Introduction
This report covers the content moderation activities of X’s international entity Twitter International Unlimited Company (TIUC) under the Digital Services Act (DSA), during the date range 21 October, 2023 to 31 March, 2024.
We refer to “notices” as defined in the DSA as “user reports” and “reports”.
Our content moderation systems are designed and tailored to mitigate systematic risks without unnecessarily restricting the use of our service and fundamental rights, especially freedom of expression. Content moderation activities are implemented and anchored on principled policies and leverage a diverse set of interventions to ensure that our actions are reasonable, proportionate and effective. Our content moderation systems blend automated and human review paired with a robust appeals system that enables our users to quickly raise potential moderation anomalies or mistakes.
Policies
X's purpose is to serve the public conversation. Violence, harassment, and other similar types of behaviour discourage people from expressing themselves, and ultimately diminish the value of global public conversation. Our Rules are designed to ensure all people can participate in the public conversation freely and safely.
X has policies protecting user safety as well as platform and account integrity. The X Rules and policies are publicly accessible on our Help Center, and we are making sure that they are written in an easily understandable way. We also keep our Help Center regularly updated anytime we modify our Rules.
Enforcement
When determining whether to take enforcement action, we may consider a number of factors, including (but not limited to) whether:
When we take enforcement actions, we may do so either on a specific piece of content (e.g., an individual post or Direct Message) or on an account. We may employ a combination of these options. In most cases, this is because the behaviour violates the X Rules.
To enforce our Rules, we use a combination of machine learning and human review. Our systems are able to surface content to human moderators who use important context to make decisions about potential violations. This work is led by an international, cross-functional team with 24-hour coverage and the ability to cover multiple languages. We also have a complaints process for any potential errors that may occur.
To ensure that our human reviewers are prepared to perform their duties we provide them with a robust support system. Each human reviewer goes through extensive training and refreshers, they are provided with a suite of tools that enable them to do their jobs effectively, and they have a suite of wellness initiatives available to them. For further information on our human review resources, see the section titled “Human resources dedicated to Content Moderation”.
Reporting violations
X strives to provide an environment where people can feel free to express themselves. If abusive behaviour happens, we want to make it easy for people to report it to us. EU users can also report any violation of our Rules or their local laws, no matter where such violations appear.
Transparency
We always aim to exercise moderation with transparency. Where our systems or teams take action against content or an account as a result of violating our Rules or in response to a valid and properly scoped request from an authorised entity in a given country, we strive to provide context to users. Our Help Center article explains notices that users may encounter following actions taken. We will also promptly notify affected users about legal requests to withhold content, including a copy of the original request, unless we are legally prohibited from doing so.
X employs a combination of heuristics and machine learning algorithms to automatically detect content that violates the X Rules and policies enforced on our platform. We use combinations of natural language processing models, image processing models and other sophisticated machine learning methods to detect potentially violative content. These models vary in complexity and in the outputs they produce. For example, the model used to detect abuse on the platform is trained on abuse violations detected in the past. Content flagged by these machine learning models are either reviewed by human content reviewers before an action is taken or, in some cases, automatically actioned based on model output. Heuristics are typically utilised to enable X to react quickly to new forms of violations that emerge on the platform. Heuristics are common patterns of text or keywords that may be typical of a certain category of violations. Pieces of content detected by heuristics may also get reviewed by human content reviewers before an action is taken on the content. These heuristics are used to flag content for review by human agents and prioritise the order such content is reviewed.
Automated enforcements under the X Rules and policies undergo rigorous testing before being applied to the live product. Both machine learning and heuristic models are trained and/or validated on thousands of data points and labels (e.g., violative or non-violative) that are generated by trained human content reviewers. For example, inputs to content-related models can include the text within the post itself, the images attached to the post, and other characteristics. Training data for the models comes from both the cases reviewed by our content moderators, random samples, and various other samples of pieces of content from the platform.
Before any given algorithm is launched to the platform, we verify its detection of policy violating content or behaviour by drawing a statistically significant test sample and performing item-by-item human review. Reviewers have expertise in the applicable policies and are trained by our Policy teams to ensure the reliability of their decisions. Human review helps us to confirm that these automations achieve a level of precision, and sizing helps us understand what to expect once the automations are launched.
In addition, humans proactively conduct manual content reviews for potential policy violations. We conduct proactive sweeps for certain high-priority categories of potentially violative content both periodically and during major events, such as elections. Agents also proactively review content flagged by heuristic and machine learning models for potential violations of other policies, including our sensitive media, child sexual exploitation (CSE) and violent and hateful entities policies.
Once reviewers have confirmed that the detection meets an acceptable standard of accuracy, we consider the automation to be ready for launch. Once launched, automations are monitored dynamically for ongoing performance and health. If we detect anomalies in performance (for instance, significant spikes or dips against the volume we established during sizing, or significant changes in user complaint/overturn rates), our Engineering (including Data Science) teams - with support from other functions - revisit the automation to diagnose any potential problems and adjust the automations as appropriate.
Art. 15.1.c: TIUC Terms of Service and Rules Restricted Reach Labels - 21/10/23 to 31/3/24 | |||
Policy | Auto-Enforced | Manually Enforced | Total |
21,853 | 21,853 | ||
437,410 | 38,298 | 475,708 | |
5,359 | 5,359 | ||
Total | 437,410 | 65,510 | 502,920 |
Important Note: The table lists actions of visibility filtering on content potentially violative of our rules in accordance with our Freedom of Speech, Not Reach enforcement philosophy. We did not apply any visibility filtering based on illegal content. In cases where we receive a request for illegal content and the post is also found to be in violation of the policy where FOSNR is applied, our enforcement for illegal content will always be applied.
Important Notes about Action based on TIUC Terms of Service and Rules Violations:
Art. 15.1.a: Removal Orders Received - 21/10/23 to 31/3/24 | |||
Member State | Unsafe and/or Illegal Products | Illegal or Harmful Speech | Total |
France | 8 | 8 | |
Italy | 1 | 1 | |
Spain | 4 | 4 | |
Total | 8 | 5 | 13 |
X provides an automated acknowledgement of receipt of removal orders submitted by law enforcement through our Legal Request submission portal. As a consequence of this immediate acknowledgement of receipt, the median time was zero hours.
The median handle time to resolve removal orders during the reporting period was 4.1 hours.
Important Notes about Removal Orders:
X provides an automated acknowledgement of receipt of information requests submitted by law enforcement through our Legal Request submission portal. As a consequence of this immediate acknowledgement of receipt, the median time is zero.
The median time to resolve information requests during the reporting period was 74 hours.
Important Notes about Information Requests:
ACTIONS TAKEN ON ILLEGAL CONTENT:
ACTIONS TAKEN ON ACCOUNTS FOR POSTING ILLEGAL CONTENT: We suspended accounts in response to 11,268 reports of Intellectual Property Infringements. This was the only type of violation of local law that resulted in account suspension as many types of illegal behaviour are addressed in our policies, such as account suspensions for posting CSE.
REPORTS OF ILLEGAL CONTENT
REPORTS RESOLVED BY ACTIONS TAKEN ON ILLEGAL CONTENT
REPORTS OF ILLEGAL CONTENT MEDIAN HANDLE TIME
The median time to resolve illegal content notices during the reporting period was 2.7 hours.
Important Notes about Actions taken on illegal content:
Art. 15.1.d: Illegal Content Complaints - 21/10/23 to 31/3/24 | |||
Volume | Overturns after complaint | Complaints Rejected | |
Complaints | 667 | 190 | 477 |
The median time to resolve illegal content complaints is 2.8 hours.
The median time to resolve TOS complaints is 0.34 hours.
Important Notes about Complaints:
The possible rate of error of the automated means used in fulfilling those purposes, and any safeguards applied.
Important Notes about indicators of accuracy:
To date, zero disputes have been submitted to the out-of-court settlement bodies.
We received zero reports from Article 22 DSA approved trusted flaggers during the reporting period. Once Article 22 DSA awarded trusted flaggers information is published, we are prepared to enrol them in our trusted flaggers program, which ensures prioritisation of human review.
Today, we have 1,726* people working in content moderation. Our teams work on both initial reports as well as on complaints of initial decisions across the world (and are not specifically designated to only work on EU matters).
*This number was updated on 5/31/2024
X’s scaled operations team possesses a variety of skills, experiences, and tools that allow them to effectively review and take action on reports across all of our rules and policies. X has analysed which languages are most common in reports reviewed by our content moderators and has hired content moderation specialists who have professional proficiency in the commonly spoken languages. The following table is a summary of the the number of people in our content moderation team who possess professional proficiency in the most commonly spoken languages in the EU on our platform:
Art. 42.2: Linguistics Expertise* | |
Primary Language | People |
Arabic | 27 |
Dutch | 0 |
English | 1,535 |
French | 52 |
German | 59 |
Hebrew | 2 |
Italian | 2 |
Portuguese | 16 |
Spanish | 29 |
Latvian | 1 |
Bulgarian | 2 |
Polish | 1 |
*This table (Art. 42.2: Linguistics Experience) was updated on 5/31/2024
X has built a specialised team made up of individuals who have received specific training in order to assess and take action on illegal content that we become aware of via reports or other processes such as on our own initiative. This team consists of different tier groups, with higher tiers consisting of more senior, or more specialised, individuals.
When handling a report of illegal content or a complaint against a previous decision, content and senior content reviewers first assess the content under X’s Rules and policies. If no violation of X’s Rules and policies is determined warranting a global removal of the content, the content reviewers assess the content for potential illegality. If the content is not manifestly illegal, it can be escalated for second or third opinions. If more detailed investigation is required, content reviewers can escalate reports to experienced policy and/or legal request specialists who have also undergone in-depth training. These individuals take appropriate action after carefully reviewing the report or complaint and available context in close detail. In cases where this specialist team still cannot determine a decision regarding the potential illegality of the reported content, the report can be discussed with in-house legal counsel. Everyone involved in this process works closely together with daily exchanges through meetings and other channels to ensure the timely and accurate handling of reports.
All teams involved in solving reports of illegal content closely collaborate with a variety of other policy teams at X who focus on safety, privacy, authenticity rules and policies. This cross-team effort is particularly important in the aftermath of tragic events, such as violent attacks, to ensure alignment and swift action on violative content.
Content reviewers are supported by team leads, subject matter experts, quality auditors and trainers. We hire people with diverse backgrounds in fields such as law, political science, psychology, communications, sociology and cultural studies, and languages.
Training and support of persons processing legal requests
All team members are trained and retrained regularly on our tools, processes, rules and policies, including special sessions on cultural and historical context. Initially when joining the team at X, each individual follows an onboarding program and receives individual mentoring during this period, as well as thereafter through our Quality Assurance program (for external employees), in house and external counsels (for internal employees).
All team members have direct access to robust training and workflow documentation for the entirety of their employment, and are able to seek guidance at any time from trainers, leads, and internal specialist legal and policy teams as outlined above as well as managerial support.
Updates about significant current events or rules and policy changes are shared with all content reviewers in real time, to give guidance and facilitate balanced and informed decision making. In the case of rules and policy changes, all training materials and related documentation is updated. Calibration sessions are carried out frequently during the reporting period. These sessions aim to increase collective understanding and focus on the needs of the content reviewers in their day-to-day work.
The entire team also participates in obligatory X Rules and policies refresher training as the need arises or whenever rules and policies are updated. These trainings are delivered by the relevant policy specialists who were directly involved in the development of the rules and policy change. For these sessions we also employ the “train the trainer” method to ensure timely training delivery to the whole team across all of the shifts. All team members use the same training materials to ensure consistency.
There is a robust training program and system in place for every workflow to provide content moderators with the adequate work skills and job knowledge required for processing user cases. All agents must be trained in their assigned workflows. These focus areas ensure that X agents are set up for success before and during the content moderation lifecycle, which includes:
X’s training programs and resources are designed based on needs, and a variety of modalities are employed to diversify the agent learning experience, including:
Classroom training is delivered either virtually or face-to-face by expert trainers. Classroom training activities can include:
When agents successfully complete their classroom training program, they undergo a nesting period. The nesting phase includes case study by observation, demonstration and hands-on training on live cases. Quality audits are conducted for each nesting agent and agents must be coached for any mis-action spotted in their quality scores the same day that the case was reviewed. Trainers conduct needs assessment for each nesting agent and prepare refresher training accordingly. After the nesting period, content is evaluated on an ongoing basis with a team of Quality Analysts to identify gaps and address potential problem areas.
When an agent needs to be upskilled, they receive training of a specific workflow within the same pillar that the agent is currently working. The training includes a classroom training phase and nesting phase which is specified above.
Refresher sessions take place when an agent has previously been trained, has access to all the necessary tools, but would need a review of some or all topics. This may happen for content moderators who have been on prolonged leave, transferred temporarily to another content moderation policy workflow, or ones who have recurring errors in the quality scores. After a needs assessment, trainers are able to pinpoint what the agent needs and prepare a session targeting their needs and gaps.
During the period from 21 October, 2023 through 31 March, 2024 there were an average of 109,191,304 active recipients of the service (AMARS) in the EU.
The AMARS for the entire EU over the past six months is 66.1M. The difference between the total AMARs for the EU and the cumulative total AMARs for all EU member states is due to double counting of logged out users accessing X from various EU countries within the relevant time period.
- - - - - - - - - - - - - - - - - - Appendix - - - - - - - - - - - - - - - - -