Recalibrating Artificial Intelligence for Social Biases: Four New Priorities for Datasets, Models, and Metrics
Lucy Havens, Melissa Terras, Benjamin Bach, Beatrice Alex · 2026
This paper demonstrates the need for a recalibration of Artificial Intelligence (AI) and provides guidance on its implementation. Through a literature review, we identify priorities of predominant AI approaches and explain how they prevent the efficacy of efforts to mitigate harms from socially biased AI systems. We suggest alternative priorities to recalibrate AI and provide case studies to demonstrate their feasibility. Predominant approaches to AI prioritize quantity, convenience, efficiency, and universal thinking. Often, only after AI systems are developed and deployed are they scrutinized for biased behavior. We argue that to more effectively prevent harms from such biases, approaches to creating AI datasets, models, and metrics should instead prioritize quality, representativeness, accuracy, and situated thinking. While prior work has critiqued existing AI values and practices, this work rarely offers, let alone implements, an alternative. Building on these critiques, we draw upon critical theorizing and existing work that has taken alternative approaches to creating AI systems to develop practical, decision-making guidance on how to implement AI differently. The four new priorities, or recalibrations, we propose for guiding AI research and development encourage a proactive approach to mitigating harms from socially biased datasets and models. We developed the recalibrations with researchers, policymakers, and business leaders in mind, providing evidence of the possibility and efficacy of alternative approaches to AI innovation.