前沿的人工智能治理

Executive Summary As artificial intelligence diffuses throughout society, policymakers are faced with thechallenge of how best to govern the technology amid uncertainty over thefuture ofAIdevelopment. To meet this challenge, many stakeholders have put forth variousproposals aimed at shaping AIgovernance approaches.This report outlines ananalytic approach to help policymakers make sense of such proposals and takesteps to govern AI systems while preserving future decision-making flexibility.Ourapproach involves analyzing common assumptions acrossvariousproposals(as theseassumptions are foundational elements for the success of multiple proposals),as wellas unique assumptionsacross individual proposals,by answering three questions: 1. What risks areimportant to mitigate and who should have primary oversight offrontier AI? 2. Who is delegated tasks and able to play a role? 3. Would the proposed mechanisms or tools actually achieve the proposal’sobjectives? We apply this analytic approach to five U.S.-centric AI governance proposals thatoriginate from industry, academia, civil society, and thefederal andstate governments.These proposals are generally aimed at governing frontier AI systems, which possesscutting-edge capabilities and therefore pose some of the most challenging questionsfor AI governance.Our analysis reveals that most proposals view AI-enabling talentand AI processes and frameworks as important enablers of AI governance. However,proposals lack consensus regarding thetechniques that are most effective atmitigating AI risks and harms. Ouranalysisalso bearslessonsthat are broadly applicable topolicymakers seeking toanalyze anyproposal.Our case studies demonstrate that1)policymakers shouldleverage proposals’assumptions to more precisely understand disagreements andshared views among stakeholdersand2)policymakers can take action in anuncertain and rapidly changing environment by addressing common assumptionsacross proposals. By adopting our analytic approach,U.S.policymakers can moveaway from rhetorical debates about AI governance and better prepare the UnitedStates for a range of possibleAIfutures. Table of Contents Executive Summary................................................................................................................................1Introduction...............................................................................................................................................3Report Scope............................................................................................................................................5Guiding Questions..............................................................................................................................5Selection Criteria.................................................................................................................................6Limitations............................................................................................................................................6Proposal Case Studies..........................................................................................................................7Industry: OpenAI’s Approach to Frontier Risk...........................................................................7Civil Society: AI Now Institute’s Zero Trust AI Governance..................................................9Academia: Frontier AI Regulation: Managing Emerging Risks to Public Safety.........11State Government: California’s Safeand Secure Innovation for Frontier ArtificialIntelligence Models Act.................................................................................................................13Federal Government: Framework to Mitigate AI-Enabled Extreme Risks....................17Analysis...................................................................................................................................................19Deriving Assumptions....................................................................................................................19Shared Assumptions......................................................................................................................20What Risks Are Important to Mitigate And Who Should Have PrimaryOversight of Frontier AI?...........................................................................................................21Who Is Delegated Tasksand Able to Play a Role?.........................................................22Would the Proposed Mechanismsor Tools Actually Achieve the Proposal’sObjectives?....................................................................................................................................24Unique Assumptions......................................................................................................................30Discussion.........................................................................