Issue-Specific Reporting Best Practices

This section and the corresponding charts in the appendix will look at how well different companies are doing at applying the above best practices to the six different categories of content takedown reporting, while also identifying a few individual best practices that only apply to specific categories of reporting.

For each category of takedown and reporting, we will also address any category-specific best practices that should be considered.

Charts 2a and 2b highlight how companies have been reporting on government and other legal content demands, including instances of content and IP restrictions and blocking. Generally, most companies report on the number of requests they received, and break down requests by country. They also generally report on how they responded to requests.

We don’t have any category-specific best practices to offer here. However, it is worth noting that demands to DNS intermediaries to block content via the Domain Name System (DNS) are becoming increasingly prevalent. However, reporting on this type of content blocking is not yet common. We urge domain name registrars and registries to adopt the practice of transparency reporting going forward.

Charts 3a and 3b highlight how companies have been reporting on copyright requests.

In addition to following the general best practices previously described, the best transparency reports on copyright requests also report on the following aspects:

Counter-Notices and Impact of Counter-Notices for DMCA-Based Requests: A DMCA counter-notice is a notice submitted by a party who has been accused of infringing copyright in response to a DMCA takedown notice as a mechanism to counter the takedown and reinstate the content. There is debate over whether DMCA counter-notices are necessary or useful when it comes to protecting against improper removals, but there is little data to inform this debate. More reporting on the number of counter-notices that companies receive as well as the impact of those counter-notices would provide some much-needed perspective on this issue. Some companies that currently exhibit this best practice are:

Medium: In its latest transparency report, which covers the year 2015, Medium reports on the number of counter-notices received as well as the number of posts reinstated as a result.

Tumblr: In its latest transparency report on copyright and trademark requests, which covers the second half of 2017, Tumblr has a dedicated section for DMCA counter-notices which highlights the number of counter-notices received, the number of counter-notices which were deemed valid, the number of accounts whose content was restored, the number of posts restored and the number of pieces of content restored per month during this reporting period.

Twitter: In its latest transparency report on copyright notices, which covers the second half of 2017, Twitter has a dedicated section for DMCA counter-notices which highlights the number of counter-notices received, the percentage of material restored as a result of the counter-notices, as well as the number of tweets restored and the number of pieces of media restored. This data is broken down on a monthly basis within each reporting period.

Format or Medium of Content Targeted by Copyright Requests: For copyright requests, the category of objectionable content targeted by demands is already evident: it is content that allegedly infringes on a copyright holder’s rights. Therefore, when it comes to specifying the category of objectionable content in transparency reports on copyright requests, we instead encourage companies to specify the format or medium of the targeted content: for example, whether the targeted content is a video, an image, specific text, a specific link, etc. Copyright notices can target a range of content types including text, images, audio and video. This can be seen in practice in the example below:

Tumblr: In its latest transparency report on copyright and trademark requests, which covers the second half of 2017, Tumblr offers numbers of copyright notices received, broken down by the format or medium of the content targeted by the notice, including content types such as images, texts, audio, video, links, asks, quotes and chats.

Data on Who is Submitting Copyright Requests: As already described above, some companies have begun reporting on the top copyright reporters they receive notices from or otherwise highlighting the types of copyright holders who have submitted requests to the company, and we encourage this as a best practice. We also encourage companies, especially American companies, to report on non-DMCA copyright takedown requests they receive. Currently, Medium does this. However, little detail is provided on what the applicable laws and regulations around these requests were, and what the impact of these requests was.

Charts 4a and 4b in the appendix highlight how companies have been reporting on trademark requests. In our survey we did not identify any best practices for reporting that were specific to trademark requests. As a result, Charts 4a and 4b only reflect performance against the general best practices for content takedown reporting mentioned previously. There was little standardization across the companies that were reporting on this issue area.

Currently, some companies, such as Facebook and Etsy, also report on other forms of intellectual property-related requests, such as those related to counterfeit and patent takedowns. Although these reports are not surveyed in this toolkit, we encourage platforms to report on these intellectual property-related takedowns if appropriate for their service.

Charts 5a and 5b highlight how companies have been reporting on network shutdowns and service interruptions. For telecommunications companies, this is usually in the form of a demand from a government to shut down or otherwise disrupt its offering of internet service, often to a specific geographic area. For internet platforms, these events are typically implemented by government agencies directly pressuring telecommunications to block the service for them, and as a result the affected companies themselves do not receive requests to execute such actions. Therefore, internet platform-related data in these charts primarily pertains to internet companies reporting on externally-implemented shutdown and disruption events that have impacted their services. Because network shutdowns and service interruptions are system- and service-wide disruptions, the best practice of reporting on the number of accounts and items specified in demands is not applicable to this content category.

Currently, only a handful of internet and telecommunications companies report on network disruptions and shutdowns, and there is little standardization across these forms of reporting. However, many of the companies that report on network disruptions and shutdowns do so in a detailed manner, and provide a large amount of qualitative information on these events.

In addition to reporting on the general best practices previously described, the best reporting on network shutdowns and service interruptions also includes the following information.

The date or date range of the network shutdown or service interruption: Highlighting the date or date range of each network disruption or shutdown event a company has experienced enables a company to demonstrate the extent of the impact on their business and their users. It also enables companies to highlight any ongoing social, political, or economic events that may have contributed to these disruption or shutdown events.

Facebook: In its latest transparency report, which covers the second half of 2017, Facebook reports that there were 2 nationwide network disruptions in Togo, the first between September 5, 2017 and September 10, 2017 and the second between September 19, 2017 – September 22, 2017, although it offers no other context for these disruptions.

Google: In its latest traffic and disruptions transparency report, Google specifies the date a network disruption event took place. For example, Google reports that on February 20, 2014, Google’s web search experienced a disruption in Syria due to a internet shutdown there, and links to external coverage of that shutdown. In addition, in its reporting Google also specifies whether the disruption event is ongoing. For example, Google reports that there are ongoing disruptions to its web search feature in Congo-Brazzaville and China.

The duration of the network shutdown or service interruption: Highlighting the specific duration of a network disruption or shutdown event enables a company to demonstrate the impact each of these events has had on their services.

Facebook: In its latest transparency report, which covers the second half of 2017, Facebook reports that they experienced 14 internet disruptions in India. These disruptions lasted a total duration of 5 weeks, 2 days, 21 hours. In its reporting, Facebook also provides the duration of each of these individual 14 internet disruptions. For example, Facebook reports that between July 6, 2017 and July 7, 2017, Facebook’s services experienced a disruption in the Kashmir Valley for 13 hours.

Google: In its latest traffic and disruptions report, Google specifies the duration of network disruptions in days, hours and minutes. For example, on September 5, 2017, Google’s web search experienced a disruption in Togo that lasted 4 days, 19 hours and 55 minutes.

The reasons behind network shutdowns and service interruptions: Because the best practice of reporting on categories of objectionable content targeted by demands is not directly applicable during network shutdowns and service interruptions, companies who report on this category should instead report on the reasons behind their shutdowns and interruptions, if known.

Millicom: In its latest Law Enforcement Disclosure Report, which covers the year 2017, Millicom reports that since a specific date in 2014, authorities in El Salvador and Honduras have required telecommunications companies to shut down services or reduce signal capacity in and around prisons in order to prevent criminal gangs from using smuggled cell phones. Their discussion of this shutdown highlights that it is ongoing and because it is legally mandated will likely continue indefinitely.

Chart 6 highlights how companies have been reporting on Right to be Forgotten delisting requests. It reflects company performance against the general best practices for content takedown reporting outline above.

In regard to the best practice of reporting on the categories of objectionable content targeted, because Right to be Forgotten delisting requests only apply to search engines and their results, and only to particular types of data, the categorizations will likely be different than with other types of reporting. The most detailed and specific breakdown of reporting on this best practice we’ve seen so far is in Google’s latest transparency report on search removals under European privacy law, where it provides numbers for a variety of different “categories of content requested for delisting”. The categories of content include insufficient information, personal information, professional information, self-authored, crime and professional wrongdoing. Google also provides numbers broken down by “categories of websites hosting requested for delisting,” including categories like news, directory and social media sites. We encourage other services to follow these best practices.

We also encourage all platforms that receive Right to be Forgotten delisting requests to report on those requests, even if they don’t offer search services and don’t believe believe the current regulations apply to them. If complainants are trying to use the Right to be Forgotten outside of the bounds of the law, that information should be publicized.

Notably, search engines themselves aren’t the only ones that might have Right to be Forgotten-relevant data worth reporting. For example, the Wikimedia Foundation has a dedicated page where it posts “notices of delisted project pages” it has received from search engines that have delisted Wikimedia content as a result of Right to be Forgotten requests. However, because it does not report on requests it has directly received related to the Right to be Forgotten, it does not receive credit for doing so in Chart 6.

Reporting on content takedowns based on a company’s Community Guidelines (also often known as Terms of Service or Community Standards) is one of the newer and emerging forms of transparency reporting. Until recently, the few companies that shared any data at all about their Community Guidelines-based content takedowns did so in an inconsistent manner, either failing to publish regularly or focusing only on takedowns of a particular kind of guidelines-violating content rather than comprehensively looking at the breadth of guidelines-based takedowns. The following companies offer examples of such limited reporting.

Automattic: Beginning in July 2017, Automattic began reporting on the number of reports of terrorist and extremist propaganda in violation of its terms that they received from government Internet Referral Units (IRU’s) around the world. (We treat these as Community Guidelines-based removals because they are not legal demands.) Automattic also reports on the “number of notices where sites/content were removed as a result” and the “percentage of notices where sites/content were removed as a result”. Automattic does not, however, report on the amount of extremist content they identify through other means like user flagging or automated systems and then remove based on their Community Guidelines (which they refer to as “User Guidelines” or “Terms of Service”), and it does not report on any other c aside from terrorist or extremist content.

Etsy: In its latest transparency report, which covers the year 2017, Etsy reports on the total number of accounts closed because of violations of its Seller Policy other than intellectual property violations. According to the report, “violations may range from reselling commercial items to listing prohibited items to repeatedly providing poor customer service service.” The Etsy team uses a combination of “automated systems, community flags, and proactive review” to enforce this policy. In the report, Etsy highlights that they received 394,564 flags regarding “potentially noncompliant listings, including both internal and community generated flags.” Etsy also automatically processed 110,037 of those flags (28%) and the Etsy team manually reviewed 369,694 of those flags (94%), including some automatically flagged items to ensure accuracy. The report breaks down flags by source (community or Etsy) and breaks down accounts reviewed by region (Americas, EMEA and Asia Pacific). However, the report does not offer any granularity in terms of the types of policy violations at issue.

Medium: In its latest transparency report, which covers the year of 2015, Medium reports on takedown requests received “based on violations of Medium’s Terms of Service or Rules, including harassment and privacy” under the category “other takedown requests”. Medium’s reporting includes quantitative data on “total other takedown requests”, “posts removed in whole or part” and “account suspensions”. However, it does not provide a breakdown based on content category, nor does it indicate how much content was flagged by users, by governments, or by automated systems.

Microsoft: In its latest transparency report, which covers the second half of 2017, Microsoft reports on the number of “requests for removal of nonconsensual pornography (“revenge porn”)” it has received. This reporting includes quantitative data on the total number of “requests reported”, “requests accepted” and the “percentage of requests accepted”. Microsoft however does not provide such reporting for other content categories and also does not indicate how much content was flagged by users, by governments, or by automated systems.

Twitter: In its latest reporting on “Government TOS reports”, which covers the second half of 2017, Twitter reports on “worldwide reports” the company received from “official government representatives through our standard customer support intake channels to review content against our terms of service. This section covers government requests to remove content that may violate Twitter’s Terms of Service (TOS) under the following Twitter Rules categories: abusive behavior, copyright, promotion of terrorism, and trademark. It does not include legal demands regardless of whether they resulted in a TOS violation.” For each of these categories, Twitter provides data on the number of “reports received” from governments, the number of “accounts reported” and the “percentage of accounts actioned.” However, Twitter does not report on the amount of content it or other non-government actors have flagged for removal based on the company’s Terms of Service or Community Guidelines. And although Twitter provides reporting on government requests regarding the above-specified content categories, it does not report on the full complement of content categories the platform prohibits.

Especially because the vast majority of takedowns that occur on major platforms are based on those platforms’ Community Guidelines, such limited reporting is not nearly sufficient. The need for in-depth reporting is made even greater by the fact that much of that takedown activity is done unilaterally by the companies—rather than in response to legal process subject to political checks and balances—and because that activity is often automated, raising unique due process, accuracy, and bias concerns.. Therefore, much greater transparency is needed on how companies identify and remove (or otherwise restrict) content based on their own Community Guidelines.

In April 2018, both Facebook and Google’s video service YouTube released transparency reports on the enforcement of its community standards and related content takedown efforts. Thus far these are the most comprehensive transparency reports on Community Guidelines-based content takedowns, and represent the latest best practices in this fast-evolving area.

Facebook: Facebook’s Community Standards Enforcement Report covers three key metrics: how prevalent are Community Standards violations on Facebook (measured by estimating the percentage of views violating content receives), how much content the company takes action on,and how much violating content the company identifies before users do. The company also reported that it is working on instituting a fourth metric — how quickly it takes action on violations. The report covers six content categories: graphic violence, adult nudity and sexual activity, terrorist propaganda (ISIS, al-Qaeda and affiliates), hate speech, spam and fake accounts. These are the biggest categories of guidelines-violating content, but not every category; we hope that future reports are even more comprehensive.

Prior to the release of its Community Standards Enforcement Report, Facebook also announced it was broadening its appeals process, which enables users to appeal the removal of their content. Greater transparency on the number of appeals the company receives and the impact of these appeals is a logical next step in the transparency reporting process.

Google: Google’s Community Guidelines enforcement report focuses on Community Guidelines-based content takedowns on YouTube. Google issues this report on a quarterly basis as opposed to bi-annually like the rest of Google’s content-related transparency reports. The report highlights the number of videos removed based on the type of flagger (human or automated), the percentage of videos that were flagged by automated flagging tools and were removed before and after they were viewed, the percentage of human flaggers by type (government agency, NGO, user and individual trusted flagger), the top 10 countries by human flagging volume, and examples that provide further insight into the flagged video management process. The report also breaks down by category of violation the content removed as a result of human flagging. Unfortunately, it does not offer the same breakdown by category of violation for content originally flagged by automated means, which is especially concerning since automated systems flagged the vast majority of objectionable content.

The release of both Google and Facebook’s Community Guidelines content takedown reports are vital first steps towards increased transparency around content removals online. However, civil society will continue to push companies for greater transparency and accountability in this area. Currently, the best recommendations for reporting on Community Guidelines-based content takedowns are the Santa Clara Principles on Transparency and Accountability Around Online Content. These principles were developed by New America’s Open Technology Institute and a coalition of organizations, advocates and academic experts who support the right to free expression online, with the intention of sparking further in-depth dialogue on Community Guidelines-based transparency reporting in the future. These principles outline minimum levels of transparency and accountability that companies should be providing around enforcement of their Community Guidelines. However, companies should also continue to develop their own unique metrics and forms of reporting, such as Facebook’s prevalence metric, as these also provide valuable insights into how content is removed and how content policies are enforced.

Below are the three Santa Clara Principles—Numbers, Notice, and Appeal—that were intended by the document’s signers “as initial steps that companies engaged in content moderation should take to provide meaningful due process to impacted speakers and better ensure that the enforcement of their content guidelines is fair, unbiased, proportional, and respectful of users’ rights.” As these principles represent some of the latest thinking from civil society and academic experts on the issue of content takedown transparency and accountability, and are mostly applicable to all categories of takedowns and not just Community Guidelines-based takedowns, they are an appropriate place for us to conclude our report. Go forth and be more transparent!