AI智能体伦理治理研究报告（摘要版）

(SUMMARIZED VERSION) RESEARCHREPORT July 2026 CONTENTS Preface The Basic Logic of Ethical Governance for Al Agents 1.1 Technical Foundations and Operational Characteristics of Al Agents...011.2TypologyandCapabilityLevelsofAlAgents.021.3Social Embeddedness ofAlAgents and theNeedforEthical Governance...31.4OverviewofGlobal GovernancePolicie.04 Main Ethical lssues Triggered by Al Agents 2.1Physical Safetyand Health Risks.082.2PrivacyProtectionandData Ethics Risks..092.3 Human Dignity and Human-Machine Relationship Risks.102.4 Fair Use and Differential Impact Risks.122.5 Spatial Order and Public SafetyRisks.132.6ResponsibilityAttribution andDamage Relief Risks14 Challenges in Applying Rules to the Ethical Governance ofArtificial Intelligence Agents 3.1The Boundaries for ApplyingData GovernanceRules inContinuous InteractionScenarios Remain Unclear.173.2TheExpansionofAutonomousActionCapabilitiesPlacesNewPressureonGovernanceRules..193.3PlatformGovernanceRules Showan ImbalanceinResponsibilityAllocation inAgentEcosystemOperations.203.4 Ethical ReviewMechanisms Still ProvideInsufficient CoverageforHigh-RiskDynamicApplications.223.5Tool-InterfaceRules Need StrongerConstraints on the RiskofUltraViresActions....233.6Responsibility andRemedy Rules HaveInstitutional Shortcomings in AllocatingtheConsequences of Agent Behavior....24 Constructing a Framework forthe Ethical Governance of Al中Agents 4.1 Establishing theHuman-Centered Ethical GovernancePrinciple for Agents....264.2Establishing a Risk-Based TieredGovernance StructureforAgents......274.3Constructing Ethical GovernanceMechanismsThroughoutthe Lifecycle....284.4Improving BehaviorControlMechanisms ConcerningTool Invocation....4.5EstablishingCollaborativeGovernanceMechanismsforMulti-Agent Systems....314.6OptimizingResponsibilityAllocationMechanisms inAgentEthicalGovernance....32 05Implementation Mechanism of Ethical Governance forArtificial Intelligence Agents 5.1ImprovingtheStandards frameworkforEthicalGovernanceofAlAgents.......345.2EstablishinganEthicalRiskAssessmentSystemforAlagents355.3StrengtheningComplianceManagement intheOperationProcessofAlagents....365.4ImprovingUserRightsProtection andRemedyMechanisms.5.5Promotingthe Synergyof EthicalandTechnical Governance...385.6Buildinga Multi-StakeholderCollaborativeGovernanceFramework.c Conclusion.43 Preface In recent years, generative artificial intelligence has been shifting fromcontentgenerationtowardtaskexecution.Alagentscapableof planning,tool useenvironmental perception,sustained interaction,andautonomous executionhaveenteredofficework,manufacturing,transportation,healthcare,andpublic-servicesettings.Whereas conventionalgenerativeAlmainly respondsto one-off instructions,anAlagentdecomposesobjectivesintotasks,formulatesplans,invokestools,andadjusts its actions as the environment changes.Byintegrating a largelanguage modelwith memory, planning, and tool modules, it evolves from an information-generationtoolintoanaction-orientedsystem. Governanceof generativeAl haslargelyfocusedoncontent authenticity,datacompliance,and algorithmic bias.Al agents can read data,run programs,controldevices, and collaborate with other systems, extending their effects to individual rights,markettransactions,institutional operations,and publicadministration.In high-impactfields such as healthcare, transportation, and finance, governance therefore reachesfundamental values including life and health, human dignity,fairness and justice, socialtrust, andthepublic interest.Chinahas adopted rulesonalgorithmicrecommendation,deep synthesis,generativeAl services,and ethical review,creating an initial frameworkcenteredondata,algorithms,platforms,ethics,and responsibility;the European Unionandinternationalorganizations arelikewiseadvancingrisk-basedgovernance. However, most existing rules were designed around data processing, contentgeneration, and platform services, and do not adequately cover the long-termoperation,dynamic decision-making,and tool use of Al agents.Their conduct mayinvolvemodeldevelopers,platforms,tool providers,deployingorganizations,andusers,makingtheallocationof responsibilitydifficult.Moreover,risks often emerge during liveoperation;becauseactionpathwaysandoutcomesareuncertain,exanteassessmentand static regulation alone are no longer sufficient.This report examines thetechnicalcharacteristics,ethical risks,and institutionalfit of Al agents,identifies their corecapabilities,analyzes issues of safety,privacy,fairness,public order,andresponsibility,andproposesacorrespondinggovernanceframework. The Basic Logic of Ethical Governancefor Al Agents Al agentscanperceive,decide,invoketools,and act in pursuit of agoal.Clarifying their technical foundations, capability structure, and modes ofsocial embeddedness is therefore a prerequisite for analyzing ethical risks andgovernance pathways. l 1.1 Technical Foundations and Operational Characteristicsof Al Agents Al agents arebuilt onlarge models,but theirdefining distin