RTFP

Meta Platforms Ireland Ltd. (“Meta Ireland”) operates under a structured governance framework designed to ensure effective oversight and accountability. Meta Ireland is led by a Board of Directors which is the relevant management body for the purposes of the Digital Services Act (DSA). The Board of Directors comprises “those charged with governance” under applicable corporate and company law and, as such, has primary and overarching responsibility for ensuring MPIL’s compliance with obligations arising in connection with the DSA. Meta Ireland has also established a committee of Meta Ireland directors and senior executives to oversee statutory compliance with the requirements of the DSA. Moderation decisions are made by dedicated teams, following established policies and procedures that prioritise user safety and transparency. In complex cases, moderation decisions may involve cross-functional input from multiple stakeholders to ensure consistency. This governance structure enables Meta Ireland to maintain robust oversight of compliance with its DSA obligations.

Every day, we remove millions of violating pieces of content and accounts on Instagram. In most cases, this happens automatically, with technology to detect and remove content and restrict accounts that may go against our Community Standards, Advertising Standards, Commerce Policies, and Content Monetization Policies. In other cases, our technology selects content for human review. Our review teams review a blend of user reports and content surfaced by our technology. Our technology also supports the review teams by prioritising the most critical content to be reviewed, based on severity, virality, and likelihood of a violation. Our review systems use technology to prioritise high-severity content with the potential for offline harm (e.g., posts related to terrorism and suicide) and viral content that is spreading quickly and has the potential to reach a large audience, in order to prevent as much harm as possible. Our technology is trained to identify violations of our Community Standards, Advertising Standards, Commerce Policies, and Content Monetization Policies. There are three primary forms of technology used to detect possible policy violations. First, we employ rate limits (speed limits) on how rapidly accounts can perform multiple actions on our platforms, including making posts, to prevent the usage of bots. Next, we have matching technology that identifies identical or near identical copies of URLs, text, images, audio, and videos that we have previously identified as violating our policies. When we match the content exactly or we determine it is near identical, we will typically remove the content. Finally, we also use artificial intelligence (AI) to augment and scale our human review capacity with appropriate oversight. Depending on the results of the review against our policies, typically the AI will either take no action, demote the content or remove it. We also use AI to select the content and/or account for human review on the basis of severity, virality, and likelihood of a violation. As with matching technology, AI operates on URLs, text, images, audio, and videos. Unlike technologies that can only match violations they’ve seen before, AI has the potential to identify certain violations it has never seen before. In the context of advertisements, when advertisers place an order, each ad is reviewed against our policies. Our Advertising Standards provide policy detail and guidance on the types of ad content we allow and the types of ad content we prohibit. Our Advertising Standards also provide guidance on advertiser behaviour that may result in advertising restrictions being placed on a business account or its assets (an ad account, Page, or user account). Our ad review system relies primarily on automated tools to check ads and business assets against our policies. Of all content removed on Instagram in H2 2025, 70% was never viewed before removal.

Our human resources dedicated to content moderation are reflective of reviewers who work on EU-specific content across EU official languages For EU content that is in a language other than the 24 official languages of the EU, there are additional language-based content reviewers. For languages that are widely spoken outside of the EU, e.g. French, English, Spanish, Portuguese, there are additional content reviewers that review reports from non-EU countries in these languages.

Our content review team is global and reviews content 24/7 in over 80 languages. The team includes reviewers with language expertise to enforce our policies in cases where certain words or content require additional contextual understanding. Our teams also include experts in enforcement in policy areas such as child safety, hateful conduct, and counterterrorism. We partner with companies to help with content review, which allows us to scale globally with coverage across time zones, languages, and markets. For content that requires specific language review in the EU, there are dedicated teams of reviewers that perform content moderation activities specifically for that content. For EU languages that are widely spoken outside the EU, like Portuguese and Spanish, we have content moderation teams that provide global coverage.

Our technology learns and improves from human decisions. Over time – after learning from thousands of human decisions – the technology gets better. When reviewing violating content, review teams manually label the policy guiding their decision, which means that they mark or “label” the relevant policy that the content, account, or behaviour violates. This labelling of data helps us improve the quality of our algorithms that proactively detect and remove harmful content, accounts, and behaviour. To ensure and improve the quality, i.e., how accurate the technologies mentioned above are in enforcing Community Standards and other policy violations, there are ongoing quality evaluation processes in place. We use overlapping techniques and systems for maintaining a high overall accuracy for our automated content moderation. Prior to fully launching any new rate limit (speed limit), matching technology or AI, we use the technology to only log how the technology would have behaved instead of immediately acting. We then use human reviewers to assess the accuracy against current content, accounts, or behaviour, rather than just historical ones, as we did during the technology’s training. After launching rate limits, matching technologies, or AI, we monitor the volumes of actions and appeals by the user who posted the content, as well as the rate at which appeals are granted. If the metrics we monitor are abnormal, our engineering teams may investigate. For each primary form of automation technology, the investigation of abnormal metrics can vary. For our matching technologies, if an entry in our list of previously identified instances of policy violations has abnormal signals, we will re-review the entry to confirm it continues to go against our policies. Similarly, if one of our AI tools has abnormal signals, we will either send a sample of the AI tool’s recent results to human labelling to confirm the accuracy rate or deprecate the AI tool if abnormal signals indicate a clear breakage. In addition, many of our machine learning classifiers are reassessed for accuracy after human review. This classifier reassessment is an example of the general feedback loop between human review and technology. The content labelling decisions taken by human reviewers are used to train and refine our technology. As a part of this process, the review teams manually label the policy guiding their decision, i.e., they mark the policy that the content, account, or behaviour violates. This helps to improve the quality of our AI algorithms and our lists of known policy-violating content used by our matching technology. To maintain quality control across these decisions, we regularly audit random samples of decisions taken by the algorithm and our content reviewers and measure them against our expectations for policy enforcement. In the context of automation relating to language, some automation is developed to support specific languages whilst others are language agnostic. While various types of automation necessitate different and overlapping techniques for assessing performance, we are able to measure precision, recall, and accuracy across all types of automation. To measure precision, recall, and accuracy we assume all pieces of content actioned using automated means, that are later restored, are true automated mistakes. To measure recall and accuracy, we also assume all reactively actioned and human actioned content are true automated misses. While these assumptions are not true in all cases, they enable the creation of directionally approximate indicators for precision, recall, and accuracy.

We use technology to help us proactively detect and take action on content on our services that might be harmful and violate our Community Standards, Advertising Standards, Commerce Policies, and Content Monetization Policies. These technologies run on accounts, posts, comments, photos, and other pieces of content uploaded to Instagram. They determine how probable or likely it is that the content violates a certain policy, based on those signals or patterns, and if the content should be automatically enforced on. Our technology is trained to identify violations of our Community Standards, Advertising Standards, Commerce Policies, and Content Monetization Policies. There are three primary forms of technology used to detect possible policy violations. First, we employ rate limits (speed limits) on how rapidly accounts can perform multiple actions on our platforms, including making posts, to prevent the usage of bots. Next, we have matching technology that identifies identical or near identical copies of URLs, text, images, audio, and videos that we have previously identified as violating our policies. When we match the content exactly or we determine it is near identical, we will typically remove the content. Finally, we also use AI to augment and scale our human review capacity with appropriate oversight. Depending on the results of the review against our policies, typically the AI will take no action, demote the content or remove it. We also use AI to select the content and/or account for human review on the basis of severity, virality, and likelihood of a violation. As with matching technology, AI operates on URLs, text, images, audio, and videos. Unlike technologies that can only match violations they’ve seen before, AI has the potential to identify certain violations it has never seen before.

To ensure and improve the quality, i.e., how accurate the technologies mentioned above are in enforcing Community Standards and other policy violations, there are ongoing quality evaluation processes in place. We use overlapping techniques and systems for maintaining high overall levels of accuracy for our automated content moderation. In addition, many of our machine learning classifiers are reassessed for accuracy after human review. This classifier reassessment is an example of the general feedback loop between human review and technology. The content labelling decisions taken by human reviewers are used to train and refine our technology. As a part of this process, the review teams manually label the policy guiding their decision, i.e., they mark the policy that the content, account, or behaviour violates. This helps to improve the quality of our AI algorithms and our lists of known policy-violating content used by our matching technology. To maintain quality control across these decisions, we regularly audit random samples of decisions taken by the algorithm and our content reviewers and measure them against our expectations for policy enforcement. In the context of automation relating to language, some automation is developed to support specific languages whilst others are language agnostic.

As described in this report, we use technology to help us proactively detect and take action on content on our services that might be harmful and violate our Community Standards. These technologies run on accounts, posts, comments, photos, and other pieces of content uploaded to Instagram. They determine how probable or likely it is that this content violates a certain policy, based on those signals or patterns, and if the content should be automatically enforced upon. Our technology also helps us prioritise the most critical content to be reviewed, based on severity, virality, and likelihood of a violation.

Instagram maintains a set of globally applicable Community Standards that define what is and is not allowed on Instagram. In addition, we have Advertising Standards and Commerce Policies in place for advertising and commerce content, respectively; and Content Monetization Policies. If we determine that a user's content goes against our policies, we will remove it. We will also notify users so they can understand why we removed the content, how to avoid posting violating content in the future, and how to appeal our content moderation decisions. We appreciate that some content can create a negative experience for people even though it does not violate our policies. Actions we may take on such content to promote a safe and positive experience are not taken on the basis that the content is illegal or violates our terms and policies.

We recognise that reviewing content can be challenging work. Keeping people safe online sometimes means review teams have to look at content that may be objectionable or graphic. We respect the difficulty of this work and work with industry leading vendors to ensure reviewers have access to the resources they need to do their job and support their health. There is a robust and diverse program to support human reviewers. Our vendor contracts mandate high quality support in a variety of areas, including pay, benefits, work environment, and wellbeing and psychological support. The assistance model depends on what type of content reviewers work on. Such assistance, for example, can take the form of psychological support, including individual and group sessions, and a 24/7 independent support program which includes a range of offerings that include clinical services.

Human reviewers come from different backgrounds and reflect our diverse community.Human reviewers undergo extensive training when they join and are regularly trained and tested beyond this initial training, with specific examples, such as how to uphold the Community Standards and take the correct action on a report. We also do our own proactive audits, where we conduct re-reviews that help us figure out if we are getting it right. Human reviewers who review content alleged to be illegal receive distinct training based on the nature of their respective work. Every member receives several weeks of training focused heavily on operational proficiency and in preparation for processing such content. For example, the reviewers who review content for defamation receive training specifically on assessing defamation