RTFP

X's content moderation governance is structured around a hybrid approach combining automated detection with human review to enforce the X Rules and policies. This work is led by an international, cross-functional team providing 24-hour coverage and multilingual capabilities. For illegal content reports (including under DSA), specialized tiered teams assess under X Rules first, escalating to policy/legal specialists or in-house counsel for complex local law cases, ensuring timely and consistent handling through daily collaboration. Content moderation functions operate within this scaled operations framework, blending engineering, policy, and legal input for proportionality and freedom of expression safeguards.

X uses a combination of automated detection methods on its own initiative, including machine learning models (natural language processing, image processing, and other sophisticated machine learning techniques trained on past violations), heuristics (patterns of behavior, text, or keywords for rapid response to emerging violations), hash-matching technologies (e.g., PhotoDNA for known CSAM), and keyword/phrase blocking, to detect X Rules violating content. A vast majority of all accounts that are suspended for the promotion of terrorism and CSE are proactively flagged by a combination of technology and other purpose-built internal proprietary tools. When we remove CSE content with these automated systems, we immediately report it to the National Center for Missing and Exploited Children (NCMEC). NCMEC makes reports available to the appropriate law enforcement agencies around the world to facilitate investigations and prosecutions. Our current methods deploy a range of internal tools and and third party solutions that utilises industry standard hash libraries (e.g., PhotoDNA) to ensure known CSAM is caught prior to any user reports being filed. We leverage the hashes provided by NCMEC and industry partners. We scan media uploaded to X for matches to hashes of known CSAM sourced from NGOs, law enforcement and other platforms. We also have the ability to block keywords and phrases from Trending and block search results for certain terms that are known to be associated with CSAM. Content detected by X’s automated detection methods is either automatically actioned (based on model accuracy) or routed for human review. Proactive human sweeps also occur for high-priority categories (e.g., during elections or for CSE/violent extremism). Enforcement actions include content removal, account suspensions, visibility restrictions, and soft measures like labels or warnings. For high-priority illegal content, proactive detection often leads to immediate removal and reporting to authorities. X’s moderation of content follows an "information-first" approach to reduce bias and increase enforcement consistency by having moderators get to an enforcement decision by answering a set series of questions, rather than having them immediately make a decision, with appeals available.

Manual content moderation resourcing requirements can experience fluctuations based on a variety of challenges such as trending issues and product feature changes. To address this, recurring operational capacity review meetings are held that consider incoming volumes, our meet rate against service legal agreements, any case backlog accumulation, and assessment of risk. As a result of this analysis, moderation resources may be reallocated, removed or reserves committed to address emergent crises and opportunities.

Moderators are recruited using a standard job description that includes a language requirement which states that the candidate should be able to demonstrate written and spoken fluency in the language and have at least one year of work experience for entry-level positions. In the interview and application process, each agent candidate must meet certain linguistic standards to be considered “language qualified”. This determination is made through multiple tests (i.e. written, oral, etc.) of the candidate’s respective language, to determine their respective proficiency level. Candidates must also meet the educational and background requirements in order to be considered, as well as demonstrate an understanding of current events for the country or region of content moderation they will support.

We use a variety of performance metrics to monitor performance in detecting TOS violating content on-platform, which may include but is not limited to volume, false positive rate, appeal rate, appeal overturn rate, user reports and more. For example, we have feedback loops for our automated detection systems to monitor their performance using the rate at which human content reviews agree with the automated system decision. Reviewers have expertise in the applicable policies and are trained by our policy specialists 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. Thresholds are set according to the severity of the violation and remediation. We set precision targets that balance the rate of potential false positives with the ease of overturning the enforcement decision as well as the severity of missing violative content. This strategy holds both before we launch a new control, such as restricted reach labelling, and when we reassess the effectiveness of an existing measure. Given the inherent risks of false positives and false negatives, our automation systems 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 teams - with support from other functions - revisit the automation to diagnose any potential problems and adjust the automations as appropriate.

X employs automated means through a hybrid system combining machine learning models, heuristics and large language models (“LLM”) to detect and moderate content violating the X Rules and policies. Machine learning models include natural language processing for text analysis, image processing models for visual content, and other sophisticated algorithms trained on historical violation data (e.g., past abuse cases) to identify patterns indicative of violations such as hateful conduct, abuse, child sexual exploitation, violent extremism, or spam. These models output confidence scores or violation probabilities, with parameters set to balance precision and recall—typically high-confidence thresholds trigger automatic actions (e.g., removal or restriction for known high-accuracy categories like CSE via hash-matching), while lower-confidence flags route to human review. Heuristics use predefined parameters like keyword/phrase lists, behavioral patterns (e.g., rapid posting, coordinated activity), or common violative structures for quick detection of emerging threats. Parameters are tuned via rigorous testing: models are trained/validated on thousands of labeled data points (violative/non-violative, including human-moderated cases), with inputs incorporating post text, attached media, user/account signals, and context. Before launch, automations undergo statistical sampling and item-by-item human review to confirm acceptable precision levels; post-launch, parameters are dynamically monitored for anomalies (e.g., spikes in volume, changes in appeal/overturn rates), with adjustments by engineering/data science teams. For high-priority illegal content (e.g., CSE/terrorism), parameters prioritize proactive, near-immediate action using tools like PhotoDNA hash-matching and keyword blocking to minimize exposure. LLMs are used to classify and annotate content on the platform as part of X’s Safety defence system.

Our approach to Safety focuses on effective and efficient risk mitigation utilising technology enhanced by human resources as appropriate. We use automation (combination of heuristics and models) to automatically detect content that we believe violates the X Rules and policies enforced on our platform. A vast majority of all accounts that are suspended for the promotion of terrorism and CSE are proactively flagged by a combination of technology and other purpose-built internal proprietary tools. X applies multiple safeguards to automated means (models, heuristics, and proprietary tools) to ensure reliability, minimize errors, and protect freedom of expression. These include rigorous pre-launch testing: models are trained/validated on thousands of human-labeled data points, with statistically significant test samples undergoing item-by-item human review by policy-expert reviewers to confirm acceptable precision levels before deployment. Content flagged by automated systems is routed to human review for lower-confidence cases or complex context, while only high-accuracy outputs enable automatic actioning. Post-launch, automations are dynamically monitored for performance anomalies, triggering adjustments. A robust appeals system allows users to challenge decisions quickly. Human oversight blends throughout, including proactive sweeps for high-priority risks (e.g., CSE/terrorism via hash-matching like PhotoDNA) and scaled investigations into evasion tactics.

X applies automated means - including machine learning models (natural language processing, image processing, and other sophisticated algorithms trained on historical violation data), heuristics (predefined patterns of text, keywords, behaviors for rapid response to emerging violations), and proprietary tools (e.g., hash-matching like PhotoDNA and keyword blocking) - precisely to proactively detect and enforce against violations of the X Rules and policies on its own initiative. The purposes are to automatically identify potentially violative content, flag it for human review or take direct action (e.g., removal, restriction), prioritize high-risk cases, and enable quick reactions to new violation forms, thereby scaling moderation while blending with human review. Risks addressed: X applies automated means primarily to address high-priority risks of dissemination of illegal and harmful content, including: - Dissemination of child sexual exploitation material. - Promotion of terrorism and violent extremism. - Hateful conduct, abuse & harassment, and violent content. - Sensitive/adult media and platform manipulation/spam.

Own-initiative moderation activities X’s 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. X engages in content moderation on its own initiative through a combination of automated detection (using machine learning models, natural language processing, image processing, and heuristics) and human review to proactively identify and act on violations of the X Rules and policies and reactively act on violations of the X Rules and policies following user reports. Types of restrictions applied Restrictions include: removal of content; reduction in visibility or reach; account suspensions or terminations; and geo-blocking in specific regions where required. Post-level enforcement We take action at the post level when a specific post violates the X Rules, including posts that share or reproduce other posts by posting screenshots, quote-posting, or sharing post URLs that violate our Rules. A few of the ways in which we might take action at the post level include: Limiting post visibility: Where appropriate, we will restrict the reach of posts that violate our policies and create a negative experience for other users by making the post less discoverable on X. This can include: Excluding the post from search results, trends, and recommended notifications Removing the post from the For you and Following timelines Restricting the post’s discoverability to the author’s profile Downranking the post in replies Restricting Likes, replies, Reposts, Quote posts, bookmarks, share, pin to profile, or Edit post Requiring post removal: When we determine that a post violated the X Rules and the violation is severe enough to warrant post removal, we will require the violator to remove it before they can post again. They will need to go through the process of removing the violating post or appealing our removal request if they believe we made an error. The post will be hidden from public view with a notice during this process. Account-level enforcement We take action to suspend an account if we determine that a user has engaged in repeated violations of our policies and/or violated specific policies that cause significant risk to X (i.e. posting illegal content, attempts to manipulate our platform or spam users, using our platform to incite violence, etc.) or pose a threat to our users (fraud, user privacy violations, violent threats, targeted harassment, etc.). Outline of soft moderation restrictions Soft measures include the attachment of labels, contextual warnings, or notes (e.g., sensitive media warnings for graphic or adult content, labels on synthetic/manipulated media for authenticity context, and warnings for potentially violative posts to provide additional user context without full removal): Sensitive media warnings for graphic or adult content: Under X’s Violent Content and Adult Content policies, users posting content depicting adult nudity, consensual sexual behavior and graphic violence (e.g., bodily fluids, serious injury, death depictions) have to label it as such. Content labeled as sensitive is placed behind content warnings or interstitials (e.g., "This media may contain sensitive material" or specific adult/graphic alerts). To view the content, users must acknowledge the content warning or interstitials and click through it. Users who regularly post sensitive content are required to adjust their media settings for automatic warnings. X will apply sensitive media labels to content depicting adult nudity, consensual sexual behavior and/or graphic violence which has not been proactively labeled by users. Content subject to sensitive media labels are age-restricted from users 18 or with no birth date blocked. Labels on synthetic/manipulated media for authenticity context: Under X’s Authenticity policy, X applies “Manipulated Media” or similar labels to media is shared as authentic but is significantly and deceptively altered, manipulated, or fabricated in a way that fundamentally changes its meaning and can result in widespread confusion on public issues, impact public safety, or cause serious harm, as well as to media that is not manipulated, but is shared in a deceptive manner or out-of-context or with intent to deceive people about the nature or origin of the content and can result in widespread confusion on public issues, impact public safety. Users can click through the “Manipulated Media” and similar labels to obtain more information or content, and sometimes links to trusted sources. The application of these labels is often applied simultaneously with visibility limits.

The entire team has access to online resources and regular onsite group and individual sessions related to resilience and well-being. These are provided by mental health professionals. External content reviewers also participate in resilience, self-care, and vicarious trauma training as part of our mandatory wellness plan during the reporting period.

Training and support of persons processing legal requests All team members, i.e. all employees hired by X as well as vendor partners working on these reports, 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 (QA) program, 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, by allowing content moderators to ask questions and discuss aspects of recently reviewed cases, X’s Rules and policies, and/or local laws. 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. QA is a critical measure to the business to help ensure that we are delivering a consistent service at the desired level of quality to our key stakeholders, both externally and internally as it pertains to our case work. We have a dedicated QA team within our vendor team to help us identify areas of opportunity for training and potential defect detection in our workflow or Rules and policies. The QA specialists perform quality checks of reports to ensure that content is actioned appropriately. The standards and procedures within the QA team ensure the team’s QA is assessed equally, objectively, efficiently and transparently. In case of any mis-alignments, additional training is scheduled, to ensure the team understands the issues and can handle reports accurately. In addition, given the nature and sensitivity of their work, the entire team has access to online resources and regular onsite group and individual sessions related to resilience and well-being. These are provided by mental health professionals. Content reviewers also participate in resilience, self-care, and vicarious trauma training as part of our mandatory wellness plan during the reporting period. Training and Support provided to those Persons performing Content Moderation Activities for our XIUC Terms of Service and Rules Training is a critical component of how X maintains the health and safety of the public conversation through enabling content moderators to accurately and efficiently moderate content posted on our platform. Training at X aims to improve the content moderators’ enforcement performance and quality scores by enhancing content moderators’ understanding and application of X Rules through robust training and quality programs and a continuous monitoring of quality scores. Training Process 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 content moderators must be trained in their assigned workflows. These focus areas ensure that content moderators are set up for success before and during the content moderation lifecycle, which includes: - Training analysis/design focused on agent and learning needs; - Classroom training with expert trainers; - Nesting period to apply new skills; - Cross-skilling opportunities; - Upskilling opportunities; - Refresher programs; - New launch/update roll-outs process; and - Remediation plans. Training Analysis and Design Before commencing design work on any content moderators program or resource, a rigorous learner analysis is conducted in close collaboration with training specialists and quality analysts to identify performance gaps and learning needs. Each program is designed with key stakeholder engagement and alignment. The design objective is to adhere to visual and learning design principles to maximise learning outcomes and ensure that agents can perform their tasks with accuracy and efficiency. This is achieved by making sure that the content is: - Easy to experience; - Easy to understand; and - Easy to apply. X’s training programs and resources are designed based on needs, and a variety of modalities are employed to diversify the content moderators learning experience, including: - Self-led learning: microlearning, scenario-based learning, e-learning modules, and gamification (where appropriate); - Virtual live instructor-led trainings; - Face-to-face classroom training; and - Videos. Classroom Training Classroom training is delivered either virtually or face-to-face by expert trainers. Classroom training activities can include: Instructor-led policy training; Interactive e-learnings; Scenario-based learning sets; Shadowing sessions with seasoned agents; Guided casework sessions with trainers; and Knowledge checks, quizzes and assessments. Onboarding and Ramp Up When content moderators successfully complete their classroom training program, they undergo an onboarding period. The onboarding phase includes case study by observation, demonstration and hands-on training on live cases. Onboarding activities include content moderator shadowing, guided case work, Question and Answer sessions with their trainer, coaching, feedback sessions, etc. Quality audits are conducted for each onboarding content moderator and content moderators 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 onboarding content moderator and prepare refresher training accordingly. After the onboarding period, content is evaluated on an ongoing basis with the QA team to identify gaps and address potential problem areas. There is a continuous feedback loop with quality analysts across the different workflows to identify challenges and opportunities to improve materials and address performance gaps. Up-Skilling When a content moderator needs to be upskilled they receive training of a specific workflow within the same pillar that the content moderator is currently working. The training includes a classroom training phase and onboarding phase which is specified above. Refresher Sessions Refresher sessions take place when a content moderator 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 content moderator needs and prepare a session targeting their needs and gaps. New Launch/Update Roll-Outs There are also processes that require new and/or specific product training and certification. These new launches and updates are identified by X and the knowledge is transferred to the content moderators. Remediation Plans There are remediation plans in place to support content moderators who do not pass the training or onboarding phase, or are not meeting quality requirements.