Abstract

Social media platforms review billions of new posts every day and remove ones that involve illegal content (e.g., exploitative or copyrighted material). Given the scale of this setting, responsive and reliable decision-making is operationally challenging. As a result, social media platforms invest billions of dollars on an AI–human pipeline, in which black-box AI models complement thousands of human reviewers. This talk focuses on a recent collaboration with a major social media platform to develop better scheduling algorithms for its human review system. The cost of delaying the human review of a malicious post is proportional to the number of views the post receives over time. A key challenge is that the number of views is ex-ante uncertain, which the existing queueing literature does not capture. To tackle this challenge, we introduce a new queueing model, where the uncertainty of post views resolves over time. On the theoretical front, we develop an asymptotically optimal algorithm. Moreover, simulations based on real data show that the algorithm consistently outperforms status quo heuristics.

A preprint of the corresponding paper can be found here: https://arxiv.org/pdf/2505.21331.

Video Recording