ToxiSight: Human-AI Content Moderation

We sought to better understand how to build a collaborative tool that harmonizes strengths and weaknesses between human and AI in the goal to facilitate and improve content moderation for online video games. In this talk, we will demo the research conducted behind ToxiSight, an internal tool developed at Ubisoft La Forge, that leverages human nuance and AI scale to improve judgements on toxicity in online game chat. We will present a demo that showcases how a human and AI model will interact to provide all available context and knowledge gaps as well as simultaneously re-train and fine tune a model. We also introduce how we can integrate a multi-disciplinary approach in this endeavour by leaveraging key theory in psychology to learn more about the support humans need and the shortcomings of the AI model.