Measurement as governance in and for responsible AI

Measurement choices encode hidden governance decisions: The paper argues that how we measure things in AI systems (like fairness or risk) isn't neutral—it implicitly makes governance decisions about "which systems are fair, which individuals belong in which categories." These decisions are often obscured from public view, making them difficult to scrutinize or challenge.

Mismatches between what we measure and what we intend to measure create real harms: Fairness-related harms emerge when there's a gap between the concept we claim to be measuring (like "creditworthiness") and what the system actually measures. Understanding this distinction is critical for identifying where AI systems can unfairly disadvantage people.

"Content validity" and "consequential validity" reveal feedback loops in social categorization: The paper highlights two measurement concepts that help expose how AI systems don't just measure social categories—they actively construct and reinforce them. For example, a system measuring "recidivism risk" may create feedback loops that entrench certain groups in criminal justice categories.

Measurement processes are where values get embedded in AI systems: Social, cultural, and political values are implicitly encoded during the measurement process in sociotechnical systems. Making this process visible and explicit is essential for understanding whose values are being prioritized and how.

Treating measurement as governance enables better oversight: By recognizing measurement as a governance mechanism rather than a technical detail, organizations and regulators can design more effective interventions and accountability mechanisms for responsible AI.