通往負責任的 AI 之路：新興的最佳實踐和監管觀點.pdf

1、P A N E L D I S C U S S I O N The Road to Accountable AI:Emerging Best Practices and the Regulatory View Bojana BellamyPresident,Centre for Information Policy Leadership(CIPL)Kate CharletGlobal Head of Privacy,Safety&Security Policy,GoogleProf.Dieter KugelmannPresident,Authority for Data Protection

2、and the Freedom of Information,Rhineland-Palatinate,GermanyDr.Mark LeiserAssitant Professor,Vrije Universiteit,Amsterdam Law&Technology InstituteACCOUNTABLE AI IN PRACTICEPANEL DISCUSSION THE ROAD TO ACCOUNTABLE AI Building Accountable AI Programs:Mapping Emerging Practices to the CIPL Accountabilit

3、y Framework CIPL Paper 2024Provides concrete empirical evidence from accountable AI programs being deployed on the groundShares best practices and lessons learned from leading global organizationsBuilds global consensus on accountability in AI governancePromotes accountability as an effective strate

4、gy for the responsible development and deployment of AIInforms the global debate on AI regulation,governance,oversight,and enforcementAccountable AI Programs Mapped to CIPL Accountability FrameworkAccountabilityEffective compliance,business sustainability,protection for individualsRisk AssessmentPol

5、icies and ProceduresTransparencyTraining and AwarenessMonitoring and VerificationResponse and EnforcementLeadership and OversightImplementing protective,mitigating measures proportional to the likelihood and severity of the risks of harm while enabling the benefits of AI technologiesCreating effecti

6、ve governance structures,appointing appropriate personnel to oversee them,and promoting awareness and support across all functionsEstablishing internal,written AI policies and procedures that operationalize ethical principles,standards,and legal requirements into concrete actions and controlsProvidi

7、ng contextualized,relevant information and notices to stakeholders regarding organizational practices and offered AI technologiesProviding employee training to ensure awareness of an organizations AI policies and procedures and understanding of relevant topicsEnsuring internal implementation,complia

8、nce,and effectiveness of an organizations AI programAddressing incidents of noncompliance or security and data breaches,and responding to inquiries from customers,regulators,or auditors General findings from CIPLs Report on Accountable AI ProgramsTop 10 Findings of Accountable OrganizationsAI transf

9、ormation,coupled with accountability in AI,is a top priority and business imperative1Accountable governance of AI is a smart business investment for long-term sustainable and competitive business2“Tone from the top”is crucial for responsible AI programs3Guidance from regulators and policymakers is w

10、elcome as organizations prepare to implement emerging standards and regulations4A risk-based,technology-agnostic approach is the most effective and appropriate approach for AI governance5Convergence around common AI-related terminology is urgent and essential6Organizations are adapting and updating

11、their governance frameworks to address new issues and risks,including those raised by generative AI(GenAI)7Multidisciplinary and diverse teams are the foundation for building and implementing accountable AI governance programs8There is great value in building consensus on the appropriate elements of

12、 accountable AI governance and benchmarking against peers9Accountable development and deployment of AI is a continuous journey and an iterative process10Mapped to the CIPL Accountability ElementsExamples of Best Practices in Accountable AI Programs“Tone from the top”to demonstrate commitment to ethi

13、cs,values,and specific AI principlesAI/ethics/oversight boards/committeesResponsible AI lead/officerCentralized governance framework with flexibility within internal teams Multidisciplinary,cross-functional AI teamsDeveloping algorithmic/fairness impact assessmentsRequiring risk assessments througho

14、ut AI lifecycle Risk-based approach based on likelihood&severity of risksCreating risk taxonomy Documenting considerations&tradeoffs PETs/anonymization techniquesSpecifying use of protected data in model trainingIncorporating“privacy,AI ethics,or security by design”principlesInternal glossary of AI-

15、related termsDefining escalation steps for reviewing&reporting high-risk AI Ideation phase with all relevant internal stakeholders Pilot testing AI models before releaseDue diligence/self-assessment checklists for business partners using toolsTranslating internal principles into third-party agreemen

16、ts Cleaning data sets through automated/manual checks before trainingTailoring transparency measures for different audiences/contextsProvide counterfactual informationFactsheets and model cardsTiered transparency understanding customers expectations and deploying based on their readiness to embrace

17、AIConsidering accessibility needs Benchmarking opportunities,public engagement,regulatory sandboxes Focused ethics training for technicians(e.g.,limiting&addressing bias)Cross-functional training(e.g.,privacy professionals and AI engineers)Incentivizing compliance with ethics training(e.g.,promotion

18、s,bonuses,pay raises)Compilation of AI case studiesTailored ethics&fairness training Ongoing monitoring,validation&checks throughout AI lifecycle Human in the loop in design,oversight&redress Human audit of input&output Red-teaming&adversarial testing Monitoring the data ecosystem Understanding wher

19、e AI is being deployed Human review of individual decisionsRedress mechanisms to remedy AI decisionsInternal supervision of AI deploymentRedress through a human,not a botCommunication channels for internal(e.g.,employees)&external(e.g.,end users,customers,regulators)reporting,feedback,complaints,req

20、uestsPolicies and ProceduresLeadership and Oversight Risk AssessmentTransparencyTraining and AwarenessMonitoring and VerificationResponse and EnforcementINTERPRETING AND EVOLVING THE GDPR IN THE CONTEXT OF AIPANEL DISCUSSION THE ROAD TO ACCOUNTABLE AI AI and Data Protection PrinciplesCIPLs Report on

21、 AI and Data Protection-https:/bit.ly/2QUP2xyData Protection RequirementsArtificial IntelligenceTensions To ResolveConsentInsufficient/limited variety of legal bases may undermine full range and stages of AIData minimisationPurpose specification and limitationTransparencyRetention limitationIndividu

22、al rightsLegal basis for processingNot practical to obtain consent for the processing of personal data(including sensitive data)Needs sufficient volumes and diversity of data for research,analysis,operation,training and to avoid biasUses data for new and unforeseen purposes beyond original scopeMay

23、produce unexplainable and unanticipated outcomes;hard to provide meaningful noticeNeeds to retain data for AI training,traceability,audit and oversight Difficult to facilitate access,correction,deletion or explanation of the logic involved Rules on automated decision-making Automated decision-making

24、 capabilities are inherent to AICross border data transfer restrictionsNeeds to use diverse and geographically disperse dataPOLICY AND REGULATORY RECOMMENDATIONS PANEL DISCUSSION THE ROAD TO ACCOUNTABLE AI Did you enjoy this session?Is there any way we could make it better?Let us know by filling out

25、 a speaker evaluation.1.Open the Cvent Events app.2.Enter IAPP AIGG24(case and space sensitive)in search bar.3.Tap Schedule on the bottom navigation bar.4.Find this session.Click Rate this Session within the description.5.Once youve answered all three questions,tap Done.How Did Things Go?(We Really Want To Know)