Abstract
Artificial intelligence (AI) is beginning to participate in producing the judgment for which fiduciaries are legally responsible. This Article examines that problem in the setting of registered investment advisers' proxy voting under Rule 206(4)-6 and the Advisers Act's fiduciary standard. AI-assisted proxy voting is not merely a new technological tool; in some configurations, AI can gather information, apply voting policies, generate recommendations, explain those recommendations, or cause votes to be cast without individualized human review. The central question is therefore not whether advisers may use AI, but when AI-assisted reliance becomes AI-driven substitution.
The duties that govern that question are familiar: care, loyalty, retained responsibility, and the recordkeeping and compliance architecture that makes voting decisions reviewable. What AI changes is not the content of those duties but the operational account of how they are satisfied when part of the analytic process has moved into a computational system.
The Article makes three principal contributions. First, it identifies the fiduciary principle that AI places under pressure: a fiduciary may rely on outside inputs but may not substitute those inputs for its own judgment. Second, it reconceives reviewability as reconstructability, recasting the inquiry from whether the adviser can inspect a model's internal reasoning to whether it can reconstruct the inputs, instructions, policy, output, human review, conflict controls, and records that produced the vote. Third, it offers a four-mode typology of AI use in proxy voting: research assistant, policy-execution tool, recommendation engine, and autonomous or default voter. The typology maps increasing AI participation in voting judgment onto increasingly demanding fiduciary controls. Together, these contributions yield a framework, organized around ex-ante governance, use-stage controls, and ongoing monitoring, for determining when AI-assisted voting remains consistent with fiduciary judgment.
Download This Paper