RTFP

Amazon's content moderation governance structure is designed to provide a safe, trustworthy shopping experience through a multi-layered approach, integrating specialized teams across the organization. Senior leadership oversees strategy and performance through regular business reviews, with dedicated teams responsible for content moderation decisions, such as in the areas of intellectual property protection, product safety, customer review authenticity, and seller verification. Amazon’s content moderation processes combine automated systems using advanced machine learning, with human review by trained expert investigators. Clear escalation paths enable program managers to coordinate across operational teams, legal, and business stakeholders for consistent policy application. Amazon employs thousands of people globally - including machine learning scientists, software developers, and expert investigators - dedicated to protecting customers, brands, and selling partners from counterfeit, fraud, and abuse. Decision-making authority is distributed across specialized functions with controls in place to ensure actions are taken diligently, objectively, and proportionately. The structure incorporates regular audits, regular analysis of false positive and negative reviews, continuous improvement processes, data validation, and leadership oversight to refine enforcement accuracy. This governance framework is supported by documented change management processes, business review forums, and cross-functional collaboration, so Amazon can adapt its content moderation processes as risks evolve.

See row #2 above

We count the number of operations headcount that handle DSA Articles 16 and 22 notices and DSA Article 20 complaints.

Amazon employs machine learning scientists, data analysts, software developers, and expert investigators dedicated to protecting customers, brands, selling partners, and our store from illegal content, including counterfeit, fraud, and other forms of abuse. All Amazon staff, including staff dedicated to content moderation, are required to meet Amazon's Leadership Principles. The Leadership Principles are a set of guidelines that Amazon employees use every day to solve problems, evaluate trade-offs, and make decisions. There are 16 in total, and they are the framework of how we evaluate potential candidates for jobs and set the expectations of performance across Amazon. In addition, the level of qualification and expertise our content moderators have is diverse and varies depending on their specific job role. Most of our content moderators have a bachelor's degree in relevant fields of study, including computer science, information technology, data science, information security, finance, foreign studies, intellectual property, and risk management. All staff dedicated to content moderation have demonstrated experience performing research on a variety of topics, including fraud, abuse, trust, and risk; and have the ability to investigate complex and highly technical problems, and perform root cause analysis.

Accuracy Definition: The ratio of all correct model predictions (both positive and negative) to total predictions. Formula: Accuracy = (TP + TN) / (TP + TN + FP + FN) Recall Definition: The models’ ability to find all positive instances. Of all actual positives, how many did we correctly identify? Formula: Recall = TP / (TP + FN) Precision Definition: The accuracy of positive predictions - of all items flagged as positive, how many were correct? Formula: Precision = TP / (TP + FP)

See row #2 above

To safeguard against potential errors made by our automated tools, we implement processes to ensure that we have a high confidence rate that our automated tools operate as intended and to minimise mistakes. We do this by ensuring our automated tools meet a high bar of accuracy before they are launched by testing the provision of the control, and by continuously auditing our automated tools after they launch and removing from use automation that does not maintain a sufficiently high level of accuracy. We also constantly improve our automated tools by training them using new information, including internal learnings and developments (including outcomes of expert manual decisions) and external risk signals, so they can learn and constantly get better at proactively identifying and blocking non-compliant products automatically.

Seller verification: Amazon uses advanced technology and expert human reviewers to verify the identities of potential sellers. When prospective sellers apply to sell in Amazon’s store, they are required to provide a form of government-issued photo IDs, along with other information about their business. We employ advanced identity detection methods like document forgery detection, image and video verification, and other technologies to quickly confirm the authenticity of government-issued IDs and whether they match the individual applying to sell in our store. In addition to verifying these, Amazon’s systems analyze numerous data points, including behavior signals and connections to previously detected bad actors, to detect and prevent risks. Similarly, throughout the selling experience in our store, Amazon’s systems monitor selling accounts to identify anomalies or changes in account information, behaviors, and other risk signals. In the event Amazon identifies a risk of fraud or abuse, we promptly initiate an investigation using automated and/or human review, request additional information where helpful, and swiftly remove bad actors from our store. Product Safety and Compliance: Our content moderation systems aimed at product compliance include controls that function through automated rules to identify and remove non-compliant products. We employ thousands of keyword-based algorithms and machine learning models that are continuously run against the EU store?s product catalogue, considering linguistic differences and local compliance requirements by EU storefront location, to identify potential policy violations. These controls aim to prevent non-compliant products from being listed or flag them for Amazon?s expert investigators so listings can be stopped if compliance issues are found or additional information is needed from sellers. Automated brand protections Amazon’s Intellectual Property Policy prohibits listings that violate rights owners’ IP rights. Amazon Brand Registry, a free service launched in 2017, enables brands to more effectively protect their IP, whether or not they sell on Amazon. Through Brand Registry, brands can share IP and product data, which Amazon uses to prevent potential infringements. The purpose of these automated brand protections is to detect content that likely infringes the IP rights of brands and other rights owners. Amazon’s automated technology scans billions of attempted changes to product detail pages daily for signs of potential abuse, including the creation of new listings and changes to existing listings. For example, our tools use advanced machine learning to prevent the attempted listing of counterfeit or infringing products—scanning keywords, text, and logos which are identical or similar to registered trademarks or copyrighted work. We use the data and learnings gathered throughout these processes to innovate and improve our proactive protections. When we receive a valid notice of infringement or a customer complaint, our machine learning algorithms use this information to learn and improve protections for brands. Advertising: We proactively detect and remove advertising content that violates our Ads Policies, which are designed to maintain a high customer experience bar for ads on the store. We require all advertising content to comply with all applicable laws, rules, and regulations; to be appropriate for a general audience, and honest about the products or services that ads promotes. For example, we prohibit deceptive, misleading or offensive ads, as well as violent and certain sexual content. We invest heavily in people and technology to protect customers, brands, and advertisers and the EU store from fraud and other forms of abuse. Amazon deploys a number of measures to ensure compliance with our Ad Policies and detect infringing ads, including through automated moderation tools that check millions of ads and their visible ad elements per day worldwide (including advertiser-supplied images, product listing titles and images, and product descriptions). For example, we implement deny lists on certain products that block all ads for customers who search for specific query terms e.g. guns. Ads Policies also block specific listings for being viable for advertising. To complement our automated measures, expert teams also conduct human reviews of ads to identify any potential non-compliance and apply the learnings as feedback to continually improve our automated moderation tools. Trustworthy Reviews: Our moderation processes for community content include machine learning models that detect content that violates our Community Guidelines and prevent it from being published. We strictly prohibit fake reviews that intentionally mislead customers by providing information that is not impartial, authentic, or intended for that product or service. We invest significant resources to proactively stop fake reviews. This includes machine learning models that detect risk, including relationships between accounts, sign-in activity, review history, and other indications of unusual behaviour, as well as expert investigators that use sophisticated fraud-detection tools to analyse and prevent fake reviews from ever appearing in our store. Our machine learning models analyse millions of reviews each week using thousands of data points to detect risk. The review ranking algorithm considers signals from Amazon?s fraud-detection tools related to the authenticity of a review. When we strongly suspect that a review is inauthentic, we suppress the review completely, so it is not displayed in the Amazon EU Store. Offensive and controversial products: Amazon prohibits the sale of products that promote, incite, or glorify hatred, violence, racial, sexual, or religious discrimination or promote organizations with such views; contain pornography, glorify rape or paedophilia or promote the abuse or sexual exploitation of children; or graphically portray violence or victims of violence, and advocates terrorism; among other material deemed inappropriate or offensive. We leverage machine learning and automation to filter listing submissions that we suspect of potential policy violation, and then our content moderation teams manually review these suspect listings. We use machine learning and manual review to filter potentially policy-violating listings.

Our voluntary controls use advanced machine learning techniques and automation to monitor different aspects of our store for potentially fraudulent, infringing, inauthentic, non-compliant or unsafe products or content to maintain a trustworthy shopping experience. Our automated detection tools operate continuously throughout every step of selling in our store, starting from when a prospective seller begins their registration process to listing or updating a product, changing key account information, receiving a funds disbursement, and more. In most cases, bad actors are stopped from even creating an account or listing a single product for sale, and prohibited content is stopped before a customer ever sees it. Our automated tools help us scale our protections and take action more quickly. They help us operate at scale to prevent bad actors from registering an account, and to detect and remove listings or other content that violate our policies or the law. These automated tools range from text-based algorithms that identify specific keywords to sophisticated image recognition and machine learning models. Once our tools have identified potentially infringing or illegal content, we use a mixture of automated tools and expert investigators to determine the appropriate enforcement action. When our automated tools identify prohibited content with a high degree of confidence then they automatically take enforcement action. We use the data and learnings gathered from these technologies and valid notices of infringement or illegal content to innovate and improve our controls.

See row #10

Our expert investigators dedicated to content moderation, including the administration of Amazon's notice and action mechanisms, complaints and appeals procedures, are trained to identify illegal content and content that infringes our terms and conditions. Our investigators receive: an exhaustive onboarding process to familiarize themselves with the underlying policies and standardized operating procedures, which must be completed before they are able to take their own moderation decisions; robust continued on- the- job training and periodical knowledge tests, including on any new tools or processes; and when needed, support from subject matter coaches and escalation paths to team managers. This includes training in the relevant subject matter to better protect our customers and, additionally, regular training on our company's policies, terms and conditions, and their specific area of expertise, whether that's product safety and compliance, IP and brand protection, controversial content, or misleading customer reviews. For example, investigators dedicated to evaluating IP infringement notices receive training and support in accurately identifying different types of infringing listing content, including trademarks, copyright, design and patents. Similarly, investigators creating product listing rules have detailed knowledge of Amazon's catalogue and are trained to accurately develop and apply product listing rules.