested material. Although this is a simple self-help procedure allowing users to

easily contest improper takedown removals, alleged infringers hardly employ it in

practice.

321

Bruce Boyden has recently found that the Motion Picture Association

of America (MPAA) sent twenty-five million takedown notices to search engines

potential impact of Google’s algorithm on the Internet economy is such that it is worth

exploring ways to ensure that the editorial policy guiding Google’s tweaks is solely intended to

improve the quality of the results and not to help Google’s other businesses.”); Richard Waters,

Unrest over Google’s Secret Formula, F

IN. TIMES (July 12, 2010),

http://www.ft.com/cms/s/0/1a5596c2-8d0f-11df-bad7-00144feab49a.html (last visited May 8,

2016).

314. See Elec. Frontier Found., The CFAA: Blocking Competition and Stifling

Innovation, https://www.eff.org/files/filenode/cfaa-stifling-innovation.pdf

[https://perma.cc/97R3-PCT3] for examples of how anti hacking laws prohibit reverse

engineering, yet with emphasis on innovation, not accountability.

315. 18 U.S.C. § 1030 (2012).

316. 18 U.S.C. §§ 2510-2522 (2012).

317. 18 U.S.C. §§ 2701; 3121 (2012).

318. Id.; see also 18 U.S.C. § 1030(a)(4) (2012), which addresses the access and fraudulent

use of a protected computer and is triggered if the value of the use obtained exceeds $5,000.

319. Eric J. Sinrod & William P. Reilly, Cyber-Crimes: A Practical Approach to the Application

of Federal Computer Crime Laws, 16 S

ANTA CLARA HIGH TECH. L.J. 177, 181 (2000).

320. A counter-notification must include the following: (A) a physical or electronic

signature; (B) identification of the material removed and its former location; (C) statement

under penalty of perjury that the user has a good faith belief the material was mistakenly

removed; (D) the user’s name, address, and phone number; and (E) consent to the jurisdiction

of a federal district court. See 17 U.S.C § 512(g)(3) (2012).

321. Seng, supra note 7, at 426.

and cyber-lockers during a six-month period in 2013, while receiving only eight

counter notices challenging its requests.

322

According to Twitter’s transparency

report, Twitter received 2,453 DMCA takedown notices in respect of the Twitter

platform during the month of January 2015, while receiving only two counter

notices in respect of all Twitter services (i.e. including Vine) in return.

323

Different explanations for this discrepancy in takedown notices and counter

notices include users’ disinclination to disclose their identities in a counter notice

and thereby submit themselves to U.S. jurisdiction;

324

the intimidating language

of takedown notices, or simply users’ “ignorance or unawareness of the possible

responses” to a notice of takedown.

325

For whatever reason, the bottom line is

that even when a counter notice procedure is available,

326

it is rarely employed,

rendering it practically useless in enhancing the accountability of algorithmic

B. Accountability Enhancing Strategies

As the previous Subpart demonstrated, when aiming to enhance the

accountability of algorithmic copyright enforcement mechanisms, different

technical, legal, and practical obstacles must be taken under consideration. In the

final Subpart of this Article, we will examine several strategies that may address

some of the barriers discussed above. Generally, these strategies can be pursued by

four distinct players: (1) individual users; (2) privately organized groups or

“watchdogs”; (3) online intermediaries; and (4) regulators.

1. Encouraging Public Participation

When thinking of enhancing the accountability of algorithmic copyright

enforcement, it is natural to put the initial focus on users and ask what they can

do to watch copyright enforcement algorithms. The answer splits into two: first,

innocent target users—namely, users whose online content has been improperly

restricted—should always avail themselves of the counter notice procedure. While

the counter notice procedure is insufficient by itself to generate adequate

accountability,

327

it is still better than having no challenging opportunities

whatsoever. Perhaps, if it were taken more seriously, it could discourage

322. Bruce Boyden, The Failure of the DMCA Notice and Takedown System: A

Twentieth Century Solution to a Twenty-First Century Problem (Dec. 2013)

http://cpip.gmu.edu/wp-content/uploads/2013/08/Bruce-Boyden-The-Failure-of-the-

DMCA-Notice-and-Takedown-System1.pdf [https://perma.cc/33TZ-T3U4].

323. Copyright Notices: Jan. 1-Jun. 30, TWITTER TRANSPARENCY REPORT (2015),

https://transparency.twitter.com/copyright-notices/2015/jan-jun [https://perma.cc/53SZ-

SV7L].

324. Seng, supra note 7, at 430.

325. Id.

326. As we explained, the counter notice procedure is not always available. See supra notes

191-192 and accompanying text.

327. See supra Part III.A.

526 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

repetitive filers of takedown notices from harassing innocent users. Realizing that

target users fight back to force intermediaries to reinstate their non-infringing or

otherwise fair use content, copyright holders and their representatives might be

dissuaded from expending their time and resources on filing bogus copyright

infringement claims.

Second, researchers and the public in general should be encouraged to engage

in reverse engineering research to reveal the essence of algorithmic enforcement

systems.

328

In this regard, the DMCA’s anti-circumvention provisions should be

excused under 17 U.S.C § 1201(g). This exception authorizes “encryption

research”—”activities necessary to identify and analyze flaws and vulnerabilities of

encryption technologies applied to copyrighted works, if these activities are

conducted to advance the state of knowledge in the field of encryption technology

or to assist in the development of encryption products.”

329

Arguably, the act of

reverse engineering the code underlying copyright enforcement algorithms may

be considered as “encryption research” aimed to investigate the code in order to

advance the state of the knowledge in the field of algorithmic copyright

enforcement.

Fortunately, it appears that reverse engineering is increasingly receiving

positive treatment. For instance, when a Princeton graduate was threatened with

a DMCA lawsuit after publishing a report documenting weaknesses in a CD copy-

protection technology developed by SunnComm, public outcry and negative press

attention caused the plaintiff to retreat from its threats.

330

Similarly, Hewlett-

Packard resorted to DMCA threats when researchers published a security flaw in

HP’s Tru64 UNIX operating system, but after widespread press attention, HP

ultimately withdrew the DMCA threat.

331

Furthermore, when the act of reverse engineering does not result in

damaging a computer system, general anti-hacking could also be excused.

332

Researchers and scholars who engage in reverse engineering of copyright

enforcement algorithms should therefore avoid causing any damage to the

underlying code by limiting their hacking efforts to challenging a given software,

without attempting to tamper with its operation. Finally, trade secrecy law

regards reverse engineering as a lawful way to acquire know-how that the

product’s manufacturer may claim as a trade secret.

333

Accordingly, the protection

of trade secrecy law could vanish where the threshold of copyright enforcement

328. Google’s secrecy about its website-ranking algorithm has provoked investigations in

Europe. See Waters, supra note 313, at 22 (“Prompted by three complaints, the European

Commission this year began an informal investigation, the first time that regulators have pried

into the inner workings of the technology that lies at the heart of Google.”).

329. 17 U.S.C. § 1201(g)(1)(A) (2012).

330. DMCA: Ten Years of Unintended Consequences, ELEC. FRONTIER FOUND. (Oct. 27, 2008),

https://www.eff.org/wp/unintended-consequences-ten-years-under-the-dmca

[https://perma.cc/57VJ-FKXZ].

331. Id.

332. See 18 U.S.C. § 1030(a)(5) (2012).

333. See, e.g., Kewanee Oil Co. v. Bicron Corp., 416 U.S. 470, 484-92 (1974).

algorithms is revealed as a result of independent reverse engineering efforts.

334

2. Watchdogs

Another important role in enhancing accountability in algorithmic copyright

enforcement includes that of the private initiatives that are committed to

protecting online free speech and free flow of information. Private initiatives may

boost the accountability of copyright enforcement algorithms by retrieving

information about improper speech restrictions, commenting on it, and

distributing it to the public.

335

A prominent example is the Electronic Frontier

Foundation (EFF),

336

which is a leading nonprofit organization defending civil

liberties in the digital world. Through “impact litigation, policy analysis,

grassroots activism, and technology development,” the EFF promotes “user

privacy, free expression and innovation.”

337

Armed with technological experts,

activists, and attorneys, the EFF increases the awareness of policymakers, the

press, and the public to online free speech violations by posting comprehensive

analysis and educational guides, and engaging in transparency enhancing

initiatives.

338

For instance, its Takedown Hall of Shame

339

project collects and displays

bogus copyright (and trademark) complaints that have threatened all kinds of

creative expression on the Internet. By highlighting interesting stories of

improper DMCA takedown requests, the EFF’s Takedown Hall of Shame can

raise public awareness and generate public pressure on intermediaries. In that

sense, it has an ex ante shaming effect in discouraging unjustified takedown

requests, as well as ex post impacts in fostering public oversight and contest.

Another important project of the EFF relates to promoting quantitative Fair

Use Principles.

340

These principles recommend that content only be blocked if

both the audio and video tracks match the same work and ninety percent or more

of the uploaded content comes from a single work.

341

Moreover, they also

advocate compliance with the DMCA’s N&TD procedures when removing

334. UNIF. TRADE SECRETS ACT § 1 cmt. 2, 14 U.L.A. 437, 438 (1990).

335. Zarsky, supra note 34, at 1535.

336. ELEC. FRONTIER FOUND., https://www.eff.org [https://perma.cc/LVP2-YH46] (last

visited May 8, 2016).

337. About EFF, ELEC. FRONTIER FOUND., https://www.eff.org/about

[https://perma.cc/67UF-VL7S] (last visited May 8, 2016).

338. Id.

339. Takedown Hall of Shame, ELEC. FRONTIER FOUND., https://www.eff.org/takedowns

[https://perma.cc/M4E7-7X8D] (last visited May 8, 2016).

340. Fair Use Principles for User Generated Video Content, ELEC. FRONTIER FOUND.,

http://www.eff.org/issues/ip-and-free-speech/fair-use-principles-usergen

[https://perma.cc/U3AQ-SKTA] (last visited May 8, 2016); See Julian Sanchez, EFF Seeks

Mashup Makers to Fight YouTube Filtering, A

RS TECHNICA (Feb. 3, 2009, 2:44 PM),

http://arstechnica.com/telecom/news/2009/02/eff-seeksmashup-makers-to-fight-youtube-

filtering.ars [https://perma.cc/6TAE-EL54].

341. Fair Use Principles for User Generated Video Content, supra note 340.

528 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

content, including enabling counter notices.

342

Finally, the Fair Use Principles

encourage UGC sites to enable dialogue between content owners and users to

resolve fair use takedowns.

343

Unfortunately, thus far, these well-intended

recommendations have not been widely implemented. As we showed,

intermediaries do not publish how they determine fair use; content is occasionally

blocked ex ante, in direct opposition to the DMCA; and fair use takedowns are

hardly resolved through online dispute resolution systems.

344

Another valuable private initiative that helps holding online copyright

enforcement mechanisms accountable is Lumen (formerly the Chilling

Effects).

345

Founded by the Berkman Center for Internet and Society at Harvard

University, Lumen offers an invaluable clearinghouse for researchers and the

public in general.

Basically, Lumen is “an independent 3rd party research project

studying cease and desist letters concerning online content,” especially “requests

to remove content from online.”

346

Its goals are:

to educate the public, to facilitate research about the different kinds of

complaints and requests for removal—both legitimate and questionable—that are

being sent to Internet publishers and service providers, and to provide as much

transparency as possible about the ‘ecology’ of such notices, in terms of who is

sending them and why [they are being sent], and to what effect.

347

Yet, while Lumen receives a growing number of removal notices—especially

from Google and Twitter, who made public commitments to publish their notices

in Lumen, but also from private recipients—it does not publish all notices of

removal received by all existing intermediaries.

348

Indeed, even Google states that

a copy of each legal notice “may be sent to the Lumen project for publication and

annotation,”

349

allowing that some removal notices may remain unpublished.

Hence, further supporting and encouraging the participation of online

intermediaries in this project is an important step in ensuring researchers have

access to data that can teach them about the practices of online intermediaries

engaging in algorithmic copyright enforcement, which in turn promotes public

literacy and accountability.

342. Id.

343. Id.

344. See supra Part III.B.

345. LUMEN, https://lumendatabase.org [https://perma.cc/K6UD-FRN5] (last visited

May 8, 2016).

346. See About, LUMEN, https://lumendatabase.org/pages/about [https://perma.cc/6H37-

S8J3] (last visited May 8, 2016).

347. Id.

348. Copyright Policy, TWITTER HELP CTR., http://support.twitter.com/articles/15795

[https://perma.cc/BRQ8-YZLB] (last visited May 8, 2016); Legal Removal Requests, G

OOGLE

SUPPORT, https://support.google.com/legal/answer/3110420?hl=en&ref_topic=4556931

[https://perma.cc/RYE9-36JX] (last visited May 8, 2016); Seng, supra note 7, at 379.

349. Legal Removal Requests, supra note 348 (emphasis added).

3. Intermediaries

Intermediaries themselves may also assist in enhancing the accountability of

algorithmic copyright enforcement because they stand in the best position to

increase the transparency of their own, private mechanisms. Arguably, if

intermediaries voluntarily disclosed information about their fair use policies and

their quantitative thresholds, they could foster public scrutiny over their

algorithms’ automatic determinations. If intermediaries published all takedown

notices they receive; they could also assist in generating public literacy by

contributing to the clearinghouse of raw data available for researchers and policy

makers. Google

350

and a few other intermediaries, including Mapbox, Medium,

Reddit, Twitter, Wikimedia and Wordpress,

351

already disclose such information

about DMCA takedown notices they receive in their transparency reports.

Despite the advantages recited above, it is important to stress the pitfalls of

such voluntary transparency reports.

352

First, voluntary sharing of information

may allow the release of partial information that might be biased and misleading.

Second, establishing public oversight exclusively on intermediates’ self-reporting

may entrust online intermediaries with public functions, thus further

strengthening their power in the public sphere. Some extent of regulatory

intervention may therefore be required to ensure accurate reporting, as we

suggest below.

4. Regulators

A fourth player in enhancing the accountability of algorithmic copyright

enforcement is the regulator. Regulation might be necessary to impose some level

of mandatory disclosure on online intermediaries who engage in copyright

enforcement, and to set minimal standards for such disclosure duties. Specifically,

given the un-scalability of the current, top-down DMCA regulation, which fails

to keep pace with the robustness of algorithmic copyright enforcement, we

advocate a collaborative-dynamic regulation that considers the experiences of

350. Google, supra note 207.

351. See Medium’s Transparency Report, MEDIUM (2014),

https://medium.com/transparency-report/mediums-transparency-report-438fe06936ff

[https://perma.cc/7EXT-Y62D] (last visited May 8, 2016); Reddit, Reddit Transparency

Report: Requests for User Information and for Removal of Content (Jan. 29, 2015)

https://www.redditstatic.com/transparency/2014.pdf [https://perma.cc/DUA6-7P3W] (last

visited May 8, 2016); Transparency Report, M

APBOX, https://www.mapbox.com/transparency-

report [https://perma.cc/88BT-C3QK] (last visited May 8, 2016); Transparency Report,

WITTER, https://transparency.twitter.com [https://perma.cc/WAF3-EZ85] (last visited

May 8, 2016); Transparency Report, W

IKIMEDIA FOUND., https://transparency.wikimedia.org

[https://perma.cc/SS6R-KWWR] (last visited May 8, 2016); Transparency Report, W

ORDPRESS,

http://transparency.automattic.com [https://perma.cc/Q3VC-F8EJ] (last visited May 8, 2016).

352. Perel & Elkin-Koren, supra note 284 (explaining that disclosures cannot practically

facilitate adequate accountability in a world of robust algorithmic enforcement and suggesting

using the methodology of black box tinkering for extracting valuable information about the

practices of algorithmic mechanisms of enforcement).

530 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

specific intermediaries in order to continuously update the standards it sets.

353

Indeed, it has been recognized that a system that produces accountability through

bottom-up efforts “can curb discretion, promote consistency, allow for

monitoring, and create incentives for high-quality performance.”

354

Ideally, enhancing intermediaries’ accountability in enforcing copyrights

through regulation should rely on three components previously described in

Bamberger’s work on automated risk-management technologies

355

: (1) increasing

transparency as to the decisions intermediaries make in structuring their

administrative regulator itself, both in terms of technical expertise and in terms of

computing capacity; and (3) encouraging cooperation between the regulated

intermediaries and the regulator in the ongoing development of an effective

regime of algorithmic copyright enforcement.

356

We explain each of these

elements henceforth.

First, regulation may enhance the accountability of algorithmic copyright

enforcement by setting standards of disclosure, such as requiring intermediaries

to disclose the criteria their enforcing algorithms consider when determining

percentage of the copyrighted work must be used to cause content restriction) and

fair use policies.

357

Perhaps it would even be advisable to standardize measures of

the different intermediaries, which may further assist in preventing users from

attempting to reload their removed content on alternative platforms.

358

353. Michael C. Dorf, The Domain of Reflexive Law, 103 COLUM. L. REV. 384 (2003)

(reviewing J

EAN L. COHEN, REGULATING INTIMACY: A NEW LEGAL PARADIGM (2002)). Dorf

essentially proposes a model of reflective law in which insights drawn from experience at the

relatively local level are continually refined and transmitted to the standard-setter, which uses

these insights continually to update the standards all must meet. Id. We propose to encourage

similar cross-fertilization between the regulator who sets the standards of disclosure and the

online intermediaries that must apply them in order to develop standards that on the one hand,

allow for adequate disclosure, and on the other hand, do not overly burden the intermediaries.

354. Rabinovitch-Einy, supra note 35, at 269 (using the term “structural accountability” to

describe “a system that produces accountability through bottom-up efforts”).

355. Bamberger, supra note 15.

356. Id. at 735-38.

357. In the EU, for instance, specific legal rules in the Data Protection Directive provide

users with a supplemental right to receive information about the underlying logic of automated

processing of their personal data. See Council Directive

95/46/EC, art. 12(a), 1995 O.J. (L 281)

42 (EC); see also Douwe Korff, Data Protection Laws in the EU: The Difficulties in Meeting the

Challenges Posed by Global Social and Technical Developments (Centre for Public Reform, Working

Paper No. 2, 2010),

http://ec.europa.eu/justice/policies/privacy/docs/studies/new_privacy_challenges/final_report

_working_paper_2_en.pdf [https://perma.cc/ADG5-XU96].

358. One example relates to the EFF’s Fair Use Principles discussed earlier. Although these

specific recommendations are yet to be implemented in any meaningful way, this sort of

standardization accompanied with regulated disclosure would generally allow target users to

better understand why their content uploads had been taken down or targeted for take down,

while further facilitating public scrutiny over the enforcing algorithms’ determinations. See

Furthermore, regulation can also impose reporting obligations on online

intermediaries, requiring them to publish all takedown notices they receive in

order to allow interested parties to follow and review their conduct. Note,

however, that intermediaries should not be forced to disclose their actual source

code—not only because it is extremely unlikely that lay users would be able to read

and understand it, but especially because the source code may be legitimately

protected under trade secrecy law.

359

As a general rule, when deciding on a

specific mandatory disclosure policy, it is important not to overburden

intermediaries with requirements. Too much transparency may cause

intermediaries to make conservative, inefficient, and unfair decisions, fearing

negative criticism.

360

Indeed, an unbalanced mandate of compulsory disclosure

may result in tilting the algorithm’s default postulation exceedingly towards

tolerating copyright infringement.

Second, increasing the transparency of copyright enforcing algorithms must

be supplemented with increasing the competence of the regulator

361

(for

instance, a trusted advisory committee within the Federal Trade Commission).

The immense volume of takedown notices, combined with the fact that these

notices are processed by machines, calls for the hiring of enough staff with

sufficient technical expertise to propose policy modifications that can be tailored

to algorithms. In other words, a suitable regulator must be able to appreciate

which policies algorithms can achieve, and which they objectively cannot (because

they are too discretionary, for instance). This is especially so if the regulator

wishes to standardize copyright infringement and fair use policies and translate

them into machine-readable code.

Third, regulators should engage in “robust, albeit collaborative, participation

in the ongoing development” of accountable mechanisms of algorithmic copyright

enforcement.

362

Indeed, “[c]ollaboration seeks to facilitate a variety of

arrangements that might capitalize upon the know-how and abilities of

nongovernmental groups in ways that reconfigure those groups and their

relationships, while also providing adequate accountability.”

363

For instance,

regulators can collect data from intermediaries regarding which algorithms work

best, and how to deal with gray-area cases. By collaborating with intermediaries

in developing regulations and acknowledging the practical limitations of

computer codes as enforcement mechanisms, regulators might profoundly change

the enforcing algorithms’ default assumptions and reduce the likelihood of false

positives.

supra notes 337-343 and accompanying text.

359. See supra notes 303-313 and accompanying text.

360. Zarsky, supra note 34, at 1537.

361. Bamberger, supra note 15, at 734.

362. Id. at 735.

363. Jody Freeman, Collaborative Governance in the Administrative State, 45 UCLA L. REV. 1,

31 (1997) (emphasis removed).

532 STANFORD TECHNOLOGY LAW REVIEW [Vol. 19:473

V. CONCLUSION

This Article engaged in a critical discussion about algorithmic governance

and how it intersects with conventional proxies of accountability. Conceiving

algorithms as independent players in automated enforcement regimes that

employ non-transparent learning capacities demands rethinking how to

accumulate the important virtues of accountability. In this Article, we focused on

algorithmic copyright enforcement by online intermediaries to learn about

deficiencies in algorithmic accountability. We identified the various causes that

prevent adequate public scrutiny and explored current and possible strategies for

enhancing accountability in algorithmic copyright enforcement.

The automation of online copyright enforcement—algorithmic applications

of the statutory Notice and Takedown regime, and especially algorithmic systems

that are capable of blocking allegedly infringing content ex ante—raises serious

challenges to the basic notions standing at the heart of accountable enforcement

regimes: transparency, due process and public oversight. Private, profit-

maximizing mega-players’ use of opaque codes to implement discretional legal

doctrines, such as fair use and copyright infringement, can have worrying impacts

on freedom of speech and the rule of law.

However efficient an invisible hand can be in coordinating online content, it

may occasionally be arbitrary and even biased. To secure the free flow of

information and protect online users’ right to create and liberally enjoy the fruits

of their own creations, we must enable adequate checks on algorithmic copyright

enforcement employed by online intermediaries. Especially because online

individuals to understand how mechanisms of algorithmic copyright enforcement

exercise their power, their decision-making criteria, and how their decisions may

be challenged. Otherwise, individuals could be deprived of their right to choose

the content they upload and the online platforms they use.

Current practices of algorithmic copyright enforcement, however, do not

seem to pursue these objectives successfully. As we have demonstrated, affected

individuals lack sufficient knowledge as to the specific thresholds that trigger

algorithmic mechanisms of online copyright enforcement. Oftentimes, no

procedural safeguards are available to allow affected individuals to challenge

algorithmic determinations, and even when challenging opportunities do exist,

they are frequently inefficient.

The public as a whole is also largely incapable of efficiently monitoring

algorithmic copyright enforcement by online intermediaries. Because ex post

content removals are so pervasive, it is impractical to expect the public to

promptly review all removals and generate meaningful public pressure to cause

the reposting of improperly removed content. Moreover, a large amount of

content is being automatically restricted by ex ante filters, such as YouTube’s

Content ID, making it absolutely impossible for general members of public to

know about content restrictions and subsequently contest them.

Enhancing the accountability of algorithmic copyright enforcement faces

several barriers. In addition to the inherent non-transparency of algorithms and

their constantly improving learning capacities, there are several legal barriers—

particularly the DMCA’s anti-circumvention provisions, trade secrecy law and

general anti-hacking laws—that may obstruct the public’s ability to deconstruct

the code underlining algorithmic enforcement mechanisms and hold those

mechanisms accountable. As a result, it is imperative to hone current

accountability enhancing strategies, such as private watchdogs that police and

report improper content restrictions and private disclosure initiatives, and begin

thinking of more structured forms of accountability—either by embedding

internal scrutiny-enhancing mechanisms in the design of enforcement

algorithms, or by subjecting online intermediaries to some degree of external

regulatory inspection—to improve the accountability of algorithmic copyright

enforcement.