Machine learning (ML) is reshaping how societies make decisions, prompting a growing focus on policy, regulation, and institutional design. Two contrasting governance models dominate the debate. The "monocentric" approach advocates a single, central regulator or a unified international treaty that can coordinate standards, deter forum‑shopping, and provide clear accountability ("The monocentric vision imagines a single overarching authority…"), but it faces serious challenges. Because AI systems are opaque, regulators often rely on the very developers they must oversee, creating a risk of regulatory capture ("Regulatory capture … the oversight body comes to depend upon … those it is meant to govern").
An alternative is polycentric governance, which distributes authority across multiple, interacting centres. In this model, dyadic relationships between pairs of institutions establish shared protocols or mutual‑recognition agreements, allowing local adaptation while still enabling interoperability ("Polycentricity … its elementary unit is the dyad: two centres that establish a workable interface…"). This structure addresses the Hayekian knowledge problem: no single body can gather the dispersed, context‑specific knowledge needed to regulate a technology whose applications are highly heterogeneous ("a single centre cannot assemble the dispersed, context‑specific knowledge…").
Current evidence shows that both public‑sector and private actors are experimenting with risk‑based frameworks. AWS highlights its support for standards such as ISO 42001 and collaborates with initiatives like the G7 AI Hiroshima Process to develop regulatory structures that protect civil rights while fostering innovation ("AWS supports the development of effective risk‑based regulatory frameworks for AI…"). Similarly, public‑policy tools such as algorithmic impact assessments, risk‑based evaluations, and post‑deployment monitoring are being adopted to detect bias and ensure fairness ("algorithmic impact assessments, risk‑based evaluation procedures, and post‑deployment monitoring … seek to identify potential disparate impacts before deployment").
Regulatory activity is already underway. New York City enacted the United States' first algorithmic accountability law in 2017, creating a task force to increase transparency and address harms ("New York City passed the first algorithmic accountability bill … requiring creation of a task force to share information on agency automated decision systems"). More recently, Executive Order 14110 (October 30, 2023) mandates a coordinated government‑wide strategy to develop trustworthy AI and mitigate risks such as discrimination and security threats ("Executive Order 14110 … emphasizes safe, secure, and trustworthy development and use of AI, outlining a coordinated, government‑wide approach…").
Trade‑offs and risks include the tension between innovation and oversight, the potential for capture in centralized regimes, and the difficulty of maintaining consistent standards across a polycentric landscape. Practical implications for stakeholders are clear: developers should embed documentation, impact assessments, and monitoring into the ML lifecycle; policymakers must balance uniform rules with flexible, local mechanisms; and institutions should design governance structures that can evolve as AI technologies and market dynamics change. [1] [2] [3]