Antagonistic Predictions

Why do we use models to predict the most likely outcome? Integrating model predictions into high-stakes decision-making pipelines can amplify biases and introduce flawed decisions when human caseworkers and models compete over the same prediction target. Learning to anticipate the most likely outcome risks reproducing biases embedded in historical data, while delivering such predictions in human-AI decision-making raises issues of cognitive biases and algorithmic aversion versus overreliance. In this perspective, I argue for Antagonistic Predictions, a structurally different paradigm where models are tasked to act as sparring partners that challenge human pre-conceptions and show reachable futures rather than anticipating the most common pattern. I envision instantiations of Antagonistic Predictions that can counteract historical disadvantages, social stereotyping, and cognitive biases and sketch potential implementations in predictive ML and generative AI settings.