RTFP

Good corporate governance is critical to our approach to content moderation. The foundation for managing this is a clear internal structure, together with accompanying processes and tools to manage the needs of Google's diverse products and services with consistency and appropriate flexibility. Our governance structures include: - Board governance: The Board of Directors of Google Ireland Limited is responsible for overseeing the management of systemic risks. Risk management and mitigation issues are regularly reported to and considered by the Board. At least once a year, the Board approves the strategies and policies for taking up, managing, monitoring and mitigating the risks identified. - Executive oversight: Senior management has direct oversight of content risks, ensures appropriate resourcing for dedicated teams, reviews significant incidents or escalations, and reports these items to the Board as necessary. - Google Trust and Safety: We pioneered the now industry-wide practice of investing in Trust and Safety specialists who are trained to analyse bad actors, abusive practices, content issues, and the effectiveness of existing policies. Today, our Trust and Safety teams consist of experts, specialists, and engineers working to keep people safe online by using the latest technology (including AI and LLMs) to enforce our policies and moderate content. These teams partner with external experts and teams across Google to carry out our mission to keep people safe online and protect our services and products from abuse. These governance structures work to identify and address the risk to users across the product lifecycle: from product development (where we embed user safety into the design of our products) to product launch, followed by ongoing monitoring and evaluation. This is enabled by internal policies, guidance, regular training, clear escalation paths, and processes designed to monitor emerging trends and address new harm vectors before they can become larger issues).

While qualitative information about Article 23 suspensions imposed to protect against misuse is not specifically requested in the DSA Template, we are providing it here to aid interpretation of relevant metrics provided on the 7_appeals_and_recidivism and the 7_appeals_and_recidivism_Ads tabs. Article 23 Suspensions imposed to protect against misuse Requirements for Article 23 suspensions apply to Online Platforms but not Hosting Service Providers and Intermediary Service Providers more generally. To protect users from significant harm and unlawful activity, Google suspends user accounts when we detect egregious content (e.g., child abuse) or repeated violations of our services' policies. Suspended user accounts are unable to access Google products and, depending on the suspension reason, may not contribute to Google services or sometimes engage in specific Google processes (e.g., submission of complaints through dedicated complaint channels). Depending on the severity of the detected violation and involvement of legal enforcement authorities, users may receive a warning and/or remedial instructions to remove the violating content before their account is suspended. Users who intentionally misuse webforms and processes by repeatedly filing manifestly unfounded legal notices or manifestly unfounded complaints based on legal notices will be flagged, and their requests will be closed without assessment. Users will typically receive a written warning before Google takes action. If misuse continues, the user will be suspended from reporting content for a period of up to six months and their requests will be closed without assessment. After a maximum of six months, new requests for content removal may be submitted. The Multi-Services report includes the number of Google-wide account-level suspensions of EU users who posted manifestly illegal content across Google services. These Google-wide suspensions were not necessarily limited to individual VLOPs and Online Platforms and therefore are not included in service-level reports. In addition, the number of suspensions of users who repeatedly submitted manifestly unfounded notices and manifestly unfounded legal complaints during the reporting period are provided. If applied, these would suspend the processing of a user's notices for any Google service within the relevant operation, therefore they are also not linked to a specific Google service. In many cases, specific Google services do not permit users to submit repeat complaints, therefore suspension under Article 23 is not an applicable outcome for those specific services. If repeat complaints are permitted, metrics are provided in the service report. For relevant services, Ads specific numbers reflect the number of complaints that have been suspended or not processed during the reporting period. For more information, see: - Misuse Policy: https://support.google.com/legal-help-center/answer/13949470

While qualitative information about Art 9 and 10 orders from Member States' authorities is not specifically requested in the DSA Template, we are providing it here to aid interpretation of metrics provided on the 3_member_states_orders tab. Article 9 and 10 Orders from Member States' authorities Courts and government agencies in the EU regularly request that we remove information from Google services (Removal Orders). These requests are routed to the appropriate team(s) within Google who review these requests closely to determine if information should be removed because it may violate a law or our product policies. In addition, specific EU and Member State laws allow government agencies in the EU to request user information for civil, administrative, criminal, and national security purposes (User Data Disclosure Orders). Each request is carefully reviewed to make sure it satisfies applicable laws; this review and any subsequent enforcement requires human involvement. Metrics relating to Removal Orders and User Data Disclosure Orders received during the reporting period, conforming to the requirements of Article 9 and Article 10 of the DSA respectively and pertaining to the applicable service are provided in this report. In addition, metrics around orders received relating to advertisements, including where those advertisements may have appeared on one or more Google services, are provided in both the Ads and the Multi-Services report. Where relevant, the median time needed to confirm receipt and the median time needed to take action (or non-action) in response to Removal Orders and User Data Disclosure Orders is provided in hours. Information about other requests from government authorities around the world, including requests related to illegal content or user information that are not made pursuant to Article 9 or 10, are published in our Government Requests for Content Removal Transparency Report and our Government Requests for User Information Transparency Report. Information in these reports is voluntarily provided and not necessarily directly comparable with information presented in this mandated DSA report, due to differences in methodologies. For more information, see: - Government Requests for Content Removal Transparency Report: https://transparencyreport.google.com/government-removals/overview?hl=en - Government Requests for User Information Transparency Report: https://transparencyreport.google.com/user-data/overview?hl=en

While qualitative information about complaints received through out-of-court dispute settlements is not specifically requested in the DSA Template, we are providing it here to aid interpretation of relevant metrics provided on the 7_appeals_and_recividism tab. Out-of-court dispute settlements Out-of-court dispute settlement (ODS) bodies are independent bodies, certified by an EU Member State, to handle complaints referred by EU users. Further information about Google's approach to this requirement is available on the EU Out-of-Court Dispute Resolution Help Center. Google reviews all eligible complaints submitted by ODS bodies. Duplicative complaints are not accounted for in the metrics. After reviewing a case, Google communicates the outcome of its review to the relevant ODS body which then issues a verdict. The 'Decisions upheld', 'Decisions partially reversed' and 'Decisions reversed' verdicts only reflect cases where a decision has been achieved on the merits of the dispute and cases where the ODS body has issued a 'default' decision because it accepted review of a complaint over which it has no legal authority to review. The 'Decisions partially reversed' category may include ODS disputes which contain multiple items submitted by the same user and therefore, led to different decisions on each item by the ODS body. The 'Decision omitted' category includes ODS disputes which were filed but did not lead to a decision by the ODS body. This includes situations where the dispute was canceled / withdrawn unilaterally or by agreement with the complainant, or which require no action. The median time (in hours) to complete the dispute settlement procedure is calculated from when the ODS body receives a complaint to when the ODS body issues the final verdict to Google. It is calculated based on the 'Decisions upheld', 'Decisions partially reversed' and 'Decisions reversed' verdicts only. The percentage of verdicts implemented reflects cases where an adverse verdict was issued by the ODS body and where Google implemented a decision, consistent with the verdict. Metrics relating to out-of-court dispute settlements that apply across multiple services are provided in the Multi-Services report. For more information, see: - EU Out-of-Court Dispute Resolution Help Center: https://support.google.com/european-union-digital-services-act-redress-options/answer/13535501?hl=en

While qualitative information about complaints received through the internal complaint handling systems (i.e. appeals) is not specifically requested in the DSA Template, we are providing it here to aid interpretation of relevant metrics provided on the 7_appeals_and_recidivism and the 7_appeals_and_recidivism_Ads tabs. Complaints received through internal complaint handling systems (i.e., appeals) Google works hard to maintain services that are safe and vibrant. As with any system, we sometimes make mistakes, which may result in the unwarranted removal of content from or access to our services. To address that risk, where appropriate, we make it clear to users and/or creators that we have taken action on their content and provide them the opportunity to contest that decision through designated complaint-handling systems and give us clarifications. In addition, under the DSA, EU users can submit complaints about an action that Google did not take in response to a notice/flag that they previously submitted. Complaint outcomes may include decision upheld, decision reversed, decision partially reversed and decision omitted. A decision refers to the initial enforcement of Google's terms of service or product policies. These decisions may be reversed in light of additional information provided by the appellant or additional review of the content. A partially reversed decision may include complaints which contain multiple items submitted by the same user and therefore, led to different decisions on each item by the service. If a complaint is withdrawn, if the complaint requires no action, response or decision from Google, or if the creator resolves the issue so that their content is no longer policy-violating, this is categorised as 'Decision Omitted'. Within the 'Decision omitted' category, there are some cases related to removals due to copyright law where Google services act as a neutral intermediary between the claimant and the uploader, who may choose to pursue resolution in court. Google works to provide complaint outcomes to users within a reasonable timeframe. The types of complaints vary widely, with some requiring a longer review period due to varying degrees of complexity or external factors (e.g., legally prescribed wait times). Where relevant, median times are reported in hours. In addition, not all complaints can be resolved during the reporting period and some resolved complaints may have been received prior to the reporting period, therefore the total number of complaint outcomes will not necessarily equal the total number of complaints received. For relevant services, we separately present metrics for complaints received relating to advertisements impressed on those services. However, complaints about actions taken in response to Article 16 notices that relate to advertisements (including where those advertisements may appear on one or more Google services), are provided in both the Ads and the Multi-Services report.

While qualitative information about notices received through the notice and action mechanism is not specifically requested in the DSA Template, we are providing it here to aid interpretation of metrics provided on the 4_notices tab. Notices received through notice and action mechanisms Google's content and product policies apply wherever you are in the world, but we also have processes in place to remove or restrict access to content based on local laws. Users, Trusted Flaggers (as defined by Article 22), and other entities can report content that they believe should be removed from Google's services under applicable laws. Action is taken on content that is deemed to violate applicable laws or Google policies. In the European Union, national entities called Digital Services Coordinators may award Trusted Flagger status to entities tasked with flagging allegedly illegal content on online platforms. Trusted Flaggers are likely to have expertise in one or more fields relevant to content moderation, such as privacy or child safety. The European Commission maintains a list of designated Trusted Flaggers in a publicly accessible database. Legal standards vary greatly by country/region. Content that violates a specific law in one country/region may be legal in others. Typically, Google removes or restricts access to content only in the country/region where it is deemed to be illegal. However, when content is found to violate Google's content or product policies or Terms of Service, Google may remove or restrict access globally. When a notice is reviewed and the content violates our content policies, action may be taken on policy grounds. If the content does not violate our policies, Google may take action on legal grounds, in line with local laws. As a notice may contain one or more URLs for review, multiple actions may be taken as a result of a single notice received. In some cases, no action is taken in response to a notice. The median time to take action calculation includes notices where an action was taken and notices where it was decided not to take action (with median reported in hours). Article 16 notices are processed by automated means for some, but not all, Google services. While intermediary services are out-of scope for Art 15(1)b, actions taken on a legal basis in response to legal notices received about intermediary service content are reported under Art 15(1)c for completeness. In addition, metrics around notices received relating to advertisements, including where those advertisements may have appeared on one or more Google services, are provided in both the Ads (hosting service) report and the Multi-Services report. For more information, see: - Trusted Flaggers under the DSA: https://digital-strategy.ec.europa.eu/en/policies/trusted-flaggers-under-dsa

There is a dedicated page on the Google Transparency Center to help users find ways to report harmful content on several of our services. Often the option to report will be found in the service itself or with its policies. Depending on the nature of the service and the potential policy violation, these user reports may be reviewed using automated tools, human review or a combination of both. To enforce our policies at scale, Google relies on a combination of automated and human tools to spot problematic content. While automated systems can quickly identify and take action against spam and some violative content, human judgement is needed for the many decisions that require a more nuanced determination. The context in which a piece of content is created or shared is an important factor in any assessment about its quality or its purpose. Google is attentive to educational, scientific, artistic, and documentary contexts, including journalistic intent, where the content might otherwise violate our policies. Google escalates particularly complex cases to specially-trained experts. Additionally, Google uses a corpus of human-reviewed and removed content to train machine learning technology to flag new content that may also violate product policies. Using machine learning technology trained by human decisions enables our enforcement systems to adapt and become more effective over time. More information about how Google services use automated tools, often supplemented with human review, for content moderation, along with the indicators of accuracy of any fully automated tools is included in the sections below. For more information, see: - Google Transparency Center: https://transparency.google/intl/en_uk/tools-programs/reporting-and-appeals/ Service-specific details End users on the service do not view a sizable proportion of illegal or incompatible content on the service in comparison to the overall volume of content on the service within the reporting period.

Computing the human resources who evaluate content across Google services is a highly complex process. Content can be posted by users globally or reviewed by content moderators located globally. Content moderators may review content for multiple policy violations or focus on one specific topic; they may review content that appears across one or more services; and content assigned for their review may have been posted in several different languages. In some cases and where appropriate, translation tools may be used to assist in the review process and allow us to moderate content 24/7 and at scale. Accordingly, these metrics do not necessarily reflect the language that the content was ultimately reviewed in. The number of internal moderators reflects individuals employed by Google who completed at least five content reviews during the reporting period and were available to review content on Google Shopping. All internal moderators are employed with English language expertise and are counted as such. The number of external moderators contracted by Google reflects individuals who were available to conduct reviews on Google Shopping as of 31 December 2025. All external moderators were contracted to moderate content in a specific EU Member State Language or to moderate non-language content. The total number of moderators with sufficient linguistic expertise reflects the sum of the internal moderators and external moderators contracted to moderate content in an EU Member State Language. External moderators who were contracted to review non-language content (e.g., an image) are excluded from this count. The total number of moderators with sufficient linguistic expertise is also broken down by EU Member State Language. External moderators who review non-language content (e.g., an image) are included in the 'Agnostic' category. Note that some content moderators are available to review content that appears across multiple Google services (including Google Maps, Google Play, and Google Shopping), therefore these moderators are counted under each of Google Maps, Google Play and Google Shopping.

Metrics presented in this report generally reflect our efforts and resources to moderate potentially illegal content and policy-violative content in the EU. Where that is not possible and we are reporting more broadly than the EU, we are looking to remediate this for future reports. In some cases, EEA metrics have been voluntarily provided. Metrics presented in this report generally reflect our efforts to moderate content throughout the full reporting period; in cases where that is not possible, we are looking to remediate this for future reports. Numbers reported may fluctuate between successive reports due to various reasons, including service-level changes or enhancements, changes in the number of users on a service, external events and differences in reporting periods. Therefore, report-by-report comparisons may not accurately reflect time-based improvements in our processes. Services differ in various ways, including content type on the service, underlying content moderation systems and number of users on a service, which means that in some cases, metrics may not be directly comparable. In accordance with EC guidance, all indicators reporting a percentage are reported as floating numbers in the [0,1] interval. Any value of 0.99995 or greater is rounded up to 1. In addition, all median times are reported in hours. Metrics provided in the report reflect details of the orders, notices, appeals and ODS cases received during the reporting period, and the actions taken during the reporting period. In general, metrics do not reflect details of possible outcomes that did not materialize during the reporting period. For example, where a category and sub-category of illegal/incompatible content applies to a service, but no orders or notices were received, or no actions were taken during the reporting period, these fields may be blank; or where a restrictive action is applicable to a service, but no actions were taken during the reporting period, these fields may be blank.

Human reviewers or content moderators play a key role in content moderation at Google. Technology has become very helpful in identifying some kinds of problematic content (e.g., finding objects and patterns quickly and at scale in images, video, and audio), while humans are able to apply a more nuanced approach to assessing content. To safeguard against content actions that could potentially contribute to or exacerbate adverse impacts due to allowing or removing content, Google utilises international human rights standards to guide policy and enforcement decision-making, considering how content could adversely impact the rights of an individual, community, or society as a whole, or further the understanding of social, political, cultural, civic, and economic affairs. As an example of public interest-informed content moderation, Google carves out exceptions to enforcement guidelines for material that is Educational, Documentary, Scientific, and/or Artistic (EDSA). Content that falls under those exceptions are crucial to understanding the world and to chronicling history, whether it is documenting wars and revolutions or artistic expression that may include nudity. Consequently, Google takes great care in helping reviewers understand the EDSA exceptions when reviewing flagged content. Google strives to create workplaces and economic opportunities that work for employees, as well as vendors, temporary staff, and independent contractors. While Google does not employ all of the individuals who contribute to content moderation, Google is committed to ensuring that work on Google products is conducted in environments that treat all workers with respect and dignity, ensure safe working conditions, and conduct responsible, ethical operations. For that reason, Google seeks out suppliers that embrace its values, commitment to human rights, and that support a safe working environment. Suppliers must operate in accordance with our Wellness Standards for Sensitive Content Moderation and Supplier Code of Conduct, and comply with all applicable labour protection laws, including those related to privacy, safety, health, and wages. Google also provides a framework of wellness standards that promote healthy working conditions and resources for provisioned extended workforce members and Google employees performing sensitive content moderation. Adherence to our Wellness standards is verified through periodic audits. These guidelines encompass the entire worker lifecycle, including pre-hire and onboarding — where specific requirements apply to job descriptions, interviews, and training for both managers and staff — as well as ongoing engagement and support during and after exit. Ongoing support features dedicated wellness teams, counseling, employee assistance programs, peer support groups, and wellness breaks. Additionally, suppliers must provide physical wellness spaces, resources, and opt-out options, with wellness support in the form of counseling remaining accessible to content moderators for at least six (6) months after exit. Qualifications and linguistic expertise Qualifications for Google employees who work on sensitive content may include role related knowledge in the content matter, professional experience in content moderation or sensitive workflows, linguistic expertise, and computer proficiency. The linguistic expertise required varies depending on the specific workflow of a product or service, the type of content, and languages that content is available in. Some products or services require native proficiency in global supported languages, others may use translation tools, and some videos or images do not require any language proficiency to review. Some Google employees who work on sensitive content are also subject matter specialists skilled in specialty areas, such as child sexual abuse material or violent extremism. For more information, please see: - Google's Supplier Responsibility Program: https://sustainability.google/progress/supplier-responsibility/ - Google's values: https://about.google/company-info/commitments/ - Google's commitment to human rights: https://about.google/company-info/human-rights/ - Google Supplier Code of Conduct: https://about.google/company-info/supplier-code-of-conduct/

Where relevant and feasible, indicators of accuracy for automated tools used to process legal-related content removal requests, CSAM and advertisements are provided in relevant service-specific reports. While we report fully automated tools primarily on a language-agnostic basis, where applicable and feasible for this reporting period, the indicators of accuracy are also broken down by language. Unless otherwise stated, metrics are calculated using standard statistics definitions provided below: - Accuracy, defined as the proportion of all decisions that were correctly made by the automated system = (TP + TN) / (TP + TN + FP + FN) - Precision, defined as the proportion of all identified positive cases that were correctly classified = TP / (TP + FP) - Recall, defined as the proportion of actual positive cases that were correctly classified = TP / (TP + FN) Where TP = True Positives; TN = True Negatives; FP = False Positives; FN = False Negatives. Automated tools used to process Legal-related Content Removal Requests For relevant services, accuracy, precision and recall are provided for fully automated content moderation tools used to process legal-related content removal requests. These metrics are computed based on a quality assurance audit of global reviews performed by the automated systems. As previously mentioned, intermediary services are out-of scope for Art 15(1)b and actions taken on a legal basis in response to legal notices received about intermediary service content are reported under Art 15(1)c for completeness. Therefore, accuracy indicators of fully automated content moderation tools used to process legal-related content removal requests on intermediary services are reported under 'own initiative' rather than 'NAM total' on the 8_automated_means tab. Automated tools used to combat Child Sexual Abuse Materials (CSAM) For relevant services, precision is provided for fully automated content moderation tools used to restrict CSAM. False Positives reflect the number of automated actions that were later found to be incorrect and reversed following appeals from account owners in the EU while True Positives reflect the number of automated actions that were reviewed and upheld after appeal and those that were not subject to an appeal. Accuracy and recall metrics are not being reported due to infeasibility of identifying False Negatives (i.e. where violative content exists but it is not detected by automated means). Language breakdown is not applicable as moderated content is image-based. Automated tools that affect advertisements Within the Ads (hosting service provider) report, accuracy, precision and recall is reported across all automated content moderation decisions. For relevant services, Ads is reporting accuracy, precision and recall across all automated content moderation decisions on that particular service (on the 8_automated_means_Ads tab). False Positives reflect the number of automated actions that were later found to be incorrect and reversed following appeals from advertisers in the EU while True Positives reflect the number of automated actions that were reviewed and upheld after appeal and those that were not subject to an appeal. False Negatives reflect instances where an automated system did not detect a violation, but action was subsequently taken following reports from EU users while True Negatives reflect instances where content was reviewed following EU user reports and found to be non-violative, in addition to instances where content was approved by the automated system and not subsequently reviewed. To determine if the automated system was right or wrong, Ads checks if automated decisions were overturned by human reviewers within 90 days following the ad's creation. In the VLOP reports, accuracy, precision and recall metrics are provided for each EU Member State language where sufficient data exist to calculate meaningful metrics. Service-specific automated tools Google Shopping is reporting accuracy, precision and recall across all automated content moderation decisions. These metrics are calculated using standard statistics definitions provided above. False Positives reflect the number of automated actions that were later found to be incorrect and reversed following appeals from content owners in the EU while True Positives reflect the number of automated actions that were reviewed and upheld after appeal and those that were not subject to an appeal. False Negatives reflect instances where an automated system did not detect a violation, but action was subsequently taken following reports from EU users while True Negatives reflect instances where content was reviewed and found to be non-violative following EU user reports, in addition to instances where content was approved by the automated system and not subsequently reviewed. Language breakdowns for these metrics are unavailable for this service, however country-level breakdowns are provided as an alternative.

This section describes how Google uses automated tools, often supplemented with human review, for content moderation. Automated tools used to process Legal-related Content Removal Requests Automation plays a role in legal content moderation to help Google work at scale, and focus our efforts on actionable, authentic requests. There are a few ways that automation might be used while handling a removal request. The most common way is that Google uses automation to route a request to the right team. Google has subject matter experts in different types of content and languages, and using automation ensures the request is sent to the people best positioned to review it. Once content removal requests are routed efficiently, Google also uses automation to manage the millions of URLs (web page addresses) that are sent to Google for review every day, and to complement and streamline human review. As an example, Google receives a significant number of Google Search removal requests for URLs that are not included in Google's search index, which is the vast and continuously updated pool of web page addresses from which all search results are drawn. We have automated systems that detect such URLs in removal requests, enabling our teams and processes to focus on content that does appear on our services and address complex matters requiring human review. Google also uses automation to process some legal notices. The vast majority of notices are copyright removal requests, largely from submitters with a well-established track record of submitting valid requests, allowing Google to be relatively confident in automating this processing. Certain requests to remove allegedly defamatory Local Reviews are also automatically processed. Automated tools used to combat Child Sexual Abuse Material (CSAM) Google takes its responsibility to fight child sexual abuse and exploitation online very seriously. We do this by combatting CSAM across Google's services and by detecting instances of abuse and enforcing robust policies. We also partner with non-governmental organisations (NGOs) and others in industry coalitions to share proprietary technology and drive the industry forward. Built-in protections help prevent Google services from showing abusive content and deter bad actors. To detect and report CSAM, we may use a combination of cutting-edge technology, including machine learning classifiers (to identify unknown CSAM) and hash-matching technology, as well as trained specialist teams. Hash-matching technology creates a 'hash', or unique digital fingerprint, for an image or a video so it can be compared with hashes of known CSAM. When Google finds CSAM, our services remove it, report it to the National Center for Missing and Exploited Children (NCMEC), and take action, which may include disabling the account. Google scales its impact by collaborating with NCMEC and partnering with NGOs and industry coalitions to help grow and contribute to a joint understanding of the evolving nature of child sexual abuse and exploitation. One of the ways Google contributes is by creating and sharing free tools to help other organisations prioritise potential CSAM images for human review. For example, Google's Child Safety Toolkit consists of two APIs. The first is Child Sexual Abuse Imagery (CSAI) Match, an API developed by YouTube that partners can use to automatically detect known videos of CSAM so they can flag for review, confirm, report, and act on it. The second is Google's Content Safety API that helps partners classify and prioritise novel potentially abusive images and videos for review. Detection of never-before-seen CSAM helps the child safety ecosystem by identifying child victims in need of safeguarding and contributing to the list of known digital fingerprints to grow our abilities to detect known CSAM. Google takes action not just on illegal CSAM, but also wider content that promotes the sexual abuse and exploitation of children and can put children at risk. For more information, see: - How image hashing technology helps NCMEC: https://safety.google/stories/hash-matching-to-help-ncmec/ - Tools to fight CSAM: https://protectingchildren.google/#tools-to-fight-csam Automated tools that affect advertisements Google uses a combination of automated and human evaluation to detect and remove ads which violate our policies and are harmful to users and the overall ecosystem. Advertisements can also appear across multiple services. To keep ads safe and appropriate for everyone, ads are reviewed to make sure they comply with Google Ads policies and Google Shopping Ads policies. When reviewing ad content or advertiser accounts to determine whether they violate our policies, Google takes various information into consideration, including the content of the creative (e.g., ad text, keywords, and any images and video) and the associated ad destination. Google also considers account information (e.g., past history of policy violations) and other information provided through reporting mechanisms (where applicable) in our investigation. For more information, see: - Google Ads policies: https://support.google.com/adspolicy/answer/6008942 - Google Shopping Ads policies: https://support.google.com/merchants/answer/6149970?hl=en Service-specific automated tools Products and merchants go through in-depth safety reviews before they can list on Google. Thanks to features such as the Shopping Graph (Google Shopping's data set of the world's products and sellers), Google Shopping's systems can quickly review whether a business is legitimate, and whether the products and other content follow Google Shopping's policies. This automated vetting process has helped to more efficiently and accurately review a massive amount of products. Shopping's automated systems are always monitoring for violating activity. Some examples of automated content moderation processes used include: - policy checks for harmful, regulated, or illegal content (e.g., weapons, recreational and prescription drugs, tobacco products); - product image checks for policy violations such as graphic overlays or nudity; - product data quality checks; - landing page checks; and - checks for recalled products such as those listed in the Rapid Exchange of Information System (RAPEX) or Organisation for Economic Co-operation and Development (OECD) public databases. For more information, see: - Google Shopping Graph: https://blog.google/products/shopping/shopping-graph-explained/ - Google Shopping policies: https://support.google.com/merchants/topic/7286989?hl=en&ref_topic=7259123&sjid=8355886752592660758-NC

To enforce our policies at scale, Google relies on a combination of automated and human tools to spot problematic content. While automated systems can quickly identify and take action against spam and some violative content, human judgement is needed for the many decisions that require a more nuanced determination. The context in which a piece of content is created or shared is an important factor in any assessment about its quality or its purpose. Google is attentive to educational, scientific, artistic, and documentary contexts, including journalistic intent, where the content might otherwise violate our policies. Google escalates particularly complex cases to specially-trained experts. Additionally, Google uses a corpus of human-reviewed and removed content to train machine learning technology to flag new content that may also violate product policies. Using machine learning technology trained by human decisions enables our enforcement systems to adapt and become more effective over time. Automated tools used to process Legal-related Content Removal Requests Users can appeal content moderation decisions made by automated systems to process legal-related content removal requests. These appeals are then subject to human review. Automated tools used to combat Child Sexual Abuse Materials (CSAM) Users can appeal content moderation decisions made by automated systems to take action on CSAM. These appeals are then subject to human review. Sampling is also used to constantly monitor the quality of automated enforcements. Automated tools that affect advertisements Our enforcement technologies may use automated evaluation, modelled on human reviewers' decisions, to help protect our users and keep our ad platforms safe. The policy-violating content is either removed by automated means or it is flagged for further review by trained operators and analysts who conduct content evaluations that might be difficult for algorithms to perform alone, for example, because an understanding of the context of the ad is required. The results of these manual reviews are then used to help build training data to further improve our machine learning models. Ads also offers users and Priority Flaggers an option to report violative ads, while advertisers have the option to appeal automated ad restrictions. All user reports and appeals are subject to human review. Service-specific automated tools We use a combination of automated and human evaluation to ensure that product listings comply with our policies. Our enforcement technologies use algorithms and machine learning, modelled on human reviewers' decisions, to help protect our customers and keep our platforms safe. Automation has helped us more efficiently and accurately review a significant amount of products. More complex, nuanced or severe cases are often reviewed and evaluated by our specially-trained experts. Google Shopping offers users the ability to flag violative content and provides content owners with an option to appeal fully automated decisions. All user reports and appeals are subject to human review.

See response to 'Qualitative description of the automated means'.

Across all products and services, we set clear policies for what is and is not acceptable on our platforms. These policies aim to ensure a safe and positive experience for our users and observe a high standard of quality and reliability for advertisers, publishers, and content creators alike. Content policies establish the rules of the road for what content can be created, uploaded, sent, shared, and monetised. These policies are used to guide content moderation and enforcement actions on our products. They also play an important role in maintaining a positive experience for everyone on our platforms no matter where they are in the world. User data and developer policies provide rules for how developers interact with our products and services. They also describe the privacy and security requirements for handling user data to include the full spectrum of developer actions, like requesting, obtaining, using, and sharing data. Monetised product guidelines are the policies and standards related to products Google earns revenue from and cover what can or cannot be monetised. These policies empower and protect users while promoting a thriving digital ecosystem that is safe and conducive to innovation and growth. 'Own initiative' content moderation Content moderation actions taken at Google's 'own initiative' are considered to be actions taken on content shown to or flagged by those in the EU because the content violates our policies, or where the content is illegal but action is not taken in response to an Article 9 order or Article 16 notice. These can encompass both proactive and reactive enforcement actions. Proactive enforcement takes place when potentially policy-violating content has been flagged internally, for example, via algorithms or contractors. Reactive enforcement takes place in response to external notifications, such as user policy flags or legal complaints. Google services are wide-ranging and differ in their user bases, content hosted, services provided, and expectations for enforcement. To support information and content quality on our products and services, we take a wide range of enforcement actions to maintain a trusted experience for all. Enforcement actions, like policies, differ from service to service and are tailored to the purpose of each service, based on what is reasonable, proportionate, and effective, taking into consideration the appropriate approach to freedom of expression for each service. However, the types of restrictions applied may include: (i) restrictions of the visibility of content; (ii) restrictions of the ability to monetise content; (iii) restrictions of provision of some or all features of the service; and (iv) Google-wide restrictions of an account where users can no longer log into any Google service, which may be imposed as a result of multiple legal or policy violations across one or more services. In some cases, restriction of monetisation, by itself, is not an applicable enforcement action for a service, however, a different enforcement action (e.g., restriction of provision of the service) may prevent features from being monetised. Google considers 'measures' as actions taken on moderated videos, URLs, listings, accounts, and other content types, which are of a policy-violative nature or are delisted as a result of applicable law. Where feasible, the categories and sub-categories identified by the European Commission are used to group and report enforcement actions. However, some policies do not fully align with these sub-categories, and are thus reported using additional sub-categories (i.e. using the KEYWORD_OTHER option provided by the European Commission). For relevant services, we separately present metrics relating to advertisements impressed on those services (on the 6_own_initiative_Ads tab). While intermediary services are out-of scope for Art 15(1)b, actions taken on a legal basis in response to legal notices received about intermediary service content are reported under Art 15(1)c for completeness. Finally, metrics on actions taken that impact multiple Google services are included in the Multi-Services report. For more information, see: - Google Transparency Center: https://transparency.google/our-policies/product-terms/ Service-specific details The majority of the actions that Google Shopping takes happen before the content is shown publicly, and the actions may apply to both unpaid content (e.g., free listings) and advertisements. As such, Google Shopping cannot readily distinguish between unpaid content and advertisements in these metrics, therefore they are combined. In addition, for non-Shopping content, content moderation actions on advertisements that are taken before the advertisement is surfaced on a VLOSE or VLOP are not included in this report.

Wellbeing support Google is committed to supporting the wellness of its employees that work with sensitive content through comprehensive programs and resources. Google strives for safe and healthy working conditions for all employees exposed to sensitive content and is committed to ensuring they have the highest standard of support. Google has invested significantly in these teams by: - Providing access to on- and off-site counselling for workers who need it, dedicated wellness spaces, on-site specialist counsellor support in certain Google offices, and 24/7 phone support; - Limiting content exposure for those focusing on sensitive content by providing guidance on daily review time; - Providing materials for individuals to form peer-led peer support groups and optional listening sessions if teams experience escalations or specific events that are particularly impactful; - Providing physical and mental wellbeing activities (e.g., gym space, workout classes, mindfulness app access, educational sessions on a variety of topics); and - Providing post-exit mental health support, including counselling services, for at least six months after an employee who was regularly exposed to sensitive content and situations as part of their core role exits their position at Google. Consistent with Google's Wellness Standards, members of Google's extended workforce working with sensitive content are offered counseling, peer support groups, wellness breaks, mental health and wellbeing resource support directly by their employer both during their work on sensitive content moderation and wellness support in the form of counseling remaining accessible to content moderators for at least six (6) months after exit. Research and technological innovation In addition to gathering feedback directly from workers and soliciting professional input and advice, Google is committed to driving industry-leading research and technological innovation in the field of content moderation. For instance, Google published a research paper in 2019 indicating that 'grayscale transformations' (i.e., where an image was converted to black and white) reduced the emotional impact of reviewing violent and extremist content. Based on these findings, Google built grayscaling into review tools, giving each reviewer an option to use this feature when performing reviews, based on their own preference. For more information on Google's research paper, see: https://research.google/pubs/testing-stylistic-interventions-to-reduce-emotional-impact-in-content-moderation-workers/

Google employees who work on sensitive content teams are offered subject matter specific training on a variety of topics. Employees working in sensitive content are required to complete a training on the Psychological Impact of Sensitive Content Review at the point of onboarding, and managers are required to complete an additional training on Supporting Teams who Work with Sensitive Content. Additional optional training opportunities include those on self-compassion, emotional agility, and subject matter specific training to provide a deeper dive into the unique challenges faced by each team. The training is generally conducted via e-learning with opportunities for live facilitated training.