Homophobic Violence in West Africa

To improve implementation of its policy, Meta should conduct an assessment of the enforcement accuracy of the at-scale prohibition on exposing the identity or locations of anyone alleged to be a member of an outing-risk group, under the Coordinating Harm and Promoting Crime Community Standard.

The Board will consider this recommendation implemented when Meta publicly shares the results of the assessment and explains how the company intends to improve enforcement accuracy of this policy.

Our commitment: We will measure the accuracy of our enforcement of the “harm against people” subset of the Coordinating Harm and Promoting Crime Community Standard and will further assess the feasibility of conducting a more granular assessment of our enforcement accuracy for exposing members of outing-risk groups, which lies within the “harm against people” subset.

Considerations: We continuously monitor and assess the accuracy of our enforcement measures in line with our commitment to ensuring that our policies are accurately enforced. Accordingly, we periodically conduct reviews by looking at random samples of violating content – across both human and automated review - and assessing whether we made an accurate at-scale decision. To this end, we are investing in new technology, including classifiers to improve the enforcement of our Coordinating Harm and Promoting Crime policy. These improvements should also yield improvements in enforcement accuracy of policy subsets, including the provision focused on outing-risk groups.

Currently, the various subsets of outing are enforced under the “harm against people” sub-section of our Coordinating Harm and Promoting Crime policy. This addresses the outing of risk groups and their affiliates, among other variations. When reviewers enforce against outing in the policy, they assess the specific policy violation, but only apply a label for the broader policy category: “harm against people.” In order to conduct an assessment on the “outing risk groups” policy line, which is a narrow subset of the prohibitions we maintain against “harm against people”, our operational, policy, and product teams will need to identify whether there is an effective combination of technical and procedural adjustments that would enable reviewers to conduct more granular labeling and allow for more detailed assessments of accuracy. Furthermore, as outing also pertains to other policies such as Bullying & Harassment, our teams will need to determine the best method to capture the holistic enforcement accuracy of outing across our policies.

Overall, our iterative approach to reviews allows us to gather insights on the performance of our enforcement systems, as well as provide a signal when improvements are needed. We will continue to improve our enforcement accuracy review process and share the progress of our efforts in future reports.

To increase the efficiency and accuracy of content review in unsupported languages, Meta should ensure its language detection systems precisely identify content in unsupported languages and provide accurate translations of that content to language-agnostic reviewers.

The Board will consider this recommendation implemented when Meta shares data signaling increased accuracy in the routing and review of content in unsupported languages.

Our commitment: We will continue to expand on our efforts to improve our language detection systems and better identify unsupported languages. We will expand on our strategy to improve staffing and automation processes for language-agnostic reviewers.

Considerations: We are committed to providing an equitable experience for users across our platforms and to help ensure any enforcement against users considers the necessary elements of language, culture and context. Our continued investment in improving language detection across all languages includes iterations to our language identification models which, among other things, help us provide accurate translations of that content to language-agnostic reviewers with speech-to-text translation, speech-to-speech translation, text-to-text translation and text-to-speech translation across a growing number of languages.

We have a streamlined process with dedicated staffing and resources to ensure more efficient and expedient review, particularly for high priority agnostic review. Given the specifics of regional context, we are mindful of the limitations of scaling to all languages and dialects, yet we continue to refine our detection processes to better understand the nuance of locality.

We are aware that language-agnostic reviewers often need additional signals beyond accurate translations of the content, and we continue to strive towards technology solutions that provide auxiliary support to reviewers. We are innovating across various pathways, including using artificial intelligence to identify areas in a long-form video or post where agnostic reviewers should focus their time. Additionally, we are including AI-generated contextual notes to help reviewers identify words and phrases they may misinterpret due to instances such as locally trending events or terms, slurs, historical context or celebrity bait among many others. Additionally, we are refining these functions to help agnostic- language reviewers to identify the probable violation by proposing the likely violated policy and reasoning.

As we continue to improve and refine our efforts on language detection for unsupported languages and translation across all languages for reviewers, we consider the impact and scale of our interventions and will provide further details on these in future updates.

Meta should ensure that content containing an unsupported language, even if mixed with supported languages, is routed to agnostic review. This includes providing reviewers with the option to re-route content containing an unsupported language to agnostic review.

The Board will consider this recommendation implemented when Meta provides the Board with data on the successful implementation of this routing option for reviewers.

Our commitment: We will continue to allow content reviewers to route content to the appropriate language queue, which provides more optimization than a language agnostic queue. This does not, however, preclude content reviewers from routing content to language-agnostic queues when most appropriate. Additionally, we’ll continue to work on our reviewer location strategy and language identification and translation technology to improve performance in unsupported languages.

Considerations: Meta has already enabled in its review tool the option to route content to other language review or agnostic teams. This allows reviewers to redirect unsupported, mis-routed, or content with a mix of languages to the adequate team. Currently, when there is a mix of languages in a post, reviewers will review the parts they understand and route the content with a summary and assessment in the notes for other reviewers' consideration during their review.

In the case of unsupported languages where reviewers don’t speak the same language in the content, it is often more effective to have content review in regional queues where content reviewers will still have local regional context. In agnostic queues, the reviewers would have even less language and regional context, and as a result would not be as effective in their review of content in unsupported languages. As a result, we continue to work towards having adequate regional representation in our reviewer location strategy.

In addition to continuously improving the accuracy of our translations across languages, we seek to ensure that agnostic review is best positioned for accurate enforcement by pursuing efforts to align agnostic review with regional relevance. Through our continued assessments of agnostic review processes, we learn that human reviewers with the relevant regional context can leverage their expertise to make enforcement decisions, even if they don’t speak the language. Configuring our review queues to accurately reflect this nuance is a large undertaking which requires various considerations and extensive investments in resources, however our teams continue to strategically pursue means to support regionally specialized agnostic review through routing tools and updates to reviewer training.

We will also continue to work on our language identification and translation technology, allowing reviewers to better identify the language of the content and provide accurate translation for assessment.