January 15, 2018

Questions for the Record
Senate Select Committee on Intelligence
Hearing on Social Media Influence in the 2016 U.S. Elections
November 29, 2017
Questions for the Record for Mr. Sean Edgett, Acting General Counsel, Twitter.
[From Chairman Burr]
1.

What procedures must the Russian government follow to compel the production of
customer-created content or personally identifiable information from your
company?

Twitter publishes global guidelines for law enforcement personnel that explain our
policies and the process for submitting requests for information. See https://help.twitter.com/en/
rules-and-policies/twitter-law-enforcement-support. Since 2012, Twitter has published bi-annual
Transparency Reports, reflecting the number of requests that we have received for user
information and content removal on a per-country basis, including requests from Russia. See
https://transparency.twitter.com/.
Although we have received requests for user information from Russian government
entities, the number of requests per reporting period has been relatively small compared to
requests from other jurisdictions. As the Transparency Reports indicate, Twitter has not
complied with any of those requests. See, e.g., Fig. 1 below (showing Twitter’s Transparency
Report table reflecting 0% compliance for information requests from Russia).
Fig. 1: Summary of Information Requests from Russia

2.

Has the Russian government compelled the production of customer-created content
or personally identifiable information from your company?

3.

If so, has your company complied with such efforts by the Russian government to
compel the production of customer-created content or personally identifiable
information?

4.

Has your company ever refused to comply with efforts by the Russian government
to compel the production of customer-created content or personally identifiable
information? If so, have any of these efforts been successful?

5.

Has your company provided any content created by a U.S. person or personally
identifiable information about a U.S. person to the Russian government?

6.

More specifically, has your company provided to the Russian government the
content of any direct messages sent to or from a U.S. person?

7.

Has your company provided to the Russian government any information that could
be used to determine the location of a U.S. person?
The answers to questions 2-7 are provided in response to question 1.

8.

The persona, GUCClFER2.0 (@GUCCIFER_2), emerged in June 2016 and almost
immediately started to post material purportedly hacked from the Democratic
National Committee (DNC), linking to a seemingly related WordPress site.
● When did Twitter become aware that the @GUCCIFER_2 account was posting
links to a site containing material hacked from the DNC?
● What investigation, if any, did Twitter perform into the @GUCCIFER_2
account?
● If an investigation was conducted, to whom were the results provided?
● Did Twitter provide any information about @GUCCIFER_2 to U.S. law
enforcement or the U.S. Intelligence Community?
● Why is the @GUCCIFER_2 permitted to remain active on Twitter?

Our ability to respond to these questions is limited because we are unable to comment on
whether or not we received requests related to any specific law enforcement investigations.
Twitter maintains a dedicated 24/7 team to respond to law enforcement requests. We
work closely with U.S. law enforcement agencies around the country and promptly respond to
properly scoped and valid legal process. To the extent we are permitted to do so by law, we
make available to the Twitter community and the general public bi-annual, high-level reports
summarizing the types of requests we receive and Twitter’s responses, if any. See
https://transparency.twitter.com.

2

Our Rules prohibit users from posting content that violates our private information
policy, unless those individuals provide express authorization and permission for Twitter users to
do so. An example of private information covered by the policy includes, but is not limited to, a
non-public personal email address. Sharing private information on the platform could pose
serious safety and security risks for the person whose information is shared and is a violation of
the Twitter Rules, which we take very seriously.
When we receive a complaint that private information was shared on our platform,
Twitter will take appropriate action under its policies, which can include temporarily suspending
the account in question pending removal of the private information posted in violation of the
Twitter Rules. Subsequent violations of the Twitter Rules can result in a permanent suspension.
Twitter applies these rules to all accounts, including the account referenced in the question, and
has taken action when that account or others do not comply with the Twitter Rules.
9.

Relatedly, another account, currently also named Guccifer2.0 (@Guccifer2) was one
of the first accounts to welcome @GUCCIFER_2 to Twitter, responding to the
latter’s first Tweet. The @Guccifer2 account has been identified, through Twitter’s
application programming interface (API), to have been created on June 9, 2016.
The moniker “Guccifer2.0” did not publicly exist until @GUCCIFER_2 emerged,
which did not occur until June 16, 2016.
● Has Twitter conducted an investigation as to how the account @Guccifer2 and
the @GUCCIFER_2 account interacted?
● Has the @Guccifer2 account used any other names or handles? If so, when did
the change to Guccifer2.0/@Guccifer2 occur?
● What investigation, if any, did Twitter perform into the @Guccifer2 account?
● If an investigation was conducted, to whom were the results provided?
● Did Twitter provide any information about @GUCCIFER 2 to U.S. law
enforcement or the U.S. Intelligence Community?

Our ability to respond to these questions is limited because we are unable to comment on
whether or not we received requests related to any specific law enforcement investigations.
As noted above, Twitter works closely with U.S. law enforcement to promptly respond to
properly scoped and valid legal process. To the extent we are permitted to do so by law, we
make available to the Twitter community and the general public bi-annual, high-level reports
summarizing the types of requests we receive and Twitter’s responses, if any. See
https://transparency.twitter.com. Any law enforcement requests that Twitter received in
reference to this account would be addressed consistent with Twitter’s Law Enforcement
Guidelines.
Twitter users are permitted to change their usernames (also referred to as their handles)
and display names over time while maintaining an account on our platform. Once a Twitter user
changes a username, the prior username and date of the change is no longer visible on Twitter or

3

via our application programming interface (“API”). More information about how Twitter
account holders can update this information is available through Twitter’s Help Center. See
https://help.twitter.com/en/managing-your-account/change-twitter-handle.

4

[From Vice Chairman Warner]
10.

Outside researchers suggest that as much as 15% of the accounts — or 48 million
accounts — on Twitter are automated bots, while Twitter has consistently
contended the proportion of automated accounts is much lower.
● If the researchers are inaccurate with respect to the number of fake accounts on
your platform, as you suggest, what steps will you commit to take to assist these
researchers with providing the real information?

We regularly receive and welcome input from researchers and Twitter users about ways
in which we can optimize our detection and enforcement methods with respect to false or spam
accounts. We are committed to continuing to work on refining those tools and to update the
public and the Twitter community periodically about our estimates and analysis. And Twitter is
unique in the transparency it provides about public Tweet data via our application programming
interface (“API”). Through our API, we provide developers, researchers, and other third parties
access to public Twitter content. This service is a hallmark of our commitment to transparency,
collaboration, and innovation.
Based on a review of a representative sample of accounts, we estimate that false or spam
accounts represent less than 5% of our MAUs. Our estimates are lower than those reported by
outside researchers because although we can make a significant amount of information public,
researchers do not have access to critical internal information necessary to make an accurate
determination of the scale of spam, false accounts or automated bots on Twitter. As a result,
reports from third-party researchers often overestimate the true volume of such accounts on our
platform—sometimes by large orders of magnitude.
While our detection tools for false or spam accounts rely on a number of inputs and
variables and do not operate with 100% precision, they are informed by data not available
outside of Twitter. Our internal researchers have access to and can analyze a number of different
signals including, among other things, email addresses, phone numbers, login history, and other
non-public account and activity characteristics that enable us to conduct a more thorough review
and reach more accurate conclusions as to whether the account in question is fake or spam. We
keep such information confidential and do not make it available to researchers in order to protect
the privacy of our users.
Because third-party researchers do not have access to internal signals that Twitter can
access, their bot and spam detection methodologies must be based on public information and
often rely on human judgment, rather than on internal signals available to us. One common
model for determining whether an account is fake or automated is the “Botometer model,” which
compares publicly available account features, such as Tweet count, follower count, and use of
language, to the characteristics exhibited by purportedly “known” bots. The initial evaluation of
whether an account is or is not a bot, however, relies on an individual assessment and is,
therefore, inherently imprecise.
The studies that rely on information from the Twitter API to identify automated accounts
similarly overestimate both the number and impact of these accounts because our internal

5

detection tools and filtering techniques are not available to third parties. Those tools enable us to
remove from the platform malicious automated accounts (and content generated by such
accounts), but the accounts may nevertheless appear in the data stream that researchers access
through our API, thus inaccurately reflecting the traffic on Twitter.
A study conducted by the University of Southern California and Indiana University
estimated that as much as 15% of Twitter accounts are automated, spam accounts. That estimate,
however, was based on a prediction of whether a user may or may not be an automated account
and was derived from human judgments about an account’s attributes. The authors of the study
acknowledge that detecting automated accounts “is a hard task. Many criteria are used in
determining whether an account is controlled by a human or a bot, and even a trained eye gets it
wrong sometimes.” See https://botometer.iuni.iu.edu/#!/faq#bot-threshold.
In addition to our ongoing steps to promote transparency about our platform through the
API, as we have announced, we are also committed to donating the $1.9 million we projected to
have earned from RT advertising to support external research into the use of Twitter in civic
engagement and elections, including the use of malicious automation and misinformation.
11.

What are your policies regarding the distribution of hacked and stolen emails from
your platform?

The use of Twitter to distribute hacked or stolen emails or documents may implicate a
number of the Twitter Rules, including our prohibitions on posting private information,
infringement of intellectual property rights, and/or unlawful use. For example, users are
prohibited from posting other people’s private information, unless those individuals provide
express authorization and permission for Twitter users to do so. This policy protects information
such as non-public personal email addresses and mobile phone numbers.
In addition, under the unlawful use provision of the Twitter Rules, users are prohibited
from using Twitter’s “service for any unlawful purpose or in furtherance of illegal activities” and
“[i]nternational users agree to comply with all local laws regarding online conduct and
acceptable content.” See Twitter User Agreement—Twitter Rules, https://twitter.com/en/tos.
Twitter has processes in place for law enforcement and private parties to report Twitter Rules
violations to us and to submit legal requests and court orders for removal of illegal content.
When Twitter detects content posted in violation of our rules and policies, we may take a range
of enforcement actions, including requiring the deletion of specific content, withholding an
account or specific Tweets where they are unlawful, or suspending the account.
12.

We’ve seen that bad actors are working across the various platforms to spread their
disinformation. It has proven true in the past several months that identifying fake
accounts on one platform can facilitate identification on other platforms.
● Do you share data with other firms?
● What concrete steps are you taking to share information to improve detection?

Twitter has partnered with other platforms to make progress against common threats. In
June 2017, for example, we launched the Global Internet Forum to Counter Terrorism (the
6

“GIFCT”), a partnership among Twitter, YouTube, Facebook, and Microsoft. The GIFCT
facilitates, among other things, information sharing; technical cooperation; and research
collaboration, including with academic institutions.
The GIFCT has created a shared industry database of “hashes”—unique digital
“fingerprints”—for violent terrorist imagery or terrorist recruitment videos or images that have
been removed from our individual services. The database allows a company that discovers
terrorist content on one of their sites to create a digital fingerprint and share it with the other
companies in the forum, who can then use those hashes to identify such content on their services
or platforms, review against their respective policies and individual rules, and remove matching
content as appropriate, or even block extremist content before it is posted in the first place. The
database now contains more than 40,000 hashes. Instagram, Justpaste.it, LinkedIn, Oath, and
Snap have also joined this initiative, and we are working to add several additional companies in
2018. Twitter also participates in the Technology Coalition, which shares images to counter
child abuse.
Twitter understands that, to succeed, we must combine information, knowledge, and
effort with industry partners, civil society, academic institutions, and government. We do not
compete against other companies on our ability to detect and label malicious content on our
platform; instead, we recognize that we will all be stronger if we view this as a shared threat.
We are committed to a continued collaborative approach and believe it will prove successful
going forward.
13.

Press investigations have found that for as little as $100 it is possible to buy a “bot
army” via unscrupulous underground online actors, enabling the coordinated
propagation of tweets, including of false information, to create trending topics. The
Daily Beast publication investigated and was able to buy 1000 Twitter accounts for
just $45.1
● Is it a violation of your Terms of Service to buy and sell bots?
● What safeguards do you have in place to prevent this type of activity?

Twitter strictly prohibits the purchasing and selling of account interactions on our
platform. We advise our users that, by purchasing followers, Retweets, and likes, they are often
purchasing bots, fake, or hacked accounts. Accounts found to have purchased, sold, or promoted
the selling of followers, Retweets, and likes are in violation of the Twitter Rules and may be
subject to suspension. See https://help.twitter.com/en/rules-and-policies/twitter-rules.
The Twitter Rules prohibit using automation tools for the purpose of generating spam—
unwanted content consisting of multiple postings either from the same account or from multiple
coordinated accounts. While “spam” is frequently viewed as having a commercial element,
since it is a typical vector for spreading advertising, Twitter’s Rules take an expansive view of
spam because it negatively impacts the user experience. Examples of spam violations on Twitter
include automatically Retweeting content to reach as many users as possible, automatically
1

https://www.thedailybeast.com/i-bought-a-russian-bot-army-for-under-dollarl100

7

Tweeting about topics on Twitter in an attempt to manipulate trends, generating multiple Tweets
with hashtags unrelated to the topics of those hashtags, repeatedly following and unfollowing
accounts to tempt other users to follow reciprocally, Tweeting duplicate replies and mentions,
and generating large volumes of unsolicited mentions.
Our systems are built to detect malicious automated and spam accounts across their
lifecycles, including detection at the account creation and login phase and detection based on
unusual activity (e.g., patterns of Tweets, likes, and follows). Our ability to detect such activity
on our platform is bolstered by internal, manual reviews conducted by Twitter employees. Those
efforts are further supplemented by user reports, which we rely on not only to address the content
at issue but also to calibrate our detection tools to identify similar content as spam.
Once our systems detect an account as generating spam, we can take action against that
account at either the account level or the Tweet level. Depending on the mode of detection, we
have varying levels of confidence about our determination that an account is violating our rules.
We have a range of options for enforcement; generally, the higher our confidence that an account
is violating our rules, the stricter our enforcement action will be, with immediate suspension as
the harshest penalty. If we are not sufficiently confident to suspend an account on the basis of a
given detection technique, we may challenge the account to verify a phone number or to
otherwise prove human operation, or we may flag the account for review by Twitter personnel.
Until the user completes the challenge, or until the review by our teams has been completed, the
account is temporarily suspended; the user cannot produce new content (or perform actions like
Retweets or likes), and the account’s contents are hidden from other Twitter users.
We also have the capability to detect suspicious activity at the Tweet level and, if certain
criteria are met, to internally tag that Tweet as spam or otherwise suspicious. Tweets that have
been assigned those designations are hidden from searches, do not count toward generating
trends, and generally will not appear in feeds unless a user follows that account. Typically, users
whose Tweets are designated as spam are also put through the challenges described above and
are suspended if they cannot pass.
For safety-related Terms of Service (“TOS”) violations, we have a number of
enforcement options. For example, we can stop the spread of malicious content by categorizing
a Tweet as “restricted pending deletion,” which requires a user to delete the Tweet before the
user is permitted to continue using the account and engaging with the platform. So long as the
Tweet is restricted—and until the user deletes the Tweet—the Tweet remains inaccessible to and
hidden from all Twitter users. The user is blocked from Tweeting further unless and until he or
she deletes the restricted Tweet. This mechanism is a common enforcement approach to
addressing less severe content violations of our TOS outside the spam context; it also promotes
education among our users. More serious violations, such as posting child sexual exploitation or
promoting terrorism, result in immediate suspension and may prompt interaction with law
enforcement.

8

14.

If Twitter feels it is important to allow for bots on its platform, then why not require
a disclosure that a bot account is non-human? Why not ensure that your real users
know when they are interacting with an automated bot?

Because automated content can originate with any type of account, such disclosure would
necessarily be both over-inclusive and under-inclusive. For example, certain automation tools
that do not violate our rules enable users to schedule their Tweets to post automatically at a
particular time or in response to specific activity. Since they may be utilized by individual users,
applying a categorical designation to accounts that employ those tools would not accurately
characterize such accounts.
It is important to note, moreover, that not all automation is malicious. Automation is
essential for certain informational content, particularly when time is of the essence, including for
law enforcement or public safety notifications. Examples include Amber Alerts, earthquake and
other storm warnings, and notices to “shelter in place” during active emergency situations.
Automation is also used to provide customer service for a range of companies. For example, as
of April 2017, users are able to Tweet @TwitterSupport to request assistance from Twitter. If a
user reports a forgotten password or has a question about our rules, the initial triage of those
messages is performed by our automated system—a Twitter-developed program to assist users in
troubleshooting account issues.
To maintain the integrity of our platform and to ensure a positive user experience,
Twitter’s approach to addressing the spread of malicious automation is to focus on problematic
behavior and abuse, and on accounts that engage in such behavior. To that end, we monitor
various behavioral signals related to the frequency and timing of Tweets, Retweets, likes, and
other such activity, as well as to similarity in behavioral patterns across accounts, in order to
identify accounts which are likely to be automated and acting in a malicious automated and
coordinated fashion. When we identify such accounts or abusive activity, we take action to
prevent them from interfering with the integrity of the platform and detracting from the positive
experience of our users.

9

[From Senator Collins]
15.

What provision in your Terms of Service ensures that political advertisements
targeted toward the United States are purchased by an American citizen?

As part of our ads transparency and electioneering ads efforts announced in October
2017, Twitter will require advertisers to go through an onboarding process, which will obligate
them to provide information about how they are funding their media buys. For advertisers who
do not self-identify but who run electioneering ads, we will use a combination of machinelearning models and human manual review to detect and halt these advertisers until they have
correctly onboarded with us as an electioneering advertiser.
While it is possible that foreign governments may attempt to purchase ads through
consultants or management companies, Twitter’s upcoming Transparency Center is intended to
provide identifying information about any such companies and their other advertising activities
on Twitter. That information will better enable users and outside parties to conduct their own
research or evaluation regarding particular ads.
If Twitter concluded that any advertiser was running an unlawful ad on Twitter, Twitter
would restrict the user from such further action, remove the ad, or both as appropriate.
16.

Do your Terms of Service prohibit users from influencing elections in other
countries?

Twitter is a global platform with more than 330 million monthly active users around the
world. Facilitating organic, robust debate and commentary about important events around the
world is a core part of Twitter’s mission. So long as our users do not violate the Twitter Rules or
the Twitter Terms of Service, we do not place restrictions on the organic content that they choose
to Tweet about and share with their followers, including when that content contains commentary
about elections in other countries.
By contrast, unlawful interference with elections is prohibited. As noted in the unlawful
use provision of the Twitter Rules, users are prohibited from using Twitter’s “service for any
unlawful purpose or in furtherance of illegal activities” and “[i]nternational users agree to
comply with all local laws regarding online conduct and acceptable content.” Twitter User
Agreement—Twitter Rules, available at https://twitter.com/en/tos. For example, during the
period leading up to the 2016 election, Twitter took action on Tweets that suggested to users that
they can vote by text message. Because the Tweets in question appeared to mislead users into
believing that they could vote online or vote by text, Twitter viewed them as an unlawful
interference with the voting process and took action against those accounts and Tweets.
17.

If a foreign national working on behalf of a foreign intelligence service was an
authentic user in real name on your platform, could he post divisive, but non-violent
content related to a U.S. election without violating your Terms of Service? Would he
be able to purchase political advertising?

Twitter is committed to providing a service that fosters and facilitates free and open
democratic debate and that promotes positive change in the world. Twitter has a history of
10

facilitating civic engagement and political freedom, and we intend for Twitter to remain a vital
avenue for free expression here and abroad. As a global platform, we believe that our users
benefit from the exchange of ideas with other users around the world. Accordingly, unless an
account or a user violates the Twitter Rules or the Twitter Terms of Service—including by
engaging in abusive behavior, promoting violence, harassing individuals, or using the platform
for unlawful purposes—we do not restrict the type of content they choose to share with their
followers based on their nationality or citizenship.
Advertisements generally receive a different type of review than organic Tweets. This is
because organic Tweets are generally shown to people who choose to follow the user that sends
the Tweet, while ads—Promoted Tweets—can be served to a broader audience, including users
who have not chosen to follow the user or account that generated the ad. And Twitter places
greater limitations on the type of content that can be promoted with Twitter ads compared to
organic content that our users generate. In addition, foreign nationals may be subject to legal
restrictions governing campaign contributions and electioneering. If Twitter concluded that any
advertiser, including a foreign national, was running an unlawful ad on Twitter, Twitter would
restrict the user from such further action, remove the ad, or both as appropriate.

11

[From Senator Feinstein]
18.

Twitter has conceded that the number of people exposed to content from foreign
groups online is far more pronounced through organic traffic and fake accounts
than it is through paid advertising. Troublingly, it does not appear there is a proven
method for combatting the spread of fake accounts created to sow division in
society. Although the Committee has heard testimony indicating that Twitter has a
number of ways to detect “bot-like” activity, as recently as August 2017, divisive
foreign unpaid content designed to polarize and anger the American people could be
found on Twitter.
● What specific actions is Twitter taking to combat this type of divisive unpaid
activity on an on-going basis?

We are committed to addressing the spread of misinformation on our platform—and to
prevent future attempts to interfere with U.S. elections—but we recognize that spam and
malicious automation are not limited to political content and can undermine the positive user
experience we seek to offer regardless of content. Twitter’s approach to addressing the spread of
malicious automation and inauthentic accounts on our platform is to focus on identifying
problematic behavior and abuse, not primarily on the content that such accounts attempt to
disseminate. This is not to say that the content is not important—or that content has no place in
our analysis—but we recognize that those who are seeking to influence a wide audience must
find ways to amplify their messages across Twitter. As with spam and terrorist content, these
behaviors frequently provide more precise signals than focusing on content alone.
Accordingly, we monitor various behavioral signals related to the frequency and timing
of Tweets, Retweets, likes, and other such activity, as well as to similarity in behavioral patterns
across accounts, in order to identify accounts that are likely to be maliciously automated or
acting in an automated and coordinated fashion in ways that are unwelcome to our users. We
monitor and review unsolicited targeting of accounts, including accounts that mention or follow
other accounts with which they have had no prior engagement. For example, if an account
follows 1,000 users within the period of one hour, or mentions 1,000 accounts within a short
period of time, our systems are capable of detecting that activity as aberrant and as potentially
originating from suspicious accounts.
Twitter is continuing its effort to detect and prevent malicious automation by leveraging
our technological capabilities and investing in initiatives aimed at understanding and addressing
behavioral patterns associated with such accounts. For example, in early 2017, we launched the
Information Quality initiative, an effort aimed at enhancing the strategies we use to detect and
stop bad automation, improve machine learning to spot spam, and increase the precision of our
tools designed to prevent such content from contaminating our platform.
Since the 2016 election, we have made significant improvements to reduce external
attempts to manipulate content visibility. These improvements were driven by investments in
methods to detect malicious automation through abuse of our API, limit the ability of malicious
actors to create new accounts in bulk, detect coordinated malicious activity across clusters of
accounts, and better enforce policies against abusive third-party applications.

12

In addition, we have developed new techniques for identifying patterns of activity
inconsistent with legitimate use of our platform (such as near-instantaneous replies to Tweets,
nonrandom Tweet timing, and coordinated engagement), and we are currently implementing
these detections across our platform. We have improved our phone verification process and
introduced new challenges, including reCAPTCHA (utilizing an advanced risk-analysis engine
developed by Google), to give us additional tools to validate that a human is in control of an
account. We have enhanced our capabilities to link together accounts that were formed by the
same person or that are working in concert. And we are improving how we detect when
accounts may have been hacked or compromised.
With our improved capabilities, we are now detecting and blocking approximately
450,000 suspicious logins each day that we believe to be generated through automation. In
September 2017, our systems identified and challenged an average of four million suspicious
accounts per week, which represents more than double our rate of detection at this time last year.
Over three million of those accounts were challenged upon signup, before their content or
engagements could impact other users. Between June and September 2017, we also suspended
more than 117,000 malicious applications for API abuse. Those applications were collectively
responsible for more than 1.5 billion Tweets in 2017.
We plan to continue building upon our 2017 improvements, including through
collaboration with our peers and investments in machine-learning capabilities that help us detect
and mitigate the effect on users of fake, coordinated, and malicious automated account activity.
We have also observed the expansion of malicious activity on our platform from
automated accounts to human-coordinated activity, which poses additional challenges to making
our platform safe. We are determined to meet those challenges and have been successful in
addressing such abusive behavior in other contexts. We are committed to leveraging our
technological capabilities in order to do so again by carefully refining and building tools that
respond to signals in the account behavior. For example, as of September 2017, 95% of account
suspensions for promotion of terrorist activity were accomplished using our existing proprietary
detection tools—up from 74% in 2016. These tools focus on indicia of violating activity beyond
the content of the Tweet. We are confident that the combination of our dedicated teams, our
detection tools, and other technological advancements at our disposal will prove essential in
addressing malicious human-coordinated activity as well.
19.

One of the more troubling findings from this investigation is the number of targeted
voter disengagement efforts promoted through social media.
● Can you say with certainty that foreign actors did not use U.S. voter registration
data to target individuals through both paid and unpaid activity?

As we explained in connection with the November 1, 2017, hearing before the
Committee, as a result of our retrospective review, we have identified ways in which Russian
actors engaged with our platform in the period leading up to the 2016 election. Although we
have no visibility into what information those actors had at their disposal, users posting organic
(unpaid) content cannot target other users based on any criteria, and voter registration is not an

13

available targeting criterion for ads and promoted Tweets on our platform (i.e., paid Twitter
services).

14

[From Senator Cotton]
20.

Do Twitter’s Terms of Service prohibit collaboration with Russian intelligence
services to influence an election?

Unlawful interference with elections is strictly prohibited and users who engage in
malicious activity designed to exert artificial influence on an election are likely to violate any
number of Twitter’s rules and policies. These may include our rules against spam, malicious
automation, abuse or harassment, posting private information, or advertising policies. In
addition, any illegal collaboration with Russian intelligence services to influence an election
would implicate Twitter’s prohibition against use of the “service for any unlawful purpose or in
furtherance of illegal activities.” Twitter’s Rules specify that “international users agree to
comply with all local laws regarding online conduct and acceptable content.” Twitter User
Agreement—Twitter Rules, available at https://twitter.com/en/tos.
21.

Provided an individual or entity does not violate Twitter’s Terms of Service, will
they be allowed to use your platform to work with hostile, foreign intelligence
services to potentially influence the 2018 and 2020 U.S. elections?
The answer to question 21 has been provided in response to question 20.

22.

What is Twitter’s justification for allowing entities and individuals such as
WikiLeaks, Julian Assange, and Edward Snowden to maintain Twitter accounts?

One of Twitter’s core purposes is to help advance the global, public conversation on
various topics of interest to our users. As we stated in connection with the Committee hearing,
Twitter’s values include defending and respecting the user’s voice—a two-part commitment to
freedom of expression and privacy. Twitter has a history of facilitating civic engagement and
political freedom, and we intend for Twitter to remain a vital avenue for free expression here and
abroad.
Consistent with our values and commitment to fostering an open exchange of ideas,
unless the activity or posted content violates our Terms of Service or Twitter Rules, we do not
bar controversial figures from our platform or prohibit accounts from posting controversial
content. We believe that barring controversial figures from our platform or removing their
controversial Tweets would hide important information that our users should be able to see and
debate and would detract from the public dialogue that our platform is intended to promote.
We take action against accounts for Terms of Service and Twitter Rules violations, and
we apply those rules consistently to all accounts. So long as those accounts remain in
compliance with our policies, we do not take action against their Tweets or suspend them from
the platform.
23.

Describe how Twitter came to acquire DataMinr.

Twitter did not acquire Dataminr. As has been publicly reported, Twitter owns a 5%
stake in Dataminr and maintains an ongoing commercial relationship with that company.

15

24.

Is there any portion of the DataMinr platform that shows non-public facing tweets
or messages?

Twitter provides Dataminr only with Tweets that users choose to make public. We do
not share non-public facing Tweets or messages with Dataminr.
25.

What is Twitter’s justification for labeling U.S. Intelligence Community access to
information sold by DataMinr as “surveillance”?

We prohibit the use of Twitter data for surveillance purposes by any entity whose
primary function is surveillance or the collection of intelligence. This is a longstanding Twitter
policy. Twitter is proud to work with a range of government and law enforcement agencies,
including the FBI, and to provide Twitter data for public safety, news alerting purposes, or in
response to valid legal process.
26.

Which Russian and Chinese entities currently have access to information sold by
DataMinr?

Dataminr and its business relationships with its end users are managed independently of
Twitter. Dataminr has notified Twitter that it does not have existing customers in China or
Russia.

16

[From Senator Heinrich]
27.

On November 8, CNN reported that a Twitter account — MAGA Mike King —
tweeted more than a dozen times on Election Day, November 7, 2017, a graphic
purportedly instructing Virginians how to vote by text. This graphic included the
logos of the Democratic Party and its gubernatorial candidate, Ralph Northam.
According to CNN, the account remained active for almost three hours out of the 13
hours that the polls were open in Virginia.
● Did Twitter devote additional resources to monitoring misinformation on
Election Day, November 7, 2017?
● Why did Twitter take so long to suspend the account?

When we become aware of malicious uses of our platform, we take immediate action to
enforce our rules. We also recognize the importance of events such as elections, and we commit
additional resources to respond to law enforcement inquiries concerning abuse, spam, and other
malicious uses of our platform during campaign and election periods. We regularly reexamine
staffing and resources and adjust as needed, and we are launching a process to ensure that during
election periods we are positioned to respond in the most effective and efficient way.
Due to the high volume of content posted to Twitter and the real-time nature of the
platform, Twitter is not able to proactively monitor all Tweets posted in relation to those events.
Our main line of defense against malicious activity on our platform remains our technological
tools and systems, which are capable of detecting malicious automated and spam accounts across
their lifecycles, including at the account creation and login phase, as well as when those accounts
exhibit unusual activity (e.g., patterns of Tweets, likes, and follows). Our ability to detect such
activity on our platform is bolstered by internal, manual reviews conducted by Twitter
employees, as well as by information we obtain through third-party security vendors. Those
efforts are further supplemented by user reports, on which we rely not only to address the content
at issue but also to calibrate our detection tools to identify similar content as spam.
Twitter is continuing its effort to detect and prevent malicious automation by leveraging
our technological capabilities and investing in initiatives aimed at understanding and addressing
behavioral patterns associated with such accounts. For example, in early 2017, we launched the
Information Quality initiative, an effort aimed at enhancing the strategies we use to detect and
stop bad automation, improve machine learning to spot spam, and increase the precision of our
tools designed to prevent such content from contaminating our platform. Our efforts have
already allowed us to respond more quickly to automation, spam, and malicious activity during
high-profile events.
When Twitter received reports about the referenced account, we permanently suspended
the user. Our action against this account is consistent with the approach we took against illegal
voter suppression Tweets during the 2016 election. Here, however, and well before our manual
review of the account’s activity resulted in its permanent suspension, Twitter’s automated spam
detection systems identified malicious behavior originating from this account and took action to
hide that user’s Tweets from appearing in searches and counting toward trends. Those

17

automated systems, which we continue to invest in as part of our Information Quality initiative,
help us address emerging malicious behavior even before a human reviewer can assess the
content.
28.

What percentage of Twitter content reviews are conducted by an actual human
being rather than via automated review?

Twitter dedicates significant resources to addressing malicious automation, bots, and
other coordinated activities. We believe we have the right resources and strategies in place. We
dedicated nearly the entire engineering, product, and design teams to look at these issues at the
beginning of 2017, and we regularly reexamine staffing and resources and adjust as needed.
Critical to the continued success of our efforts is our ability to leverage our technological
advancements and improvements to tackle this problem; given the scale, this is not something
that we or anyone else can review and address by hand.
But we do not depend on automated systems alone to address malicious automation, bots,
and other abusive content or activity on our platform. Our ability to detect such activity on our
platform is bolstered by internal, manual reviews conducted by Twitter employees. Those
efforts are further supplemented by user reports, which we rely on not only to address the
activity at issue but also to calibrate our detection tools to identify similar content as spam and to
enforce the Twitter Terms of Service and Twitter Rules.
While automation provides significant opportunities to scale enforcement activity, it
frequently performs better when it supplements, but does not supplant, human review of content.
For example, our automated systems flag accounts that engage in suspicious activity for further
manual review by Twitter personnel. Those systems also assist our staff in prioritizing manual
reviews of user reports. For accurate and consistent enforcement of the Twitter Rules, many of
the most sensitive and nuanced types of Twitter Rules violations are typically conducted by
Twitter employees. Sensitive reports of illegal activities, such as voter suppression, are subject
to review by a Twitter employee and in coordination with the legal department.
We also recognize that computer algorithms alone are not sufficient to address the
problem. Context matters, and Twitter’s commitment to protecting users’ opinions and
expression require that tooling alone does not try and solve these complex issues. Accordingly,
those tools are complemented by manual review teams, collaboration and information sharing
with industry peers and participants, reliance on data and intelligence from third-party security
vendors, and partnerships with other companies and civil society.
Twitter understands that, to succeed, we must combine resources, information,
knowledge, and effort with industry partners, civil society, and government. We do not compete
against other companies on our ability to detect and label malicious content on our platform;
instead, we recognize that we will all be stronger if we view this as a shared threat. We are
committed to a continued collaborative approach and believe it will prove successful going
forward.

18

29.

Are Twitter’s content review processes the same now as they were during the 2016
election? If not, how have they changed?

Twitter’s approach to addressing the spread of malicious automation and inauthentic
accounts on our platform is to focus on problematic behavior and abuse, not primarily on the
content that such accounts attempt to disseminate. We are committed to addressing the spread of
misinformation on our platform—and to prevent future attempts to interfere with U.S.
elections—but we recognize that spam and malicious automation are not limited to political
content and can undermine the positive user experience we seek to offer irrespective of the
content.
Accordingly, we monitor various behavioral signals related to the frequency and timing
of Tweets, Retweets, likes, and other such activity, as well as to similarity in behavioral patterns
across accounts, in order to identify accounts that are likely to be automated or acting in an
automated and coordinated fashion in ways that are unwelcome to our users. We monitor and
review unsolicited targeting of accounts, including accounts that mention or follow other
accounts with which they have had no prior engagement. For example, if an account follows
1,000 users within the period of one hour, or mentions 1,000 accounts within a short period of
time, our systems are capable of detecting that activity as aberrant and as potentially originating
from suspicious accounts.
Twitter is continuing its effort to detect and prevent malicious automation by leveraging
our technological capabilities and investing in initiatives aimed at understanding and addressing
behavioral patterns associated with such accounts. For example, in early 2017, we launched the
Information Quality initiative, an effort aimed at enhancing the strategies we use to detect and
deny bad automation, improve machine learning to spot spam, and increase the precision of our
tools designed to prevent such content from contaminating our platform.
Since the 2016 election, we have made significant improvements to reduce external
attempts to manipulate content visibility. These improvements were driven by investments in
methods to detect malicious automation through abuse of our API, limit the ability of malicious
actors to create new accounts in bulk, detect coordinated malicious activity across clusters of
accounts, and better enforce policies against abusive third-party applications.
In addition, we have developed new techniques for identifying patterns of activity
inconsistent with legitimate use of our platform (such as near-instantaneous replies to Tweets,
nonrandom Tweet timing, and coordinated engagement), and we are currently implementing
these detections across our platform. We have improved our phone verification process and
introduced new challenges, including reCAPTCHA (utilizing an advanced risk-analysis engine
developed by Google), to give us additional tools to validate that a human is in control of an
account. We have enhanced our capabilities to link together accounts that were formed by the
same person or that are working in concert. And we are improving how we detect when
accounts may have been hacked or compromised.
With our improved capabilities, we are now detecting and blocking approximately
450,000 suspicious logins each day that we believe to be generated through automation. In
September 2017, our systems identified and challenged an average of four million suspicious

19

accounts per week, which represents more than double our rate of detection at this time last year.
Over three million of those accounts were challenged upon signup, before their content or
engagements could impact other users. Between June and September 2017, we have also
suspended more than 117,000 malicious applications for API abuse. Those applications were
collectively responsible for more than 1.5 billion Tweets in 2017.
We plan to continue building upon our 2017 improvements, including through
collaboration with our peers and investments in machine-learning capabilities that help us detect
and mitigate the effect on users of fake, coordinated, and malicious automated account activity.
We have also observed the expansion of malicious activity on our platform from
automated accounts to human-coordinated activity, which poses additional challenges to making
our platform safe. We are determined to meet those challenges and have been successful in
addressing such abusive behavior in other contexts. We are committed to leveraging our
technological capabilities in order to do so again. For example, as of September 2017, 95% of
account suspensions for promotion of terrorist activity were accomplished using our existing
proprietary detection tools—up from 74% in 2016. We are confident that the combination of our
dedicated teams, our detection tools, and other technological advancements at our disposal will
prove essential in addressing malicious human-coordinated activity as well.
As noted above, ability to detect malicious activity and spam on our platform is bolstered
by internal, manual reviews conducted by Twitter employees. Those efforts are further
supplemented by user reports, which allow us to refine and calibrate our detection tools and
carefully review content potentially in violation of the Twitter Rules and Twitter Terms of
Service. Twitter has also devoted resources to improving our process for reviewing user
reporting, including adding better technology to improve how we rank reports for review and
adopting policies to allow more reports filed by observers of abuse to be actioned.
30.

In hiring more content reviewers, are your companies simply throwing bodies at a
specific problem, or are you fundamentally rethinking how to prioritize which user
interactions require additional human oversight and review; if so, how? What other
changes have you made in this regard?
The answer to question 30 is provided in response to question 28.

20

[From Senator Manchin]
31.

Does Twitter or any Twitter affiliate use the information security products or
services of Kaspersky Lab or any Kaspersky Lab affiliate?

Neither Twitter nor any entity controlled by Twitter uses any Kaspersky Lab or
Kaspersky Lab-affiliated products or services.
32.

Does Twitter or any Twitter affiliate sell network space to Russia Today or Sputnik
news agencies?

Twitter recently off-boarded Russia Today (“RT”) and Sputnik and will no longer allow
those companies to purchase ad campaigns and promote Tweets on our platform. As we
announced in October 2017, Twitter will donate the $1.9 million that RT had spent globally on
advertising on Twitter to academic research into elections and civic engagement.
33.

If you recently terminated any agreements with Russia Today or Sputnik, on what
date did the termination become effective?
Twitter off-boarded RT and Sputnik as advertisers on October 26, 2017.

34.

Do Russia Today or Sputnik need to purchase advertising space on your platform,
or can they freely maintain a Page or distribute web content via their own or
affiliated Twitter accounts?

As noted above, Twitter recently off-boarded RT and Sputnik as advertisers on the
platform. With respect to organic (non-paid) content, in contrast, unless those accounts violate
the Twitter Rules or the Twitter Terms of Service—including by engaging in abusive behavior,
promoting violence, harassing individuals, posting prohibited content, or using the platform for
unlawful purposes—we do not restrict the type of content they choose to share with their
followers. As a global platform, we believe that our users benefit from the exchange of ideas
with other users around the world.
Twitter places greater limitations on the type of content that can be promoted with
Twitter ads compared to organic content that our users generate. We draw this distinction
because organic Tweets are generally shown to people who choose to follow the user that creates
it, while ads—Promoted Tweets—are served to a broader audience, including users who have
not chosen to follow the user or account that generated the ad.
35.

Does Twitter prohibit, or have any concern about, foreign state-sponsored news
organizations posting content via the Twitter platform?

Twitter does not categorically prohibit state-sponsored news organizations from posting
organic content on our platform. Twitter is a global company with hundreds of millions of users
accessing and engaging with information on the platform from around the world. Access to
news and real-time media reports is an essential feature of our platform, regardless of location.
As with any other account, we permit news organizations to post content on our platform that is

21

accessible to their followers, so long as they do not engage in illegal activity or otherwise violate
our Terms of Service.

22

[From Senator Harris]
36.

Your company has produced information about Russian propaganda
advertisements. Your company has also produced information about Russian
propaganda that appeared as ordinary user content. You have not, however,
provided information about the legitimate advertisements that accompanied
Russian content.
● How long do you retain placement and billing records for advertisements on
your services?
● Have you instructed your relevant business units to retain the records of
advertisements that accompanied Russian propaganda? If you have not, will you
immediately issue that instruction?

Twitter maintains advertisers’ billing records in the ordinary course of business. Those
records are and will be retained.
● How much revenue do you estimate that you earned from the advertising that
accompanied Russian propaganda?
● Have you notified the advertisers whose advertisements accompanied Russian
propaganda?
● What do you plan to do with the revenue that you earned from the
advertisements that accompanied Russian propaganda?
Twitter’s advertising revenue is primarily driven by our Promoted Products. The way in
which advertisers use our platform, and the nature of those Promoted Products, supports an
estimate that very little revenue was generated from advertising that “accompanied” Russian
propaganda on Twitter, as described below. Our Promoted Products are designed to be
incorporated into our platform as native advertising, ideally to be as compelling and useful to our
users as organic content on our platform. Given this design, Twitter’s advertising differs from
other platforms and most Twitter advertising does not accompany particular content. For
example, Twitter does not display banner ads that accompany a news story.
Twitter’s Promoted Products include Promoted Accounts, Promoted Trends, and
Promoted Tweets. Promoted Accounts appear in the same format and place as accounts
suggested by our Who to Follow recommendation engine, or in some cases, in Tweets in a user’s
timeline. Promoted Accounts provide a way for our advertisers to grow a community of users
who are interested in their business, products or services.
Promoted Trends appear at the top of the list of trending topics for an entire day in a
particular country or on a global basis. When a user clicks on a Promoted Trend, search results
for that trend are shown in a timeline and a Promoted Tweet created by our advertisers is
displayed to the user at the top of those search results.

23

Promoted Tweets, in the vast majority of cases, appear within a user’s timeline or search
results just like an ordinary Tweet, and the advertisement is the Tweet itself. These Promoted
Tweets do not “accompany” any specific Tweet, nor are they otherwise linked to a particular
account.
A small percentage of Promoted Tweets or Accounts (approximately 7% during the 2016
election period) are served to user profiles and may appear within the profile page of an account
that a user chooses to visit. As with other Promoted Product placements, to protect user
experience, only a very limited number of Promoted Product impressions will render in a profile
when another user views it.
The results of our retrospective review allow reasonable estimates of how much revenue
could have been generated by Promoted Tweets served to profiles of accounts linked to the
Internet Research Agency (“IRA”). Our review has found that a very small fraction of the total
content on Twitter during the pre-2016 election period originated from IRA accounts. Even in
the unlikely scenario in which ads appeared in each of the account profiles identified as linked to
the IRA, the maximum amount of revenue Twitter would have earned from those ads would be
very small, as described below.
The analysis we have conducted supports that estimate. We reviewed data from the
election time period concerning the number of impressions that were generally served on user
profiles available for ad impressions as well as the average revenue collected from those profiles.
Extrapolating from that data to a set of 2,752 accounts such as those that Twitter previously
identified as linked to the IRA, would yield an estimate of approximately $400 total revenue
during the late 2016 election period for Promoted Products that would have accompanied that
number of accounts by appearing in those user profiles. While this is a rough estimate based
upon aggregate data, it provides a sense of scale for revenue from the types of ads that could
accompany this content given the nature of Twitter’s advertising offerings during the relevant
time period.
Twitter has recently off-boarded Russia Today (“RT”) and Sputnik from running ad
campaigns on our platform on the basis of their efforts to disrupt the 2016 Presidential election
(as reported by the Intelligence Community) and due to violations of our advertising policies.
Twitter subsequently announced that we will donate the $1.9 million that RT had spent globally
on advertising on Twitter to academic research into elections and civic engagement.

24

37.

The problems of inauthentic, false, and hyper-partisan content are much broader
than Russian propaganda.
● How many of the accounts on your service do you estimate are inauthentic?
● If you are aware of independent estimates of inauthentic or false content on your
platforms, please provide those estimates. If you disagree with the estimates,
please explain why.
● How much of the activity on your service do you estimate is inauthentic or false?

Based on a review of a representative sample of accounts, we estimate that false or spam
accounts represent less than 5% of our MAUs. Our estimates are lower than those reported by
outside researchers because those researchers do not have access to critical internal information
necessary to make an accurate determination of the scale of spam, fake accounts or automated
bots on Twitter. As a result, reports from third-party researchers often overestimate the true
volume of such accounts on our platform—sometimes by large orders of magnitude.
While our detection tools for false or spam accounts rely on a number of inputs and
variables and do not operate with 100% precision, they are informed by information not
available outside of Twitter. Our internal researchers have access to and can analyze a number
of different signals including, among other things, email addresses, phone numbers, login
history, and other non-public account and activity characteristics that enable us to conduct a
more thorough review and reach more accurate conclusions as to whether the account in question
is fake or spam. We keep such information confidential and do not make it available to
researchers in order to protect the privacy of our users.
Because third-party researchers do not have access to internal signals that Twitter can
access, their bot and spam detection methodologies must be based on public information and
often rely on human judgment, rather than on internal signals available to us. One common
model for determining whether an account is fake or automated is the “Botometer model,” which
compares publicly available account features, such as Tweet count, follower count, and use of
language, to the characteristics exhibited by purportedly “known” bots. The initial evaluation of
whether an account is or is not a bot, however, relies on an individual assessment and is,
therefore, inherently imprecise.
There are also studies that use the limited public Tweet data that we offer researchers
through an application programming interface (“API”). The studies that rely on information
from the Twitter API to identify automated accounts similarly overestimate both the number and
impact of these accounts because our internal detection tools and filtering techniques are not
available to third parties. Those tools enable us to remove from the platform malicious
automated accounts (and content generated by such accounts), but the accounts may nevertheless
appear in the data stream that researchers access through our API, thus inaccurately reflecting the
traffic on Twitter.
A study conducted by the University of Southern California and Indiana University
estimated that as much as 15% of Twitter accounts are automated, spam accounts. That estimate,
however, was based on a prediction of whether a user may or may not be an automated account
25

and was derived from human judgments about an account’s attributes. The authors of the study
acknowledge that detecting automated accounts “is a hard task. Many criteria are used in
determining whether an account is controlled by a human or a bot, and even a trained eye gets it
wrong sometimes.” See https://botometer.iuni.iu.edu/#!/faq#bot-threshold.
We are committed to continuing to work on refining our spam detection tools and to
update the Twitter community and the public periodically about our estimates and analysis of
these things on our platform. We regularly receive and welcome input from researchers and
Twitter users about ways in which we can optimize our detection and enforcement methods. In
addition, as we have announced, we are also committed to donating the $1.9 million we
projected to have earned from RT advertising to support external research into the use of Twitter
in civic engagement and elections, including the use of malicious automation and
misinformation.
Note that Twitter does not prohibit the use of pseudonymous accounts (accounts used by
real people for non-spam purposes under any name they choose) provided they comply with the
Twitter Rules. Many common and powerful uses of Twitter are enabled by this policy, including
allowing religious and/or political dissidents and others to engage in free expression without fear
of retribution from oppressive governments. It also allows users to have accounts focused on
specific interests and to keep those interests separate from a professional or other account
associated with their real name. Other users make creative use of this ability to Tweet in the
name of a pet or to engage in parody or satire. Such uses of Twitter do not violate our Rules.
● If the independent estimates were accurate, how much of your annual revenue
would be attributable to inauthentic or false content?
● How much of your annual revenue do you estimate is attributable to inauthentic
or false content?
● Do you have a policy of notifying advertisers when their advertisements
accompany inauthentic or false content?
● What do you do with the revenue that you earn from advertisements that
accompany inauthentic or false content?
● How much of the news content that is shared on your services do you estimate is
false?
● How much of the news content that is shared on your services do you estimate is
hyper-partisan?
Twitter’s approach to addressing the spread of malicious automation and inauthentic
accounts on our platform is to focus on problematic behavior and abuse, not primarily on the
content that such accounts attempt to disseminate. We are committed to addressing the spread of
misinformation on our platform—and to prevent future attempts of interfering with U.S.
elections—but we recognize that spam and malicious automation are not limited to political
content and can undermine the positive user experience we seek to offer irrespective of the
content.
26

Accordingly, we monitor various behavioral signals related to the frequency and timing
of Tweets, Retweets, likes, and other such activity, as well as to similarity in behavioral patterns
across accounts, in order to identify accounts which are likely to be maliciously automated or
acting in an automated and coordinated fashion. We monitor and review unsolicited targeting of
accounts, including accounts that mention or follow other accounts with which they have had no
prior engagement. For example, if an account follows 1,000 users within the period of one hour,
or mentions 1,000 accounts within a short period of time, our systems are capable of detecting
that activity as aberrant and as potentially originating from suspicious accounts.
Twitter is continuing its effort to detect and prevent malicious automation by leveraging
our technological capabilities and investing in initiatives aimed at understanding and addressing
behavioral patterns associated with such accounts. For example, in early 2017, we launched the
Information Quality initiative, an effort aimed at enhancing the strategies we use to detect and
deny bad automation, improve machine learning to spot spam, and increase the precision of our
tools designed to prevent such content from contaminating our platform.
The spread of misinformation online is neither a new phenomenon nor unique to our
platform. Twitter takes this issue seriously, but we also recognize that our ability to monitor or
control the veracity of the content our users choose to share on the platform is limited. We
cannot prevent individuals from lying or exaggerating. And given the scale of activity on our
platform—where over 330 million users are Tweeting nearly half a billion Tweets per day in
scores of languages—we are not able to assess whether each of those Tweet contains arguably
inaccurate information (or assess revenue using that criterion).
We are open to examining new solutions to addressing this problem. But we also
recognize that the Twitter community itself remains one of the most powerful tools to addressing
the spread of misinformation. While it is true that false information and rumors can spread
quickly, accurate information—particular information directed at contesting untruths—
propagates in a similarly high velocity.
We have observed our users engage with false information by refuting it: they Retweet
it, reply to it, and Tweet original content contradicting it. As we noted in connection with the
Committee hearing, in response to the attempted “vote-by-text” effort and similar voter
suppression attempts during the 2016 election, Twitter restricted as inaccessible, pending
deletion, 918 Tweets from 529 users who proliferated that content. Twitter also permanently
suspended 106 accounts that were collectively responsible for 734 “vote-by-text” Tweets.
Twitter identified, but did not take action against, an additional 286 Tweets of the relevant
content from 239 Twitter accounts, because we determined that those accounts were seeking to
refute the “text-to-vote” message and alert other users that the information was false and
misleading. Notably, those refuting Retweets generated significantly greater engagement across
the platform compared to the Tweets spreading the misinformation—8 times as many
impressions, engagement by 10 times as many users, and twice as many replies.

27

● Have you conducted any studies of how false content performs on your services?
If yes, please describe those studies and provide copies.
● Have you conducted any studies of how hyper-partisan content performs on
your services? If yes, please describe those studies and provide copies.
Twitter has not conducted such studies. As noted above, our efforts to address malicious
activity on our platform focus on behavior rather than content.
38.

In the area of state-sponsored hacking, each of your companies has a responsible
senior executive and dedicated technical experts.
● Who is the senior executive responsible for countering state-sponsored
information operations? When did that executive assume that responsibility, and
what is the scope of the responsibility?

The threats posed by state-sponsored misinformation operations have the potential to
impact many parts of our company, including consumer product, advertising, and information
security teams. Twitter’s General Counsel and Head of Consumer Product, along with our Head
of Trust & Safety and Chief Information Security Officer, are generally responsible for ensuring
that the platform remains safe. That responsibility includes overseeing and directing Twitter’s
response to state-sponsored misinformation and malicious human coordinated and automated
activity. Twitter’s Information Quality team, formed in in early 2017, reports to Twitter’s Head
of Consumer Product and is intensively focused on enhancing the strategies we use to detect and
stop bad automation, improve machine learning to spot spam, and increase the precision of our
tools designed to prevent such content from contaminating our platform.
● As of November 2016, how many of your technical employees had the primary
day-to-day task of countering state-sponsored information operations?
● As of today, how many of your technical employees have the primary day-to-day
task of countering state-sponsored information operations?
Twitter dedicates significant resources to addressing malicious automation, bots, and
other coordinated activities. We believe we have the right resources and strategies in place. We
dedicated nearly the entire engineering, product, and design teams to look at these issues at the
beginning of 2017, and we regularly reexamine staffing and resources and adjust as needed.
Critical to the continued success of our efforts is our ability to leverage our technological
advancements and improvements to tackle this problem; given the scale, this is not something
that we or anyone else can review and address by hand alone. We have been successful at
addressing other challenges in other contexts and we believe we can meet this challenge as well.
For example, as of September 2017, 95% of account suspensions for promotion of terrorist
activity were accomplished using our existing proprietary detection tools—up from 74% in 2016.
We are confident that the combination of our dedicated teams and our ability to use our detection
tools and other technological advancements at our disposal equip us well to confront this
ongoing threat.

28

We also recognize that, at this time, computer algorithms alone are not sufficient to
address the problem. Accordingly, those tools are complemented by manual review teams,
collaboration and information sharing with industry peers and participants, reliance on data and
intelligence from third-party security vendors, and partnerships with other companies and civil
society.
Twitter understands that, to succeed, we must combine resources, information,
knowledge, and effort with industry partners, civil society, and government. We do not compete
against other companies on our ability to detect and label malicious content on our platform;
instead, we recognize that we will all be stronger if we view this as a shared threat. We are
committed to a continued collaborative approach and believe it will prove successful going
forward.
39.

Much of what we now know about Russian propaganda is because of academic
researchers and investigative journalists. These groups do not currently have access
to the data that they need to inform the public and to build tools for detecting statesponsored information operations. For example, these groups generally cannot
assess the full set of public user activity associated with a specific topic, nor can they
analyze the behavior of accounts associated with state-sponsored information
operations. Providing access to this data need not come at the expense of user
privacy, since these groups could be bound by non-disclosure agreements and use
privacy-preserving algorithms to conduct their studies.
● Will you commit to, by the end of the year, providing five or more independent,
non-profit entities with access to the data they need to understand and counter
state-sponsored information operations? If you will, please provide specifics and
a timeline for how you plan to honor the commitment. If you will not, please
explain why.

Twitter is working to deepen our partnership with independent researchers. An example
of this commitment is the significant resources we are dedicating to the effort. Twitter recently
off-boarded Russia Today (“RT”) and Sputnik from running ad campaigns on our platform on
the basis of their efforts to disrupt the 2016 Presidential election (as reported by the Intelligence
Community) and due to violations of our advertising policies. Twitter subsequently announced
that we will donate the $1.9 million that RT had spent globally on advertising on Twitter to
academic research into elections and civic engagement.
We are also implementing changes to our data services that will make our public data
more accessible than before for research purposes. These changes include, for example, new
services which offer developers greater historical search access than was previously accessible.
Twitter’s data services are unique in the industry, offering insights into the Twitter platform that
other companies do not provide.
In addition, Twitter is launching an industry-leading Transparency Center that will offer
the public better visibility into who is advertising on Twitter and how those ads are targeted.
That information will better enable users and outside parties to conduct their own research or
evaluation regarding particular ads.

29

40.

Similarly, much of what we now know about inauthentic, false, or hyper-partisan
content is because of independent groups.
● Will you commit to, by the end of the year, providing five or more independent,
non-profit entities with access to the data they need to understand the prevalence
and performance of inauthentic, false, or hyper-partisan content on your
services? If you will, please provide specifics and a timeline for how you plan to
honor the commitment. If you will not, please explain why.
The answer to question 40 has been provided in response to question 39.

41.

Addressing state-sponsored information operations will continue to require
cooperation among private sector entities and with the government.
● Have you established a formal mechanism for promptly sharing actionable
information about state-sponsored information operations with other online
services, similar to the mechanisms that already exist for sharing information
about state-sponsored cybersecurity threats? If not, will you commit to
developing such a mechanism?
● The FBI is the federal agency responsible for countering foreign propaganda.
Do you have a written policy of promptly sharing what you learn about statesponsored information operations with the FBI? If not, will you commit to
developing such a policy?

Twitter agrees that cooperation with other private-sector entities and the government is
necessary to address state-sponsored information operations. Twitter engages in information
sharing with its industry counterparts on a variety of threats and is committed to maintaining
such cooperative efforts.
Twitter has partnered with other platforms to make progress against common threats. In
June 2017, for example, we launched the Global Internet Forum to Counter Terrorism (the
“GIFCT”), a partnership among Twitter, YouTube, Facebook, and Microsoft. The GIFCT will
facilitate, among other things, information sharing, technical cooperation, and research
collaboration, including with academic institutions.
The GIFCT announced a commitment to create a shared industry database of “hashes”—
unique digital “fingerprints”—for violent terrorist imagery or terrorist recruitment videos or
images that have been removed from our individual services. The database allows a company
that discovers terrorist content on one of their sites to create a digital fingerprint and share it with
the other companies in the forum, who can then use those hashes to identify such content on their
services or platforms, review against their respective policies and individual rules, and remove
matching content as appropriate, or even block extremist content before it is posted in the first
place. The database now contains more than 40,000 hashes. Instagram, Justpaste.it, LinkedIn,
Oath, and Snap have also joined this joint initiative, and we are working to add several additional
companies in 2018.

30

Twitter also engages in regular discussions with law enforcement agencies, including the
FBI. We respond promptly to properly scoped legal process and valid requests for information
from those agencies. Given the difficulty of identifying and labeling activity as state-sponsored,
we also recognize the important role of government information sharing efforts, such as the
Intelligence Community’s Report about the 2016 U.S. election.
42.

You currently have automated systems in place to detect spam and abuse.
● Do you have an automated system in place to detect state-sponsored information
operations? If yes, will you provide this Committee with a private briefing on the
system’s design and performance? If no, why not?

Twitter’s approach to addressing the spread of malicious, inauthentic automation on our
platform is to focus on problematic behavior and abuse, not primarily on the content that such
accounts attempt to disseminate. We are committed to addressing the spread of misinformation
on our platform—and to prevent future attempts to interfere with U.S. elections—but we
recognize that spam and malicious automation are not limited to political content and can
undermine the positive user experience we seek to offer irrespective of the content.
Accordingly, we monitor various behavioral signals related to the frequency and timing
of Tweets, Retweets, likes, and other such activity, as well as to similarity in behavioral patterns
across accounts, in order to identify accounts that are likely to be maliciously automated or
acting in an automated and coordinated fashion in ways that are unwelcome to our users. We
monitor and review unsolicited targeting of accounts, including accounts that mention or follow
other accounts with which they have had no prior engagement. For example, if an account
follows 1,000 users within the period of one hour, or mentions 1,000 accounts within a short
period of time, our systems are capable of detecting that activity as aberrant and as potentially
originating from suspicious accounts.
Twitter is continuing its effort to detect and prevent malicious automation by leveraging
our technological capabilities and investing in initiatives aimed at understanding and addressing
behavioral patterns associated with such accounts. For example, in early 2017, we launched the
Information Quality initiative, an effort aimed at enhancing the strategies we use to detect and
stop bad automation, improve machine learning to spot spam, and increase the precision of our
tools designed to prevent such content from contaminating our platform.
Since the 2016 election, we have made significant improvements to reduce external
attempts to manipulate content visibility. These improvements were driven by investments in
methods to detect malicious automation through abuse of our API, limit the ability of malicious
actors to create new accounts in bulk, detect coordinated malicious activity across clusters of
accounts, and better enforce policies against abusive third-party applications.
In addition, we have developed new techniques for identifying patterns of activity
inconsistent with legitimate use of our platform (such as near-instantaneous replies to Tweets,
nonrandom Tweet timing, and coordinated engagement), and we are currently implementing
these detections across our platform. We have improved our phone verification process and
introduced new challenges, including reCAPTCHA (utilizing an advanced risk-analysis engine

31

developed by Google), to give us additional tools to validate that a human is in control of an
account. We have enhanced our capabilities to link together accounts that were formed by the
same person or that are working in concert. And we are improving how we detect when
accounts may have been hacked or compromised.
With our improved capabilities, we are now detecting and blocking approximately
450,000 suspicious logins each day that we believe to be generated through automation. In
September 2017, our systems identified and challenged an average of four million suspicious
accounts per week, which represents more than double our rate of detection at this time last year.
Over three million of those accounts were challenged upon signup, before their content or
engagements could impact other users. Between June and September 2017, we have also
suspended more than 117,000 malicious applications for API abuse. Those applications were
collectively responsible for more than 1.5 billion Tweets in 2017.
We plan to continue building upon our 2017 improvements, including through
collaboration with our peers and investments in machine-learning capabilities that help us detect
and mitigate the effect on users of fake, coordinated, and malicious automated account activity.
We have also observed the expansion of malicious activity on our platform from
automated accounts to human-coordinated activity, which poses additional challenges to making
our platform safe. We are determined to meet those challenges and have been successful in
addressing such abusive behavior in other contexts. We are committed to leveraging our
technological capabilities in order to do so again. For example, as of September 2017, 95% of
account suspensions for promotion of terrorist activity were accomplished using our existing
proprietary detection tools—up from 74% in 2016. We are confident that the combination of our
dedicated teams, our detection tools, and other technological advancements at our disposal will
prove essential in addressing malicious human-coordinated activity as well.
43.

You have promised to adopt additional transparency and verification requirements
for political advertising.
● Please detail the new requirements and your timeline for implementing those
requirements.
● How do you define the political advertisements that are covered by the new
requirements? Why did you adopt the definition that you did?
● Will you commit to including within your definition, at a minimum,
advertisements that advocate for or against a specific candidate, political party,
piece of legislation, regulatory action, or ballot referendum? If not, why not?

Twitter’s approach to greater transparency in political advertising centers on two
components: a new electioneering policy and an industry-leading Transparency Center. We
expect to roll out the new policy in the U.S. during the first quarter of 2018.
To make it clear when a user is viewing or engaging with content considered to be an
electioneering ad, our policy will require that advertisers that meet the definition of

32

electioneering identify their campaigns as such. We will also change the interface of such ads
and include a visual political ad indicator (see, e.g., Fig. 2 below).
Fig. 2: Template for New Electioneering Ad

Twitter’s definition of electioneering ads will be derived from the FEC regulations’
definition of that term, which includes any broadcast, cable, or satellite communication that
refers clearly to a candidate for federal office, is published 60 days before a general election or
30 days before a primary, convention, or caucus, and is targeted to the relevant electorate (if the
candidate is running for Congress).
The goal of the Transparency Center is to offer the public increased visibility into all
advertising on the platform, and to provide users with tools to share feedback with us. With
respect to electioneering ads and the Transparency Center, we intend to better enable users and
outside parties to conduct their own research or evaluation regarding particular ads.
Electioneering ads information accessible through the Transparency Center will include, among
other things, the identity of the organization funding the campaign, all ads that are currently
running or have run on Twitter, campaign spend, and targeting demographics for specific ads or
campaigns. We plan to launch the Transparency Center as soon as feasible after rolling out our
electioneering policy in the first quarter of 2018, and we are continuing to refine the tools we
will make available in conjunction launching the Transparency Center to ensure the best
experience for our users.

33

44.

Your platform offers a range of advertisement targeting criteria.
● Which types of targeting criteria, such as demographic, behavioral, lookalike, or
email matching, did Russia use for its information operations?

In connection with our retrospective review of Russian activity on our platform in 2016,
we identified nine accounts as being potentially linked to Russia that promoted election-related,
English-language content. Of the nine accounts that we identified as being potentially linked to
Russia and promoting election-related, English-language content, the most significant use of
advertising was by @RT_com and @RT_America. Those two accounts collectively ran 44
different ad campaigns, accounting for nearly all of the relevant advertising we reviewed.
Of all of RT’s ad campaigns, 11 were targeted exclusively at English-language speakers,
and several others—including all of @RT_America’s seven campaigns—used geographic
targeting to focus on U.S. users. Though many of the campaigns did not include specialized,
non-geographic targeting, a subset of the @RT_com campaigns targeted followers of other RT
accounts or followers of other leading news organizations based in the U.S. and other countries.
A small number of short “quick promote” campaigns used keyword targeting to attempt to reach
audiences searching for particular words or phrases.
The remaining seven accounts, which collectively ran approximately 50 ad campaigns,
used a broad range of targeting strategies. We did not identify a trend across the targeting
criteria used by those accounts. The accounts sporadically used English-language targeting and
location targeting at the country level (including the U.S., Canada, the UK, France, and Ukraine).
A handful of campaigns also sought to reach followers of certain accounts.
45.

Have you seen any evidence of state-sponsored information operations associated
with American elections in 2017, including the gubernatorial elections in Virginia
and New Jersey?

Twitter is not aware of any specific state-sponsored attempts to interfere in any American
elections in 2017, including the Virginia and New Jersey gubernatorial elections. However, our
automated systems for detecting and preventing abuse of our services (including our spam and
malicious automation) continually operate with the goal of ensuring that all conversations on
Twitter—including those surrounding elections—are spam- and abuse-free. As was publicly
reported, we were made aware of a surge in automated followers for a candidate in a recent
Senate election, immediately took action, and do not have any indication that the activity was
state-sponsored.

34

46.

User reports are an important signal of when an account is not authentic.
● How frequently do you receive user reports about inauthentic accounts?
● What is your process for responding to those reports? How often does that
process usually take?
● What proportion of those reports result in an account restriction, suspension, or
removal?
● Among the reports that you decline to take action on, what proportion involve
reported accounts that you subsequently identify as inauthentic?

Twitter receives approximately 5 million reports of spam content or malicious automated
Tweets, accounts, or interactions (e.g., follows) per month. The majority of those reports
automatically trigger a signal of potential inauthentic behavior by an account seeking to
manipulate our platform.
To prevent the misuse of reporting to trigger enforcement actions against users who are
not necessarily in violation of the Twitter Rules, user reports are not the sole factor that Twitter
considers when taking action against an account. Rather, we use a variety of signals—including
overall account behavior and interaction history—to determine whether a report of inauthentic
behavior warrants further review or action. Given the broad range of signals Twitter relies on in
determining whether an account should be restricted, suspended, or removed, we cannot with any
precision assess how many of the 5 million monthly reports directly result in action against the
reported account. In addition, Twitter accounts may be reported under different policies, for
different Tweets or content, and at different points in the account lifecycle. All of these factors
influence Twitter’s review and enforcement decisions.
It is not uncommon for an account to be subject to enforcement action later even if
Twitter does not initially suspend the account based on the initial report. For example, if Twitter
does not take action in response to a user report, the reporting user may be able to submit further
information for Twitter to consider during its review. Such supplemental reports may result in
Twitter taking action where it may not have previously. In some instances, such as in reports of
impersonation or intellectual property infringement, Twitter may require that the report include
specific types of information in order to take action against the account. Twitter may not be able
to take action until it has received that information. Twitter may also continue to receive reports
from the same or other reporters regarding an account or specific content, which may result in
Twitter taking action on those new reports at some point after the initial report.
● How many of the accounts that you have identified as associated with Russian
information operations were the subject of a user report? Please provide all the
user reports associated with these accounts and the actions that you took in
response, including the specific time for the report and each action.
We received user reports for a small minority of the IRA accounts previously identified
to the Committee, prior to the suspension of these accounts from Twitter. A large percentage of
those reports related to content from two accounts: @TEN_GOP and @SouthLoneStar. While
35

many of these reports were not actioned at the time, Twitter has since made substantial changes
to its operations and policies to respond more effectively to user reports. For example, in July of
2017, we announced that we are now taking action on 10 times the number of abusive accounts
every day compared to the same time last year. We also announced that we now limit account
functionality or place suspensions on thousands more abusive accounts each day.
47.

Much of the public discussion about state-sponsored information operations on your
platforms has centered on the Internet Research Agency. That is not the only group
surreptitiously spreading state-sponsored propaganda.
● What other groups are you tracking that are affiliated with the Russian
government?
● What other countries do you believe are conducting state-sponsored information
operations on your platforms? Please describe the groups that you are tracking
for each country, including both government agencies and affiliates.

As we noted in connection with the Committee hearing, there are technological limits to
what we can determine based on the information we can detect regarding a user’s origin. Based
in part on work conducted by our Information Quality team, we are aware of the fact that, among
other things, a high concentration of automated engagement and content originates from data
centers and users accessing Twitter via Virtual Private Networks (“VPNs”) and proxy servers,
which obscure the user’s location and affiliation. Twitter’s abuse and spam detection and
prevention systems and enforcements mechanisms operate without regard to the specific country
of origin of an offending Tweet or malicious account. Users who violate the Twitter Rules
against abuse, spam, malicious automation, or other forms of prohibited behavior are subject to
enforcement, regardless whether they are affiliated with specific state actors.
Information we receive from third parties, including government agencies, security
research firms, and NGOs, may allow us to reliably associate certain accounts with particular
groups (such as the IRA). And we will continue working with the Committee and other groups
to help identify further state-sponsored actors that seek to abuse our services and manipulate
activity on our platform.
48.

You have confirmed that you have systems for assessing whether a specific account
is automated (i.e. a “bot”) and whether a specific piece of content is being amplified
through automated means.
● Twitter allows bots to operate on its social network. When and why did Twitter
make that decision?

Automation has not been categorically prohibited on Twitter for years, primarily because
we recognize that it often serves a useful and important purpose. Automation is essential for
certain informational content, particularly when time is of the essence, including for law
enforcement or public safety notifications. Examples include Amber Alerts, earthquake and
other storm warnings, and notices to “shelter in place” during active emergency situations.
Automation is also used to provide customer service for a range of companies. For example, as
of April 11, 2017, users are able to Tweet @TwitterSupport to request assistance from Twitter.
36

If a user reports a forgotten password or has a question about our rules, the initial triage of those
messages is performed by our automated system—a Twitter-developed program to assist users in
troubleshooting account issues.
To maintain the integrity of our platform and to ensure a positive user experience, we
focus on addressing the spread of malicious automation, abusive content, and accounts that
engage in such behavior. To that end, we monitor various behavioral signals related to the
frequency and timing of Tweets, Retweets, likes, and other such activity, as well as to similarity
in behavioral patterns across accounts, in order to identify accounts which are likely to be
automated or acting in a malicious automated and coordinated fashion.
● How does Twitter differentiate between permissible automated activity (i.e.
“benign bots”) and impermissible automated activity (i.e. “bad bots”)?
Twitter distinguishes between “good” and “bad” automation based on the behavior of the
account, using both algorithm-driven behavior detections and manual reviews. Many of our
spam enforcement targeting malicious automation take place automatically and look for signals
such as high-volume Tweeting, inhuman response times, and coordinated activities across
accounts.
Thus, we monitor and review unsolicited targeting of accounts, including accounts that
mention or follow other accounts with which they have had no prior engagement. For example,
if an account follows 1,000 users within the period of one hour, or mentions 1,000 accounts
within a short period of time, our systems are capable of detecting that activity as aberrant and as
potentially originating from suspicious accounts. An example of automation that typically does
not trigger our detection tools involves automatic customer service responses to user’s Tweets
that include a company’s handle.
● When a user visits a profile page, Twitter does not currently indicate whether it
believes the profile belongs to a bot. Will you commit to providing a visual
indication to users that Twitter believes that the account is a bot, so that users
can better understand and evaluate the content that they see? If not, why not?
● When a user encounters a piece of content, you do not currently indicate
whether the content is being amplified through automated means. Will you
commit to providing a visual indication to users of whether Twitter believes that
the content has been amplified through automation? If not, why not?
Because automated content can originate with any type of account, such disclosure would
necessarily be both over-inclusive and under-inclusive. For example, certain automation tools
that do not violate our Rules enable users to schedule their Tweets to post automatically at a
particular time or in response to specific activity. Since they may be utilized by individual users,
applying a categorical designation to accounts that employ those tools would not accurately
characterize such accounts.
We are committed to keeping Twitter a safe environment, and we continue to invest in
improving our systems for detecting and preventing malicious uses of our platform to amplify
content using automation. We prohibit the use of automation to artificially amplify content.
37

Once our systems detect an account as generating automated content or spam, we can take action
against that account at either the account level or the Tweet level (e.g., hiding the Tweet,
revoking a user’s ability to post content on the platform until that user deletes the Tweet in
question, or temporarily or permanently suspending the account).
49.

Inauthentic accounts can be disabled subsequent to automated or manual review.
● What role do automated and human employee review play in your decision to
disable a suspected inauthentic account?

Our systems are built to detect malicious automation and spam accounts across their
lifecycles, including detection at the account creation and login phase and detection based on
unusual activity (e.g., patterns of Tweets, likes, and follows). Our ability to detect such activity
on our platform is bolstered by internal, manual reviews conducted by Twitter employees. Those
efforts are further supplemented by user reports, which we rely on not only to address the content
at issue but also to calibrate our detection tools to identify similar content as spam.
Once our systems detect an account as generating malicious automated content or spam,
we can take action against that account at either the account level or the Tweet level. Depending
on the mode of detection, we have varying levels of confidence about our determination that an
account is violating our rules. We have a range of options for enforcement, and generally, the
higher our confidence that an account is violating our rules, the stricter our enforcement action
will be, with immediate suspension as the harshest penalty. If we are not sufficiently confident
to suspend an account on the basis of a given detection technique, we may challenge the account
to verify a phone number or to otherwise prove human operation, or we may flag the account for
review by Twitter personnel. Until the user completes the challenge, or until the review by our
teams has been completed, the account is temporarily suspended; the user cannot produce new
content (or perform actions like Retweets or likes), and the account’s contents are hidden from
other Twitter users.
We also have the capability to detect suspicious activity at the Tweet level and, if certain
criteria are met, to internally tag that Tweet as spam otherwise suspicious. Tweets that have
been assigned those designations are hidden from searches, do not count toward generating
trends, and generally will not appear in feeds unless a user follows that account. Typically, users
whose Tweets are designated as spam are also put through the challenges described above and
are suspended if they cannot pass.
For safety-related TOS violations, we have a number of enforcement options. For
example, we can stop the spread of malicious content by categorizing a Tweet as “restricted
pending deletion,” which requires a user to delete the Tweet before the user is permitted to
continue using the account and engaging with the platform. So long as the Tweet is restricted—
and until the user deletes the Tweet—the Tweet remains inaccessible to and hidden from all
Twitter users. The user is blocked from Tweeting further unless and until they delete the
restricted Tweet. This mechanism is a common enforcement approach to addressing less severe
content violations of our TOS outside the spam context; it also promotes education among our
users. More serious violations, such as posting child sexual exploitation or promoting terrorism,
result in immediate suspension and may prompt interaction with law enforcement.

38

● Do you require that a human employee review a suspected inauthentic account
before it is disabled?
Our ability to detect such activity on our platform is bolstered by internal, manual
reviews conducted by Twitter employees. Those efforts are further supplemented by user
reports, which we rely on not only to address the activity at issue but also to calibrate our
detection tools to identify similar content as spam and to enforce the Twitter Terms of Service
and Twitter Rules.
While our automated systems provide significant opportunities to scale enforcement
activity on spam (though presenting far greater challenges in other areas where account level
signals are less direct and user content itself is the focus), they frequently perform better when
supplemented by human review of content. For example, our automated systems flag accounts
that engage in suspicious activity for further manual review by Twitter personnel. Those
systems also assist our staff in prioritizing manual reviews of user reports. However, the
majority of enforcement actions against spam and malicious automated accounts are assisted by
automated systems. We continue to invest in improving these systems while ensuring that we
maintain a low rate of false positives to protect our users. And where we do not have sufficient
confidence in an automated assessment to take immediate action against a suspicious account,
that account may be reviewed by a Twitter employee.
● If so, given the rate at which inauthentic accounts can be regenerated, how do
you anticipate remaining ahead of the problem?
● What are you doing to improve automation in the process of detecting and
disabling inauthentic accounts?
● What are you doing to make it more difficult to establish inauthentic accounts?
Twitter is continuing its effort to detect and prevent malicious automation by leveraging
our technological capabilities and investing in initiatives aimed at understanding and addressing
behavioral patterns associated with such accounts. For example, in early 2017, we launched the
Information Quality initiative, an effort aimed at enhancing the strategies we use to detect and
stop bad automation, improve machine learning to spot spam, and increase the precision of our
tools designed to prevent such content from contaminating our platform.
Since the 2016 election, we have made significant improvements to reduce external
attempts to manipulate content visibility. These improvements were driven by investments in
methods to detect malicious automation through abuse of our API, limit the ability of malicious
actors to create new accounts in bulk, detect coordinated malicious activity across clusters of
accounts, and better enforce policies against abusive third-party applications.
In addition, we have developed new techniques for identifying patterns of activity
inconsistent with legitimate use of our platform (such as near-instantaneous replies to Tweets,
nonrandom Tweet timing, and coordinated engagement), and we are currently implementing
these detections across our platform. We have improved our phone verification process and
introduced new challenges, including reCAPTCHA (utilizing an advanced risk-analysis engine
developed by Google), to give us additional tools to validate that a human is in control of an
39

account. We have enhanced our capabilities to link together accounts that were formed by the
same person or that are working in concert. And we are improving how we detect when
accounts may have been hacked or compromised.
With our improved capabilities, we are now detecting and blocking approximately
450,000 suspicious logins each day that we believe to be generated through automation. In
September 2017, our systems identified and challenged an average of four million suspicious
accounts per week, which represents more than double our rate of detection at this time last year.
Over three million of those accounts were challenged upon signup, before their content or
engagements could impact other users. Between June and September 2017, we have also
suspended more than 117,000 malicious applications for API abuse. Those applications were
collectively responsible for more than 1.5 billion Tweets in 2017.
We plan to continue building upon our 2017 improvements, including through
collaboration with our peers and investments in machine-learning capabilities that help us detect
and mitigate the effect on users of fake, coordinated, and malicious automated account activity.
50.

According to recent news reports, Twitter hosted content that was intended to
disenfranchise voters in the Virginia gubernatorial election by misleading them
about how to vote.
● What automated and manual processes does Twitter have in place to identify
content that is intended to disenfranchise voters?
● How quickly does Twitter remove content that is intended to disenfranchise
voters? Please provide a histogram or quantile data.

When Twitter receives reports of illegal voter suppression content, we review the Tweets
and accounts in question and, where appropriate, take action to remove the Tweets or suspend
the accounts for violating the Twitter Rules. We also employ the use of automated systems to
identify “lookalike” posts (i.e., posts which share a known misleading image) which were not
directly reported but which should be reviewed.
For example, during the period leading up to the 2016 election, in response to the
attempted “vote-by-text” effort and similar voter suppression attempts, Twitter restricted as
inaccessible, pending deletion, 918 Tweets from 529 users who proliferated that content. Twitter
also permanently suspended 106 accounts that were collectively responsible for 734 “vote-bytext” Tweets. Twitter took action against the first reports of those Tweets within a day or two;
once we calibrated our systems to detect the vote-by-text content, our response time decreased
and we were able to hide the content rapidly. Because such content removal takes place on a
case-by-case basis and involves a fact-specific inquiry and human review, there is not a uniform
frequency or pattern to such enforcement actions that we can depict with graphs.

40

[From Senator McCain]
51.

Current campaign finance law establishes disclosure standards for television, radio,
and print media. The Pew Research Center recently found that 65 percent of
Americans identified an Internet-based source as their leading source of
information in the 2016 election.
● Under current law, to what extent is Twitter responsible for providing a similar
quality of disclosure to the public?

Current campaign finance laws require advertisers to include disclosure language on
certain “public communications” and “electioneering communications.” Compliance with FEC
regulations and guidance from advisory opinions rests with the advertiser, rather than on Twitter
or the television, radio, print or digital provider on which an advertiser runs an ad. The Federal
Election Commission—through its regulations and advisory opinions—has advised advertisers
who disseminate paid communications on digital platforms that they may be able to rely on
regulatory exceptions to the disclaimer requirements for “small items” or for communications
where including a disclaimer is “impracticable.”
Twitter will take further steps to promote transparency and public understanding through
the Transparency Center that we publicly announced last year. The goal of the Transparency
Center is to offer the public increased visibility into advertising on the platform, and to provide
users with tools to share feedback with us. With respect to electioneering ads and the
Transparency Center, we intend to better enable users and outside parties to conduct their own
research or evaluation regarding particular ads. Electioneering ads information that will be
accessible through the Transparency Center will include, among other things, the identity of the
organization funding the campaign, all ads that are currently running or have run on Twitter,
campaign spend, and targeting demographics for specific ads or campaigns. We plan to launch
the Transparency Center as soon as feasible after rolling out our electioneering policy in the first
quarter of 2018, and we are continuing to refine the tools we will make available in conjunction
with launching the Transparency Center to ensure the best experience for our users.
52.

Your platform hosts thousands of tweets per second, or billions of tweets every year.
● Given the challenge of monitoring such a vast amount of content, to what extent
is the monitoring of campaign advertisements automated?
● Do you feel that this automation is sufficient in capturing bad actors?

Twitter relies on two methods to prevent prohibited promoted content from appearing on
the platform: a proactive method and a reactive method. Proactively, Twitter relies on custombuilt algorithms and models for detecting Tweets or accounts that might violate its advertising
policies. Reactively, Twitter takes user feedback through a “Report Ad” process, which flags an
ad for manual human review.
On the proactive side, an advertisement and advertiser account that is subject to manual
review is first reviewed by a set of machine classifiers that are built to detect Twitter Policy
violations; any suspicious ads that those models flag are subsequently reviewed by Twitter
41

personnel. On the reactive side, when we receive reports through our “Report an Ad” service,
those ads are similarly subject to manual review.
We believe the balance of proactive and reactive review allows us to actively enforce our
policies through an effective process that incorporates user feedback. We continue to invest in
improving our detection tools and developing new machine-learning models to improve
detection accuracy and remain up-to-date with new trends and new Twitter policies.
As part of our ads transparency and electioneering ads efforts announced in October
2017, Twitter will require advertisers to go through an onboarding process, which will obligate
them to provide information about how they are funding their media buys. For advertisers who
do not self-identify but who run electioneering ads, we will use a combination of machinelearning models and human manual review to detect and halt these advertisers until they have
correctly onboarded with us as an electioneering advertiser. While it is possible that foreign
governments may attempt to purchase ads through consultants or management companies,
Twitter’s upcoming Transparency Center is intended provide identifying information about any
such companies and their other advertising activities on Twitter. That information will better
enable users and outside parties to conduct their own research or evaluation regarding particular
ads.
With respect to organic content, Twitter dedicates significant resources to addressing and
blocking the use of malicious automation, bots, and other coordinated activities on our platform.
We believe we have the right resources and strategies in place. We dedicated nearly the entire
engineering, product, and design teams to look at these issues at the beginning of 2017, and we
regularly reexamine staffing and resources and adjust as needed.
Critical to the continued success of our efforts is our ability to leverage our technological
advancements and improvements to tackle this problem; given the scale, this is not something
that we or anyone else can review and address by hand. We have been successful at addressing
other challenges in other contexts and we believe we can meet this challenge as well. For
example, as of September 2017, 95% of account suspensions for promotion of terrorist activity
were accomplished using our existing proprietary detection tools—up from 74% in 2016.
We also recognize that computer algorithms alone are not sufficient to address the
problem. Accordingly, those tools are complemented by manual review teams, collaboration and
information sharing with industry peers and participants, reliance on data and intelligence from
third-party security vendors, and partnerships with other companies and civil society.
Twitter understands that, to succeed, we must combine resources, information,
knowledge, and effort with industry partners, civil society, and government. We do not compete
against other companies on our ability to detect and label malicious content on our platform;
instead, we recognize that we will all be stronger if we view this as a shared threat. We are
committed to a continued collaborative approach and believe it will prove successful going
forward.

42

● What sort of appeals process is in place in order to prevent faulty sorting?
A user suspended from Twitter can file an appeal directly from the Twitter mobile
application or from our website. Twitter’s Help Center, accessible at http://help.twitter.com,
provides additional information to users about how to an appeal of a suspension.
Suspended advertisers are also able to file appeals. In many cases, an advertiser will
work with their account representative in order to do so. If an advertiser does not have a Twitter
account representative, or if the advertiser prefers to proceed independently, the advertiser can
file an appeal through Twitter’s advertiser support form.
53.

Twitter recently announced efforts to make campaign advertising more transparent,
including the development of a “transparency center,” and harsher penalties for
electioneering advertisers that violate policies.
● How do you plan to detect false disclosure information and enforce your
policies?

Twitter is developing mechanisms and processes for verifying the information we will
ask our advertisers to disclose, and policies governing enforcement of and compliance with the
applicable ads policies. For advertisers who do not self-identify, but who run electioneering ads,
we will use a combination of machine-learning models and human manual review to detect and
halt these advertisers until they have correctly onboarded with us as an electioneering advertiser.

43