DMCA Classic, DMCA Turbo: Major new empirical research on notice and takedown operations

Good data about Notice and Takedown can be hard to find. Jennifer Urban and Laura Quilter’s seminal 2006 study has long been the gold standard, combining rigorous number-crunching with what must have been incredibly tedious substantive review of the copyright claims in DMCA notices. Now Urban, along with Brianna Schofield and Joe Karaganis, is back with a new report covering more recent data. It’s called “Notice and Takedown in Everyday Practice,” and it’s a treasure trove of information about current DMCA operations.

The team conducted one qualitative study, conducting extensive confidential interviews with key players in the DMCA ecosystem: rightsholders, vendors, and Online Service Providers (OSPs). They also did two quantitative studies, based on close analysis of some 4,000 real DMCA notices held in the Lumen Database. The resulting report is long but fascinating. You should read it. Here are some highlights.

Same Statute, Different Worlds

The report illuminates a major divergence between two worlds of DMCA practice. One is what the authors call “DMCA Classic”: low-volume, high-touch notice and takedown operations, typically involving small OSPs or small rightsholders. In the second world, which I’ll call DMCA Turbo, are the souped-up DMCA processes we know from news and policy debates, often involving powerful commercial actors on both the OSP and rightsholder sides. This world includes highly automated, high volume “DMCA Auto” operations (think “robonotices” and Google Web Search). It also includes voluntary “DMCA Plus” measures, such as proactive content monitoring (think ContentID on YouTube, or Audible Magic on many other platforms).

“Classic” DMCA Operations

Researchers conducted confidential interviews with 29 OSPs, more than half of which fall in the “Classic” category. As the Report describes it, for these OSPs

the traditional DMCA notice and takedown process, involving substantive human review of takedown notices, is still the norm. DMCA Classic OSPs commonly receive dozens or hundreds of notices annually, with minimal increases year-on-year, and tend to be operating outside of the areas where there has been heightened conflict over copyright (music, video, and web search). (29)

Most of these OSPs reported a conservative approach to removals, “opting to take down content even when they are uncertain about the strength of the underlying claim” in order to avoid exposure to liability. (41) OSPs that dedicated more resources or assumed more risk in order to combat abuse typically seemed to do so for cultural reasons rather than truly commercial ones. (43) Like most OSPs in the study, many “Classic” ones reported instances of “deliberate gaming of the DMCA takedown process, including to harass competitors, to resolve personal disputes, to silence a critic, or to threaten the OSP[.]” (40)

These OSPs “described being left aside in policy debates and news accounts skewed by attention to” high profile issues affecting large commercial players.” They expressed specific concern about harm to their own legal protections based on automation and filtering efforts of major commercial players. (31)

The researchers did not have access to meaningful sample sets of DMCA notices from these smaller operators, so the Report offers no quantitative analysis of their practices. Urban et al did, however, have rich sample data from Google’s Image Search service. This service receives far fewer DMCA notices than Web Search –about one three-thousandth as many. It is not generally in the cross-hairs of major commercial or policy disputes, and typically deals with complaints from small businesses or individuals. So, although it is clearly not a “DMCA Classic” provider, its data may be a crude proxy for the “DMCA Classic” operational experience. 

Here are some numbers from the Image Search data set, contrasted with the Web Search data set.

• For the same six month period, Web Search had over 108 million notices and Image search had 33,409.
• Image search notices mostly came from individuals and small businesses. Web search notices overwhelmingly came from commercial rightsholders or their agents, and many appeared to the product of automated notice generation.
• The error or abuse rate in Image Search notices is high: 38% raised substantive issues about the validity of the copyright claim. (That’s excluding one serial notifier whose barrage of notices skewed the entire data set.  With her notices, the error rate rises to 70%.) (98-99) 
• Twenty percent of Image Search notices came from outside the US. (102) 
• Most Image Search notices target personal websites, blogs, or social media sites. (102)

(As a side note: From my own experience, I would also say that the simpler and less professionalized operations described in the Report’s “DMCA Classic” section have much in common with Notice and Takedown for operations for non-copyright claims –defamation or hate speech, for example – in countries where law requires OSPs to remove content in response to those legal claims.)

Turbo DMCA Operations

The Report looks closely at two related developments. One, which it calls “DMCA Auto,” is the use of automation by both rightsholders generating DMCA notices and OSPs processing them. The second, termed “DMCA Plus,” is OSPs’ use of filtering tools to identify and remove infringing content. The legal distinction between these two is important: DMCA Auto uses technology to comply with the statute, DMCA Plus uses it to go beyond the law’s requirements. I’m clustering them for simplicity and to highlight the contrast with the experience of small “Classic” operators, described above.

These processes are fairly well discussed in existing literature, but the Report’s qualitative section adds important detail. For instance, the researchers interviewed vendors that generate “robonotice” DMCA notifications for rightsholders. Some of them discussed spot-checking and other processes they use in an effort to avoid accidentally targeting the wrong content. (34)

Quantitative data about this world of DMCA operations comes from a sample of about 2000 DMCA requests, mostly to Google Web Search. In this sample set,

• 91.8% were sent by outsourced agents, rather than by copyright holders themselves. About half of the agents were vendors and a little under half were trade associations. (84)
• Over 2/3 of requests targeted file-sharing sites. (86)
• “One in twenty-five of the takedown requests (4.2%) were fundamentally flawed because they targeted content that clearly did not match the identified infringed work. This extrapolates to approximately 4.5 million requests suffering from this problem across the entire six-month dataset.” (88)
• 28.4% raised some question about validity, based on facial review of the notices and targeted content. 15.4% failed to comply with DMCA statutory requirements. 7.3% targeted content with potential fair use defenses. (88)
• “In about one out of twenty-five (4.2%) takedown requests, the allegedly infringed work described in the request did not match the allegedly infringing material at all.” For example, rightsholders for the musician Usher targeted a site hosting “The Fall of the House of Usher.” (90-91, acronyms omitted)

The contrast between these operations and “DMCA Classic” are striking. If rightsholders and OSPs sometimes seem to be describing Notice and Takedown experiences from different worlds, the study suggests, perhaps that’s because they are.

Other Interesting Findings

The report is a treasure trove of information. A few more things that jumped out at me:

• Sizes of teams handling notice and takedown within OSPs. (36)
• Abuse of DMCA notices to shoehorn in other legal claims not covered by the statute. (12, 37, 108)
• Rarity of Counternotice. Many OSPs in the survey received none at all. (44) 
• Counternotice abuse. It exists! One notice provider reported being counternoticed seven times: five times for errors, and twice by Russian or Ukrainian torrent sites that presumably felt themselves to be beyond the effective reach of US copyright law. (46)
• Backdoors. Some OSPs “provide senders with access to backend systems that go beyond streamlining … by allowing senders to remove content directly. Some sites allow ‘trusted’ senders to remove content directly from their hosting services without formal notices, identification of the infringed work, user notifications, or review.” (55)
• Competition issues. Interesting discussion of when and why OSPs adopt fingerprinting and filtering systems; available commercial options and their cost; and filtering requirements as barriers to market entry. (57, 64)