AI Audits Must Move Beyond Fairness Scores to Prevent 'Bias Laundering'

Current AI auditing practices are failing to catch fundamental injustices because they prioritize technical consistency over political legitimacy. While regulators increasingly rely on these audits to ensure safety, a critical gap remains: the failure to question whether a system's objective is legitimate in the first place.

Tech Policy Press argues that most audits focus on "fairness scores," which measure error rates and technical consistency across different groups. To counter this, the outlet proposes a "power test." This framework examines who defines the AI's goals, who controls the system, how benefits and burdens are distributed, and who possesses the authority to contest decisions. The goal is to prevent "bias laundering," a process where political choices are translated into technical targets and then certified as neutral by an audit.

The Cost of Technical Neutrality

The danger of relying solely on technical specifications is illustrated by a 2019 study published in Science. Researchers found that a widely used health care algorithm was designed to predict future health care costs rather than actual medical need. Because less money is historically spent on Black patients, the algorithm assigned them lower risk scores than white patients who were equally sick.

This technical choice had devastating real-world consequences. The study revealed that if the algorithm had been designed to measure medical need instead of cost, the proportion of Black patients selected for additional support would have jumped from 17.7% to 46.5%. In this case, the system may have functioned perfectly according to its technical specification—predicting costs accurately—while remaining fundamentally unjust.

The Regulatory Gap

This issue arrives as AI regulation becomes more formalized. New York City has implemented bias audits for employment tools, the EU AI Act introduces conformity assessments, and the NIST AI Risk Management Framework provides a blueprint for safety. However, these interventions typically occur after a system's objective has already been established. By certifying a system based on how well it meets a flawed goal, audits risk providing a veneer of legitimacy to unequal policies.

Why Power Tests Matter

Integrating a power test into the auditing process shifts the focus from mathematical parity to democratic accountability. As the Tech Policy Press author notes, "A fairness score asks whether a system treats groups consistently. A power test asks who has the authority to define consistency, and who has the power to challenge it."

In high-stakes sectors such as policing, welfare, and healthcare, the difference between a technical error and a policy choice is profound. Without examining the power dynamics behind the code, audits may simply validate systems that are technically accurate but socially harmful.

The Path Forward

Moving forward, the industry must determine how to make the policy choices behind AI specifications visible and contestable. The challenge remains in how to standardize these "power tests" so they can be integrated into existing regulatory frameworks. Until audits address who holds the power to set the targets, the risk of bias laundering will continue to persist in automated decision-making.