RTFP

WebGroup’s content moderation governance is structured through defined roles and review stages. The governance framework includes a Head of the Moderation Team (1 FTE), responsible for overall oversight, establishing policies and workflows, and handling complex cases requiring advanced expertise. Moderation activities are organised across several functions: basic review (20 FTE) serving as the initial review layer; advanced review (4 FTE) addressing complex cases, conducting own-initiative moderation, and contributing to procedural improvements; channel oversight (25 FTE) responsible for supervision of specific channels and related compliance issues; and a dedicated notices/complaints team (3 FTE) tasked with processing notices (including those from trusted flaggers), executing takedown actions, and handling complaints. Operationally, flagged content is processed through an internal interface or backlog (“review pool”), where items are distributed either automatically (including on the basis of skill sets) or manually. Moderators coordinate through group communication channels to share expertise and enhance decision-making. Quality governance is reinforced through periodic review of moderation decisions by the Head of the Moderation Team, promoting consistency and continuous improvement.

WebGroup’s ToS exposure rate is 0.00001%, and its illegal-content exposure rate is 0.00002%. These figures are intended to provide a concise and comprehensible indication of the limited level of recipient exposure to violating and illegal material, consistent with WebGroup’s proactive own-initiative detection measures, rapid temporary unavailability actions in high-priority cases, and human verification of automated flags.

WebGroup quantifies the human resources dedicated to content moderation using a full-time equivalent (FTE) methodology, where 1 FTE corresponds to one employee primarily responsible for implementing activities related to the moderation process. Responsibilities may overlap across moderation stages (for example, the Head of the Moderation Team may also participate in advanced review or support notices/complaints processing, and advanced reviewers may occasionally perform basic review), which can make a strict allocation of headcount by stage difficult.

Moderators are required to have advanced English proficiency and strong PC skills. Formal educational qualifications are not presented as the determining criterion; instead, suitability is primarily experience-based, reflecting time spent monitoring content and developing the judgement and sensitivity required in adult-content environments. The moderation function is supported by language coverage across EU languages (with additional language coverage within channel oversight), the use of translation tools where direct coverage is not available, and keyword databases incorporating linguistic variations across EU languages.

WebGroup describes indicators of accuracy and possible error rates for its automated moderation tools on a tool-by-tool basis and clarifies that performance may vary depending on the type of content moderated and the part of the online interface scanned. For Vercury, accuracy for identifying blocked content is 83%, and initial accuracy for flagged-only content was 2.5%, improving to up to 10% since June 2024. An estimated 15–20% error rate applies to blocked content pending review. Vercury scans videos only. For Hive, accuracy is not currently quantifiable because it is used primarily as an aid in manual review. However, approximately 10% of “highly problematic” Hive outputs are confirmed as correct upon review, while “problematic” outputs are associated with an approximate 90% error rate, reflecting the need for contextual assessment. Hive scans videos and images (pictures). For Google SafetyNet API, accuracy varies depending on threshold settings. For videos, the Google High threshold shows 100% accuracy and Google Medium 1.6% accuracy. For pictures, Google Very High shows 100% accuracy, High 75% accuracy, and Google Medium 15% accuracy. Estimated error rates are up to 98.5% for videos and 70% for pictures. The tool scans videos and images. For Safer, the accuracy rate is 10%, with a possible error rate of 90%. Safer scans videos only. For keyword search and match, data on accuracy and error rates are not available. Its scope is limited to text-based content (comments).

Regarding the qualitative description of automated means, WebGroup uses automated tools to identify, review, and manage content that may violate the ToS or legal rules across videos, images, and text, supporting real-time monitoring, prioritisation, and preventative measures. In particular, Vercury operates as a fingerprinting system relying on an internal signature database: it compares video and image signatures against that database, generates a match score, and, based on that score, content may be blocked or sent for review, primarily targeting illegal materials; Vercury scans videos only. Hive uses AI to analyse images and videos for behaviour, context, or potentially harmful activity within a scene. It is mainly used to identify ToS-violative content (e.g., alcohol, violence, firearms), and outputs may lead to blocking or flagging for review; Hive scans videos and images (pictures). Google SafetyNet API uses AI to analyse images and assign an underage risk score based on its assessment of the content. It is used to flag potentially harmful material for review and is parameterised through qualitative thresholds (e.g., Google Very High / High / Medium for pictures and Google High / Medium for videos). The tool scans videos and images. Safer maintains a fingerprint database of known content to detect and flag matches efficiently and is used to flag potentially harmful or inappropriate material for review; Safer scans videos only. For text, XVideos applies keyword search and match using block lists and grey lists: block-list matches prevent publication (with such content not being saved), while grey-list matches are flagged for further analysis and human review; this tool scans text only (comments). In addition, WebGroup uses Offlimit, which maintains a fingerprint database of known CSAM content to detect and flag matches efficiently. This tool description is included here as supplemental information provided directly by XVideos alongside the descriptions of the other automated tools.

WebGroup applies safeguards designed to reduce the risk of incorrect outcomes and to ensure appropriate human oversight, recognising that automated tools can generate false positives. Content flagged by Vercury fingerprinting, Hive AI analysis, Google SafetyNet API, and Safer matching is subject to manual review. For keyword-based tools, safeguards distinguish between block lists (where the user is shown an error message and the data is not retained) and grey lists (where content is flagged and reviewed). Most automated tools are third-party solutions, and WebGroup therefore has limited influence over their underlying development. However, it enhances its proprietary systems based on moderator feedback and regularly updates keyword databases to address evolving jargon and abusive user practices.

WebGroup uses automated tools to support the detection, prioritisation, and processing of potentially illegal or incompatible content and to enhance moderation efficiency. Automated means are used to: - detect exact duplicates of known files through flag matches for review; - fingerprint video and image content (Vercury) and compare signatures against an internal database to generate match scores that can result in blocking or referral for review, primarily in relation to illegal materials; - apply AI-based analysis (Hive) to identify behaviours or contexts linked to potentially harmful activity and ToS violations (e.g., alcohol, violence, firearms) and to support blocking or flagging for review; - apply AI-based image analysis through Google SafetyNet API to assign a risk score (notably for underage-related risks) and flag content for review; (v) to match content against fingerprint databases (Safer) and flag potential matches for review; and - use keyword search and matching (block lists and grey lists) for text moderation, where block lists prevent publication of specific terms and grey lists trigger additional review.

WebGroup Czech Republic, a.s. ("WebGroup") carries out content moderation proactively at its own initiative to maintain a safe environment and ensure compliance with applicable rules. Own-initiative moderation covers multiple content types, including text content (such as comments, wall posts and “about me” sections), images (including gallery and profile images, reviewed on a sampling basis), and videos (subject to rigorous checks to prevent dissemination of illegal or incompatible material). This proactive approach relies on a combination of automated detection mechanisms (including AI functionalities) and human oversight, with automated tools used to detect and flag potentially illegal or incompatible content and moderators performing contextual assessments and final decisions. For high-priority suspected illegal content (including CSAM, non-consensual intimate imagery and gender-based violence), the platform typically immediately unindexes the relevant URL, making it unavailable while the dedicated notice/complaints team reviews the case, and the content remains inaccessible until the notice is resolved. For efficiency, WebGroup also uses an internal tool supporting review of reported videos (e.g., suspected NCII or CSAM) by extracting screenshots at intervals, enabling quicker assessments. Depending on the outcome of own-initiative moderation, WebGroup applies restrictions affecting the availability, visibility, and accessibility of content and accounts, including: (i) content ghosting (unindexing) as a temporary visibility restriction; (ii) content takedown as “pending deletion”; (iii) permanent deletion of content; and, where applicable, account restrictions, including account termination (temporary inaccessibility) and account deletion (permanent removal).

WebGroup provides several support measures for moderators. These include a Q&A reference table addressing recurring questions, as well as dedicated group chat and communication tools that enable real-time consultation and the sharing of best practices. Moderator well-being is supported through access to ergonomic workstations and subsidised fitness or wellness facilities. Mental well-being measures include fostering an open culture that encourages discussion of challenges, offering confidential consultation options, providing flexibility in working arrangements (including, where appropriate, flexible hours, shift flexibility, and remote work), and ensuring regular breaks.

Training is ongoing and tailored to both general and specific content moderation challenges. It is delivered by the Head of the Moderation Team together with an experienced moderator. New moderators follow a mentorship model, beginning with observation of experienced colleagues and gradually assuming responsibilities under supervision, with continuous feedback throughout the process. Support remains available after moderators begin working independently. Training emphasises the responsible handling of explicit content, the ability to identify and appropriately process illegal material, and the consistent application of internal moderation procedures.