RECAST

With the widespread use of toxic language online, platforms are increasingly using automated systems that leverage advances in natural processing to automatically flag and remove comments. However, most systems, when detecting moderating language, do not provide feedback their users, let alone an avenue recourse for these users make actionable changes. We present our work, RECAST, interactive, open-sourced web tool visualizing models' predictions, while providing alternative suggestions flagged language. Our work also provides with a new path moderation tools. RECAST highlights text responsible classifying toxicity, allows interactively substitute potentially phrases neutral alternatives. examined effect via two large-scale user evaluations, found was highly effective at helping reduce toxicity as detected through model. Users gained stronger understanding underlying criterion used by black-box models, enabling transparency recourse. In addition, we focus on optimizing models instead own judgement (which is implied incentive goal deploying models), cease be classifiers compared human annotations. This opens discussion how detection should future online discourse.