Disciplining Deliberation: A Sociotechnical Perspective on Machine Learning Trade-offs

Trade-offs between accuracy and fairness/interpretability are not inevitable value conflicts. The paper challenges the common assumption that formal trade-offs in machine learning models directly translate to unavoidable tensions between social values, showing instead that these relationships depend on "critical choices and assumptions" made during AI development.

Context matters more than the models themselves. The author identifies three key considerations—validity and instrumental relevance, compositionality, and dynamics—that must be examined to understand what trade-offs actually mean for real-world social values, suggesting that the same technical trade-off can have different ethical implications depending on how it's implemented and deployed.

Sacrificing one model property can sometimes advance multiple values simultaneously. Contrary to the prevailing interpretation, the paper demonstrates that "judicious sacrifices in one model property for another can, in fact, promote both sets of corresponding values," meaning developers shouldn't assume they must choose between fairness and accuracy or interpretability and accuracy.

Current AI governance discussions may be based on incomplete assumptions. By showing that formal trade-offs don't necessarily correspond to value conflicts, this work suggests that policymakers and regulators should reconsider how they frame requirements and regulations around AI systems, potentially opening pathways to solutions previously thought impossible.

AI design involves social and technical choices, not just technical constraints. The sociotechnical perspective emphasizes that how we build and deploy AI systems reflects deliberate choices about values and priorities, not merely technical limitations, giving stakeholders more agency in achieving responsible AI outcomes.