Embracing Contradiction: Theoretical Inconsistency Will Not Impede the Road of Building Responsible AI Systems

Contradictions in AI ethics metrics are actually beneficial, not problematic. The paper argues that conflicting fairness definitions and tradeoffs between accuracy and privacy should be viewed as features rather than flaws, because they help represent the genuine complexity of responsible AI rather than forcing artificial simplification.

Multiple competing metrics better represent diverse stakeholder values. By maintaining potentially contradictory metrics simultaneously, AI systems can honor different moral perspectives and stakeholder concerns rather than privileging a single ethical viewpoint—what the authors call "Normative Pluralism."

Using conflicting metrics captures ethical concepts more completely. The paper contends that relying on multiple, sometimes contradictory measures preserves more information about complex ethical ideas (like fairness) than trying to reduce them to a single definition, maintaining what they term "Epistemological Completeness."

Balancing competing objectives can actually improve AI performance. Jointly optimizing for theoretically conflicting goals acts as a form of regularization that prevents overfitting to any single metric, potentially leading to AI systems that perform better and more robustly in real-world conditions.

The field should stop trying to eliminate inconsistency and instead define acceptable thresholds. Rather than attempting to resolve theoretical contradictions in responsible AI, practitioners should focus on characterizing which levels of inconsistency are acceptable and understanding how to work productively within those constraints.