The Denominator Problem in AI Governance

2026-04-24 · Source: Tech Policy Press · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, AI Governance & Policy · Depth: Advanced, medium

Summary

The "denominator problem" in AI governance, identified by Michael A. Santoro, poses a fundamental measurement challenge to new AI incident reporting mandates from the EU, US states like New York and Colorado, and organizations like the OECD. This problem highlights the difficulty in interpreting AI harm counts (numerators) without knowing the total opportunities for harm (denominators). For instance, a doubling of reported AI harms could indicate increased failures, improved reporting, or simply expanded AI deployment. While autonomous vehicles have accessible denominators (e.g., miles driven), domains like deepfakes and AI-driven hiring lack clear metrics for total synthetic content generated or individual AI-shaped hiring decisions. The healthcare sector presents the most complex case, with various potential denominators (model inferences, AI-assisted clinical decisions, patient encounters involving AI) each yielding different rates and implications for responsibility, further complicated by the need for stratification by race, gender, and insurance status to prevent algorithmic bias.

Key takeaway

For CTOs and VPs of Engineering tasked with AI governance and risk management, recognizing and addressing the "denominator problem" is critical. Your teams must move beyond simple incident counts to establish measurable rates of harm, particularly in high-stakes domains like healthcare. This requires defining and tracking appropriate denominators, potentially through enhanced logging in health IT systems or mandatory disclosure of AI involvement in processes like hiring, to enable auditable safety outcomes and mitigate emerging AI-related liability risks.

Key insights

Interpreting AI incident counts requires a "denominator" to measure harm rates, not just raw occurrences.

Principles

• Numerator without denominator is uninterpretable.
• Ubiquity of AI is not visibility for measurement.
• Stratification is crucial for auditing fairness.

Method

Classify AI incident types into phases by estimating harm rates and deployment scale, moving beyond raw incident counting to analytically useful metrics.

In practice

• Track AI-assisted clinical decisions in EHRs.
• Mandate disclosure of AI participation in hiring.
• Stratify denominators by demographic factors.

Topics

• AI Governance
• Denominator Problem
• AI Incident Reporting
• Algorithmic Bias
• Healthcare AI

Best for: Executive, CTO, VP of Engineering/Data, Policy Maker, AI Ethicist, Legal Professional

Related on AIssential

• AI as a Justice-System Risk Multiplier: (1) implementation mistakes, (2) questionable “fit” to the real-world population, and (3) legal/ethical fragility around data and proxies. — Small algorithmic errors in justice systems scale into systemic distortions with severe societal and individual consequences.
• The Fairness Trade Off in Machine Learning and Its Health Care Application — Fairness in ML is a complex, multi-faceted dial, not a simple on/off switch, requiring careful definition and optimization.
• From Age Gates to Accountability in AI Design — Disparate impact doctrine offers a framework to assess AI systems for disproportionate harm to children, shifting focus from access to design.
• Research Sabotage in ML Codebases — Neither frontier LLMs nor LLM-assisted humans reliably detect research sabotage in ML codebases.
• The Trump Administration’s AI Policy Framework Has an Ideology. It Just Won't Admit It. — The Trump Administration's AI policy framework prioritizes unchecked innovation over critical safeguards, driven by an unstated ideological worldview.
• Towards AI epidemiology: a measurement standardisation framework for prospective risk detection — A framework enables "AI epidemiology" by standardizing expert-AI interactions for prospective risk detection without model internals.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Tech Policy Press.