RTFP

AliExpress¡¯ content moderation governance structure consists of several teams with clear lines of responsibility and accountability. These teams include employees who develop the applied automation technologies, administer content policy and deliver enforcement. Our content moderation teams are multilingual, globally dispersed, and drawn from a wide variety of relevant specialisations (such as prohibited and unsafe products, intellectual property or consumer protection) and cultural competencies. We have separate teams for different functions. Our Platform Rules Department team designs the rules and policies governing sellers¡¯ product listing and user-generated content, with the support and input of the relevant stakeholder teams related to the topic of the relevant rule and/or policy being enacted or reviewed. The Chief Risk Office Department (CRO) enforces the rules and ensuring compliance with legal requirements and the platform rules. Its Risk Review and Resolution team is responsible for reviewing and responding to items and content flagged by the automated detection or by users through the user report system. These teams' work is reinforced by data and insights provided to them by various internal and external stakeholders. With regard to emergencies (e.g., a public health crisis), we also have a dedicated task force comprising relevant employees across different teams who classify crises and implement an appropriate content moderation response.

The methods that AliExpress uses to detect illegal or non-compliant content and used for actions that flow from detection are summarised in the value to Indicator 1 above. AliExpress has in place several metrics which track users' exposure to or reach of illegal or non-compliant content, including the following major metrics: 1. Exposure of illegal listings: the total views of all illegal listings that have been moderated after the publication during the reporting period represents only 0.5% of the total views of all EU-relevant products. 2. Exposure of incompatible with T&Cs listings: the total views of all incompatible with T&Cs listings that have been moderated after the publication during the reporting period represents only 8.6% of the total views of all EU-relevant products.

During the reporting period, we have increased staffing at our content moderation teams from 876 to 986 moderators. In order to calculate the moderators volume, we apply a mathematical formula, including with regard to staff personnel under outsourcing contract providing services under a piece rate model. We carry out regular lookback exercises involving retrospective data analysis and quality assurance sampling to assess whether benchmarks for effective moderation are being met and whether resources are properly aligned with evolving needs. We will keep adjusting the size of our moderation task force, including hiring additional human moderators when necessary and improving moderation tools, to ensure adequate capacity for the moderation of potential illegal content.

Our team¡¯s qualifications Our content moderation teams consist of experts across diverse fields. Computer scientists and technology engineers build and continuously improve our automated detection tools. The team responsible for developing our platform rules and policies includes experts in subject matters such as fraud prevention, product safety and compliance, intellectual property infringement and data privacy. They are aided and supported by a specialised task force that scans for emerging trends. We have legal professionals with expertise in, among other fields, product safety and compliance, intellectual property, consumer protection and data privacy. Our team of professionals ensure our content moderation practices are aligned with relevant global regulatory requirements. This year both the number of our internal moderators and contracted external moderators was increased to intensify our platform governance measures as well as cover additional control policies. In addition, AliExpress has several teams that actively work to understand and detect emerging risks as they appear on the platform. For example, our Emergency Service Unit (ESU) keeps a close eye on trends and issues that might indicate potential risks for AliExpress through public sources, such as news and social media platforms. We operate regularly monitored internal reporting channels for teams across the company to notify our content moderation experts of potentially illegal, non-compliant or otherwise risky content. Linguistic expertise To enforce AliExpress¡¯ content policies consistently across our global user-base, our content moderation teams review content primarily in English (which is the publication language of over 90% of relevant content). Our moderators are nevertheless supported by world-class translation technology servicing the EU languages in which EU users access AliExpress. To enhance our human moderation capabilities, we have 9 human moderators focusing on non-English content (specifically, Spanish, Polish, Portuguese, French and German). In response to the needs of our global user base, we have also extended the availability of local language support by implementing AI translation technology. This allows our 24/7 English chat service to be translated into local languages, which supports translation to Spanish, Portuguese, Italian, Dutch, German, French, and Polish.

Among content fully or partially subjected to our automated system, 95.31% of flags generated by automated means are correct (i.e., flagged content whose moderation decisions were not overturned during the reporting period, and unflagged content that did not incur moderation during the reporting period). This indicator demonstrates the overall accuracy of our automated means, including their ability to correctly flag content as compliant or requiring moderation during the reporting period. The precision rate (99.37%) reflects the percentage of all flagged content with non-overturned moderation relative to all content flagged by the automated system. This indicator demonstrates how precise our automated means when flag a content as non-compliant or illegal during the reporting period. The recall rate (99.87%) indicates the percentage of all moderations made during the reporting period that were successfully identified and flagged by the automated system relative to all non-overturned moderation. This indicator demonstrates the coverage of the flags by automated means on all illegal or non-compliant content during the reporting period.

At AliExpress we employ a comprehensive range of automated means with the goal of ensuring a safe and compliant online shopping environment. In accordance with Article 15(1)(e), this section provides a qualitative description of the use made of our automated content moderation systems. In line with that same provision we also provide a specification of the precise purposes of our automated systems, the indicators of the accuracy and the possible rates of error of our automated content moderation and any safeguards applied to these systems. We build and train specialised algorithms, advanced technology, and proactive measures to identify and mitigate risks and enforce platform policies. Automated detection and content filtering When product listings are uploaded on the platform, we utilise specialised algorithms to detect illegal and non-compliant products. The algorithms are trained to perform tasks such as text analysis, image recognition, user behaviour modelling and sales records analysis that help us in making predictions about illegal and non-compliant products. For example, our proactive content moderation comprehensively analyses text and images for suspicious signals such as logos, icons, text or cartoon images to predict whether a listing is potentially violating our intellectual property policies. With a view to ensure that the platform remains a safe environment for our users, over time we have deployed a combination of algorithms to filter and detect non-compliant or illegal contents. This includes restricted keywords, where high-risk keywords are identified for various risk types, such as Hate Speech and Child Sexual Abuse Material (CSAM), and used to identify and take action on illegal or non-compliant content. Specialised algorithms such as Optical Character Recognition (OCR), obscenity detection and terrorism detection, are also leveraged to identify non-compliant or illegal content across various platform sections, such as usernames, reviews, Q&A and instant messaging. Automated enforcement We employ automated systems to enforce our policies and penalise policy violators. These systems automatically remove illegal and non-compliant content and track penalties for repeat offenders based on their infringement history. Continuous enhancements to our processes: AI-generated content moderation We are constantly working to enhance our content moderation capabilities and that is why we keep on enhancing advanced algorithms designed to identify potential illegal content risks, including those stemming from AI-generated material. These algorithms, based on large language models (LLMs) and natural language processing (NLP), enable the platform to better analyse context and full sentences, providing a more comprehensive approach to risk detection.

At AliExpress we strive to constantly improve our automated systems by employing a combination of advanced algorithms and human reviewers. We continuously evaluate the accuracy of our automated systems and use error cases to make improvements to the technology in an iterative process. Human reviewers are essential for identifying moderation mistakes and addressing nuanced content that automation may misinterpret. For example, our risk management personnel maintain a carefully curated list of restricted keywords that enables our text detection algorithm to pick up on prohibited content that may be harder to identify by automated tools. To improve the precision of our algorithms, prohibited content detected by human reviewers is continuously used to train our technology, allowing for iterative enhancements over time. Over the years, we have also developed tailored algorithms to address novel risk scenarios. For example, our multi-modal algorithm for illegal and non-compliant products was implemented to extend the capacity for detection of illegal and non-compliant content from only text to other forms of content, such as images and video. To prioritise accurate and scalable content moderation, our automated systems translate into English any content originally published in a language other than English. Given that over 90% of relevant content on AliExpress is originally published in English, the scalability challenges of conducting content moderation in multiple languages, with low content volumes, would likely diminish the quality of our detection and review systems. Limited training data in languages other than English makes it difficult to build accurate machine learning models and obtain representative samples for quality control. To disregard these limitations would be to risk introducing bias to the enforcement of our policies. Instead, we have focused our resources towards optimising our English language content moderation systems and introducing robust safeguards to continuously strengthen our systems.

As set out in our values in Indicator 1 and 3, AliExpress uses automated means mainly in order to conduct automated detection, content filtering, antomated enforcement and to constantly enhance our content moderation capablities. All identified risks are addressed by the use of automated means.

AliExpress¡¯s own initiative content moderation is described in detail in its Risk Assessment and Mitigation Report (available at https://www.aliexpress.com/p/transparencycenter/mitigationReport.html), which has been prepared in accordance with Regulation (EU) 2022/2065 (the "DSA") . As a high-level summary, content moderation engaged in at AliExpress' own initiative combines pre-listing detection and continuous monitoring. Pre-listing detection We employ text-based automatic scanning and image-comparison algorithm, which involves the automatic detection of high-risk keywords and images pre-selected by our moderating team, drawing from past violations. If our moderators find that the item is illegal or incompatible with terms and conditions ("non-compliant"), it will be moderated and added to our internal reference database as appropriate. Once our detection methods identify suspected illegal or non-compliant content, the content will either be automatically blocked by our system or submitted to manual review by our risk management personnel. Specifically speaking, if our automated systems detect illegal or non-compliant content, and the accuracy of the automatic detection is judged to be over a certain expected high accuracy threshold and triggers the rule configured in the system as not requiring human review, then it will be automatically actioned upon. Continuous monitoring We place a strong emphasis on ensuring the quality of products on our platform and, for such purposes, deploy measures through a multi-pronged approach. This includes proactive compliance certification for certain product categories, where we require sellers to upload product certificates based on the associated product's risk level to ensure adherence to relevant local standards for these products. We also employ vigilant monitoring measures to detect and moderate products that match the information from official alert portals such as "Safety Gate" to ensure a safer marketplace. Our actions Our moderation actions include removing, blocking or deleting the illegal and non-compliant content from our platfrom. Furthermore, penalties may also be imposed by applying a sliding scale of penalty points to sellers¡¯ breaches of our rules and policies, depending on the severity of the behaviour, repeated violations, and other associated factors. These penalty points will accumulate for any relevant breaches of our platform's rules and policies, and, if the total points exceed a predetermined threshold, the associated store account will be subject to penalties, applying a sliding scale, which can go up to freezing or termination of the account. A full list of our sellers¡¯ violation and penalty rules can be found at https://rule.aliexpress.com/rule-channels/37978936/?tocUuid=lU0jOhlxYPyIWqF6.

We are committed to safeguarding the health and well-being of our review teams. Content review can involve exposure to confronting or otherwise graphic materials, and so we deeply respect and value the unique challenges of this work. AliExpress provides healthcare and a free psychological counselling hotline that employees are encouraged to take advantage of. In addition, we conduct regular training for our review teams to ensure they are well equipped to safely respond to potentially challenging content.

We understand that to ensure a high level of accuracy and efficiency of our content moderation, it is crucial for responsible teams to undergo continuous professional development, keeping them up-to-date with the latest developments in industry standards and relevant risk areas. We organise regular training sessions and awareness raising programs for our staff involved in content moderation. This includes topic-specific training on subjects such as Intellectual Property Rights (IPR), product safety and compliance and data privacy. To give one such example, IPR-related teams participate in rigorous training programs on IPR protection, including receiving training from right holders or from external legal consultants, typically arranged twice a year. There are internal trainings based on the updated IPR Standardised Operating Procedures (SOPs) from time to time. We also ensure that all moderators are aware of our SOPs and that all moderating professionals consistently review the qualifications and certifications of sellers who may sell certain products liable to be restricted. In addition, we conduct regular reviews on content moderation decisions to check the accuracy and quality of the decisions made by our review teams. Our review teams are also encouraged to offer feedback on our review policies and processes. Reviewers are on the front lines and can offer valuable insights about the practicability and effectiveness of our content review system.