Mandating Transparency: A Regulatory Framework for Algorithmic Bias and Opacity
Kaichen Zhang · Applied and Computational Engineering · 2025
Algorithmic bias and opacity pose threats to the global population by damaging social equity, as demonstrated by racial disparities in facial recognition accuracy and healthcare prediction algorithms. This is mainly because most algorithms are black boxes---few know how they operate; sometimes this is intentional to maximize company profits. While ethical principles lack enforceability, current regulatory approaches like the EU AI Act focus primarily on risk and privacy but fail to identify and address biases directly. A disclosure-based regulatory framework adapted from the US FDA's Total Product Lifecycle (TPLC), originally designed for medical algorithms, is a favorable alternative. Extended TPLC has already been proposed, which lists out potential biases throughout the entire implementation of medical algorithms. As a successfully applied visual diagnostic AI, the IDx-DR demonstrates the effectiveness of companies referencing the TPLC to identify, address, and publicly disclose their handling of biases throughout an algorithm's lifecycle. This shows the potential of this framework in resolving algorithmic bias and corporate secrecy ("black boxes") in all fields. However, biases and disclosure requirements of companies need to be generalized for wide applications. The extended TPLC should thus be simplified, involving only three stages: design, development, and deployment. Its effectiveness is demonstrated through three examples: flawed teacher evaluation algorithms (design), specialized legal AI development (development), and biased healthcare risk score monitoring (deployment). While this framework promotes algorithm transparency and accountability without stifling innovation, its effectiveness can be enhanced if governments mandate disclosure accordingly and coordinate globally to prevent corporate regulatory arbitrage.