《世界经济论坛:2023Presidio关于负责任的生成式AI的建议白皮书(英文版)(12页).pdf》由会员分享,可在线阅读,更多相关《世界经济论坛:2023Presidio关于负责任的生成式AI的建议白皮书(英文版)(12页).pdf(12页珍藏版)》请在三个皮匠报告上搜索。
1、The Presidio Recommendations on Responsible Generative AIJ U N E 2 0 2 3In collaboration with AI CommonsThe Presidio Recommendations on Responsible Generative AI2IntroductionGenerative artificial intelligence(AI)has the potential to transform industries and society by boosting innovation and empower
2、ing individuals across diverse fields,from arts to scientific research.To ensure a positive future,it is crucial to prioritize responsible design and release practices from the beginning.As generative AI continues to advance at an unprecedented pace,the need for collaboration among stakeholders to e
3、nsure that AI serves as a force for good has become increasingly urgent.On 26-28 April 2023,the summit“Responsible AI Leadership:A Global Summit on Generative AI”took place at the World Economic Forums Centre for the Fourth Industrial Revolution based in the Presidio in San Francisco,USA.The event w
4、as hosted by the Forum in partnership with AI Commons to guide technical experts and policy-makers on the responsible development and governance of generative AI systems.The summit emphasized the importance of open innovation and international collaboration as essential enablers for responsible gene
5、rative AI.The focus was on moving beyond insightful discussions to generate actionable and practical recommendations for various AI stakeholders that could significantly influence the design,construction and deployment of generative AI.Over 100 AI thought leaders and practitioners participated in th
6、e summit,including chief scientific officers,responsible AI and ethics leads,academic leaders,AI entrepreneurs,policy-makers,tech investors and members of civil society.Participants engaged in discussions on numerous aspects of generative AIs design,development,release and societal impact,and delibe
7、rated on key recommendations.These recommendations emerged from interactive panel discussions and working sessions through a bottom-up process,with participants reaching consensus on critical areas related to the governance of generative AI.This summary presents a set of 30 action-oriented recommend
8、ations aimed at guiding generative AI towards meaningful human progress.The recommendations address three key themes that cover the entire life cycle of generative AI:responsible development and release;open innovation and international collaboration;and social progress.By implementing these recomme
9、ndations,stakeholders can navigate the complexities of AI development and harness its potential responsibly and ethically.Join us in shaping a more innovative,equitable and prosperous future that leverages the power of generative AI and mitigate its risks to benefit all.The Presidio Recommendations
10、on Responsible Generative AI3Responsible Development and Release of Generative AIThis section critically assesses the necessity to protect our society from unforeseen outcomes induced by the swiftly developing generative AI systems,and accordingly advocates for responsible strategies concerning thei
11、r development and deployment.These recommendations are intended for a broad spectrum of stakeholders-ranging from AI developers to policy-makers and users.The objective is to foster accountable and inclusive processes for AI development and deployment,thereby enhancing trust and transparency as gene
12、rative AI systems continue to proliferate.01 Establish precise and shared terminologyAll stakeholders are called upon to use precise terminology when discussing the design,development,evaluation and measurement of generative AI models capabilities,limitations and issues.It is the responsibility of e
13、xperts to define and standardize this language.As soon as a consensus is reached,consistent adoption of this terminology by all stakeholders is essential.This approach will boost clarity and promote effective communication,leading to a shared understanding among different parties.Ultimately,it will
14、facilitate the establishment of strong,standards,guidelines and regulations for a range of generative AI applications.02 Build public awareness of AI capabilities and their limitationsPublic and private stakeholders should prioritize the task of enhancing public understanding.This includes making th
15、e terminology related to generative AI models understandable to the general public.Additionally,stakeholders should inform users about the probabilistic(meaning their outputs are not deterministic but based on probability)and stochastic(implying their operation involves a degree of random behavior)n
16、ature of generative AI models,while setting accurate expectations for their performance.03 Focus on human values and preferencesThe challenge to align generative AI models with human values and preferences needs to be further acknowledged and addressed.Developers of AI systems should be engaged in d
17、iscussions about normative values and preferences when designing AI models.04 Encourage alignment and participationPublic and private sector stakeholders should recognize that AI systems necessitate quality feedback that is diverse and representative of the user base to be truly aligned.Policy-maker
18、s should promote the involvement of diverse stakeholders,including non-technical stakeholders,in AI research and development to ensure alignment with human values.AI developers should work to facilitate interactions and feedback from a broad range of participants to create a more inclusive and human
19、-centric development process.05 Uphold AI accountability with rigorous benchmarknig and use case-specific testing while exploring new metrics and standards AI developers should commit to the importance of not only holding models accountable against the highest established benchmarks,but also finding
20、 new metrics beyond traditional ones and towards other human-centric dimensions.Benchmarking should be complemented by application-specific and task-defined testing to ensure a comprehensive evaluation of generative AI models.The Presidio Recommendations on Responsible Generative AI406 Employ divers
21、e red teamsRed teaming,a method of critically analysing perspective to identify potential weaknesses,vulnerabilities and areas for improvement,should be integral from model design to application and release.Diversity here implies incorporating members from varied genders,backgrounds,experiences and
22、perspectives for a more comprehensive critique.The public and private sectors should implement frameworks and methodologies to facilitate thorough red teaming.07 Adopt transparent release strategiesProducers of AI should be held accountable to release AI models responsibly,making them available to t
23、he public without compromising safety.Responsible release strategies should be initiated upstream during project ideation and product design to ensure that potential risks are identified and mitigated throughout the development process.08 Enable user feedback Users should be empowered with robust co
24、ntrols that allow them to provide real-time feedback on model outputs.Additionally,it is relevant to enable users to have a comprehensive understanding of the limits and responsibilities associated with the generated content.09 Embed model and system traceabilityDevelopers and policy-makers should a
25、lign on the importance of creating formal evaluation and auditing structures surrounding traceability throughout the entire AI life cycle,from data provenance to training scenarios and post-implementation.10 Ensure content traceabilityTo increase transparency and accountability,companies developing
26、AI-generated content should be responsible for tracing how content is generated and documenting its provenance.This will help users discern the difference between human-generated and AI-generated content.11 Disclose non-human interactionIn virtual environments,humans should know whether they are int
27、eracting with a human or a machine.AI providers should develop mechanisms to support this,for example,via watermarking.12 Build human-AI trustTo build trust in AI systems,developers and companies should prioritize transparency,consistency,and meeting and managing user expectations.AI developers shou
28、ld be transparent in their processes and decision-making,providing users with an understanding of how they reach their results.By focusing on these aspects,AI developers can create systems that foster trust and facilitate positive human-AI interactions.13 Implement a step-by-step review processPolic
29、y-makers and businesses should create a step-by-step review process for AI models and products.This should be similar to the detailed checks used in clinical trials or car manufacturing,both before and after a product goes live.There should be an independent auditor or international agency to overse
30、e this to ensure uniform evaluations and continuous monitoring.To help limit potential risks and negative impacts,certification,or licensing system could be used.The Presidio Recommendations on Responsible Generative AI514 Develop comprehensive,multi-level measurement frameworksPolicy-makers should
31、emphasize ongoing efforts and incentivize developers and standardization bodies to focus on creating and employing measurement frameworks with an emphasis on socio-technical aspects rather than solely technical performance.15 Adopt sandbox processesAI developers,standard-setting bodies and regulator
32、s should cooperate on more flexible“sandbox”development environments along with new and associated processes of governance and oversight.Sandboxing could help build trust by demonstrating that AI systems have undergone rigorous testing and evaluation to ensure safety,reliability and compliance.16 Ad
33、apt to the evolving landscape of creativity and intellectual propertyWith generative AI impacting content creation,it is essential for policy-makers and legislators to re-examine and update copyright laws to enable appropriate attribution,and ethical and legal reuse of existing content.This section
34、focuses on the importance of sharing scientific knowledge and enhancing international collaboration.As frontier research capabilities tend to be concentrated in private sector companies in a select few countries,it is vital that academic researchers remain an integral part of the exploratory process
35、,while countries worldwide participate and influence the governance of generative AI systems.These recommendations are designed for a range of stakeholders,including researchers,AI developers,standard-setting bodies and policy-makers.The overarching goal is to cultivate transparency,accountability a
36、nd inclusivity in the development,implementation and governance of generative AI.17 Incentivize public-private research coordinationPublic and private stakeholders should actively work to design incentive structures that facilitate greater coordination between academic researchers and the private se
37、ctor throughout the technology development lifecycle.Possible mechanisms to be considered include joint research programmes,data-sharing protocols and joint IP ownership.18 Build a common registry of models,tools,benchmarks and best practicesProducers and researchers of generative AI should contribu
38、te to a common and open registry of source codes,models,datasets,tools,benchmarks and best practice guidelines,to be shared within the research community,in order to have a platform for academic and private sector collaboration to build future models and systems that are transparent and accountable
39、to the public.Open Innovation and International CollaborationThe Presidio Recommendations on Responsible Generative AI619 Support responsible open innovation and knowledge sharingPolicy-makers and AI providers should contribute to frameworks to democratize AI through responsible sharing of resources
40、,including data,source code,models and research findings;also encourage the sharing certification processes,ensuring transparency and trust among stakeholders.A public-private long-term initiative could be developed to build public-facing platforms that provide open access to compute,data and pre-tr
41、ained models.This platform could be treated as a digital public good,and usage could be promoted across borders.20 Enhance international collaboration on AI standardsStandard bodies must foster international collaboration on AI standards,ensuring the participation of all AI stakeholders,including al
42、l geographical locations.21 Establish a global AI governance initiativeTo address the challenges and potential risks posed by AI technologies,policy-makers should consider devoting efforts towards creating a global AI governance initiative.This initiative should bring together experts from a wide ar
43、ray of fields.The key focus should be on promoting global understanding of responsible generative AI,ensuring broad inclusion,facilitating access to infrastructure,and fostering collaboration to harmonize response structures at the national level against AI challenges and risks.Social progressThis s
44、ection examines the hurdles tied to AI-driven transformations,spanning from workforce transitions to educational shifts,as well as the necessity of championing AI for societal benefit and advocating for equitable AI access in developing nations.The recommendations are intended for a broad array of s
45、takeholders,including educational institutions,community organizations,corporations,individuals,policy-makers and governments.The primary objective is to cultivate a society that is more informed,engaged and resilient in the face of these emerging changes.22 Prioritize social progress in generative
46、AI development and adoptionAll stakeholders must ensure that the technologys societal implications remain front and centre.This involves a focus beyond technical proficiency towards the technologys role in enhancing social progress.Comprehensive support must be provided to communities and workers af
47、fected by the shift to an AI-enabled society,encompassing learning initiatives,guidance on surmounting generative AI-specific challenges and assistance in navigating the ethical,social and technical shifts inherent in an AI-influenced environment with an active participation of workers throughout th
48、e process.23 Drive AI literacy across societyEducational bodies and community institutions must take the initiative to increase AI literacy among the general public.A proactive approach is needed to demystify generative AI tools,outline their potential uses and discuss their ethical implications.Thi
49、s will empower individuals to better understand,interact with and contribute to the evolving landscape of AI,fostering a more informed and participative society.The Presidio Recommendations on Responsible Generative AI724 Foster holistic thought approaches in AI-driven environmentsFoster diverse mod
50、es of thinking critical,computational and responsible to better equip society for the generative AI era.Encourage these core competencies across sectors and communities to empower individuals to engage critically with AI-generated content,understand the underlying technology and make responsible dec
51、isions about its use.25 Steer generative AIs transformative impactAddress the transformative influence of generative AI on societal systems.Understand its effect on human interactions,knowledge dissemination and evaluation mechanisms.Proactively adapt to the evolving landscape,supporting roles that
52、may transform due to generative AI,and explore innovative ways to evaluate its impacts within our rapidly evolving digital ecosystem,to harness its potential for driving positive societal transformation.26 Incentivize innovation for social goodPolicy-makers should encourage the development and imple
53、mentation of generative AI technologies that prioritize social good and address complex and unmet societal needs,such as in healthcare and climate change,to improve the overall quality of life.27 Address resource and infrastructure disparities Policy-makers should increase public investment in natio
54、nal and international research infrastructure.That includes work to ensure greater access to computing resources for researchers,especially those from underrepresented regions and institutions.The private sector is encouraged to contribute to the development of datasets and support governments in ma
55、king more resources available to researchers.28 Promote generative AI expertise within governmentsGovernments should invest in fostering AI expertise,ensuring an informed,effective and responsible approach to public policies and regulation of these transformative technologies.By leveraging mechanism
56、s such as targeted incentives,private sector collaborations,and exchange programs,governments can nurture AI talent.This commitment while expanding in-house AI proficiency is crucial in securing a future where these technologies advance societal progress and serve the public interest effectively.29
57、Increase equitable access to AI in developing countriesTo ensure that the benefits of generative AI technology are accessible to all,public and private stakeholders should focus on establishing initiatives that can provide support and resources at scale,particularly in developing countries where the
58、re may be limited access to digital infrastructures.Efforts should focus on providing resources,training,and expertise to make AI more accessible and inclusive,fostering national and international partnerships across sectors to promote diversity and inclusion in the development and deployment of gen
59、erative AI technology.30 Preserve cultural heritageAll stakeholders need to contribute to preserve cultural heritage.Public and private sector should invest in creating curated datasets and developing language models for underrepresented languages,leveraging the expertise of local communities and re
60、searchers and making them available.This will improve access to AI technologies to help preserve linguistic diversity and cultural heritage.The Presidio Recommendations on Responsible Generative AI8ContributorsBlaise AgueraVice-President and Fellow,Google Research,GoogleXavier AmatriainVice-Presiden
61、t of Engineering Product AI Strategy,LinkedIn CorporationStephen AugustusHead of Open Source,Cisco SystemsRicardo Baeza-YatesDirector of Research,Institute for Experiential AI,Northeastern UniversityAnthony BakHead of AI and Machine Learning,Palantir TechnologiesHouman BehzadiPresident and Chief Pro
62、duct Officer,C3 AI Kimmy Bettinger Expert and Knowledge Communities Lead,World Economic Forum Seth BergesonManager,AI and Emerging Technology,PwCJamie Berryhill Artificial Intelligence Policy Analyst,Organisation for Economic Co-operation and Development(OECD)Marc BoxserVice-President,Policy and Com
63、munications,Chegg Inc.Kirk BresnikerFellow and Chief Architect,Hewlett Packard Labs,Hewlett Packard EnterpriseJoanna BrysonProfessor of Ethics and Technology,Hertie SchoolSebastian Buckup Head of Network and Partnerships,Deputy Head.Centre for the Fourth Industrial Revolution;Member of the Executive
64、 Committee,World Economic ForumJill BursteinPrincipal Assessment Scientist,DuolingoSummit Steering CommitteeEsteban ArcauteHead of Responsible AI,Meta PlatformsYoshua Bengio Head of the Montreal Institute for Learning Algorithms,University of MontrealMona Diab Lead Responsible AI Research Scientist,
65、Meta PlatformsMichael KearnsFounding Director,Warren Center for Network and Data Sciences,University of PennsylvaniaHiroaki KitanoSenior Executive Vice-President and Chief Technology Officer;Chief Executive Officer,Sony Research,Sony Group CorporationYann LeCunVice-President and Chief AI Scientist,M
66、eta PlatformsPilar ManchnSenior Director of Engineering,GooglePeter NorvigDirector of Research,GoogleCathy Li Head of AI,Data and Metaverse;Deputy Head,Centre for the Fourth Industrial Revolution;Member of the Executive Committee,World Economic ForumBenjamin Larsen Lead,Artificial Intelligence and M
67、achine Learning,World Economic Forum Hubert Halop Lead,Artificial Intelligence and Machine Learning,World Economic Forum Lucia Velasco Lead,Artificial Intelligence and Machine Learning,World Economic ForumSummit Co-ChairsAmir Banifatemi Director,AI CommonsPascale Fung Chair Professor,Hong Kong Unive
68、rsity of Science&TechnologyFrancesca Rossi IBM Fellow and IBM AI Ethics Global Leader;AAAI PresidentJoaquin Quionero-Candela Technical Fellow for Artificial Intelligence,LinkedInAuthorsThe Presidio Recommendations on Responsible Generative AI9Cansu CancaAI Ethics Lead at Institute for Experiential A
69、I,Northeastern UniversityDiane ChangDirector of Data Science,Intuit Inc.Joshua CohenMember of the Faculty,Apple University,AppleDavid CoxDirector of Exploratory AI Research,IBM CorporationNatasha CramptonChief Responsible AI Officer,Microsoft CorporationJoris CyizereHead ad interim,Centre for the Fo
70、urth Industrial Revolution,RwandaUmeshwar DayalCorporate Chief Scientist,Senior Vice-President and Senior Fellow,Hitachi AmericaAnil Dewan Senior Advisor,US Department of Homeland SecurityDaniel Dobrygowski Head,Governance and Trust,World Economic ForumAnne Marie Engtoft LarsenTech Ambassador,Minist
71、ry of Foreign Affairs of DenmarkMojdeh EskandariFounder and President,Enovant FoundationAldo FaisalProfessor of Artificial Intelligence and Neuroscience,Imperial College LondonGilles FayadAdvisor,Institute of Electrical and Electronics EngineersRebecca FinlayChief Executive Officer,Partnership on AI
72、Kay Firth-ButterfieldExecutive Director,Centre for Trustworthy TechnologyGwenda FongDeputy Secretary(Development and Regulation,Ministry of Communications and Information of SingaporeEdward FuHead of Government Affairs,DuolingoKrishna GadeChief Executive Officer and Co-Founder,Fiddler LabsEugenio Ga
73、rcia Deputy Consul-General,San Francisco,Ministry of Foreign Affairs of BrazilTiffany GeorgievskiAI Attorney,Sony ResearchMatthew GravissChief Data Officer,US Department of StateTom GruberCo-Founder/Chief Technology Officer,Siri and Humanistic.aiPeter HallinanLeader,Responsible AI,Amazon Web Service
74、sRuimin HeChief Technology Advisor,Ministry of Communications and Information of SingaporeBrittan Heller Fellow,Digital Forensics Research Lab,The Atlantic CouncilCyrus HodesCo-Founder of AIGC Chain and Stability AI,Harvard Kennedy School of GovernmentBabak HodjatChief Technology Officer AI,Cognizan
75、t Technology Solutions US Corp.Jerremy HollandDirector of AI Research,AppleMatissa HollisterAssistant Professor of Organizational Behaviour,McGill UniversitySara HookerHead,Cohere for AIEric HorvitzChief Scientific Officer,MicrosoftXinghai HuHead of TikTok Data US,BytedanceAnil KamathFellow and Vice
76、-President AI/ML,Adobe SystemsVijay KarunamurthyVice-President of Engineering,Scale AIAnja KaspersenMember,Council on Extended Intelligence and Industry Activity on Life Science,Institute of Electrical and Electronics EngineersJeffrey LadishHead of AI Insights,Center for Humane TechnologyYolanda Lan
77、nquistDirector of AI Governance,The Future SocietyThe Presidio Recommendations on Responsible Generative AI10Federico LecumberryAssociated Professor,Universidad de la RepblicaChase Lochmiller Co-Founder and Chief Executive Officer,Crusoe Energy SystemsLeland LockhartDirector,Artificial Intelligence&
78、Machine Learning,Vista Equity PartnersDavid Luan Chief Executive Officer,Adept AIEmily McReynoldsSenior Fellow,Center for Responsible AI,New York UniversityRisto MiikkulainenProfessor of Computer Science,University of Texas,AustinSteven MillsPartner and Chief Artificial Intelligence Ethics Officer,B
79、oston Consulting GroupJoshua NewTechnology Policy Executive,IBM CorporationLoren NewmanGovernment Affairs Lead,World Economic ForumVaibhav PahwaProduct Manager,Platform Fairness and Responsible AI,TikTokGleb PapyshevPhD Candidate in Science and Technology Policy,The Hong Kong University of Science a
80、nd TechnologyVijay ParthasarathyHead of Artificial Intelligence and Machine Learning,Zoom Video CommunicationsJonnie PennAssistant Teaching Professor of AI Ethics and Society,University of CambridgeNazneen RajaniResearch Lead,Hugging FaceMartin RauchbauerCo-Director and Founder,Tech Diplomacy Networ
81、kStuart RussellProfessor of Computer Science,University of California,BerkeleySultan SaidovCo-Founder and President,Beamery Inc.Nayat Sanchez-PiChief Executive Officer,INRIA ChileSupheakmungkol Sarin Head of Data and Artificial Intelligence Ecosystems,World Economic ForumSilvio Savarese Executive Vi
82、ce-President,Chief Scientist,SalesforceYoav SchlesingerArchitect,Ethical AI Practice,SalesforceCraig ShankAdvisor,Responsible Artificial Intelligence InstituteJoanna ShieldsChief Executive Officer,BenevolentAIKaren SilvermanFounder and Chief Executive Officer,The Cantellus GroupSarvjeet SinghPrincip
83、al Engineer/Engineering Director,Google Research,GoogleNavrina SinghFounder and Chief Executive Officer,Credo AIUyi StewartChief Data and Technology Officer,data.orgJoAnn StonierChief Data Officer,MastercardMurali SubbaraoVice-President,AI Solution Success,ServiceNowCandace SueVice-President for Aca
84、demic Relations,CheggArun SundararajanHarold Price Professor of Entrepreneurship and Technology,Stern School of Business,New York UniversityJosephine TeoMinister of Communications and Information of Singapore Kellee TsaiDean of Humanities and Social Science,The Hong Kong University of Science and Te
85、chnologyThomas WolfCo-Founder,Hugging FaceAndrea WongHead of Platform Fairness,BytedanceLauren Woodman Chief Executive Officer,DataKindDaniel Wroblewski Managing Director,Head of Investment Science,Canada Pension Plan Investment BoardThe Presidio Recommendations on Responsible Generative AI11Alice X
86、iangGlobal Head of AI Ethics,Sony Research,Sony Group CorporationKevin YanceyStaff AI Research Engineer,DuolingoMasaru YarimeAssociate Professor,The Hong Kong University of Science and TechnologyGrace YeeDirector of Ethical Innovation,AI Ethics,AdobePolina ZvyaginaPrivacy and Data Policy Manager,AI/
87、ML Products,Responsible AI,Meta PlatformsWorld Economic Forum9193 route de la CapiteCH-1223 Cologny/GenevaSwitzerland Tel.:+41(0)22 869 1212Fax:+41(0)22 786 2744contactweforum.orgwww.weforum.orgThe World Economic Forum,committed to improving the state of the world,is the International Organization for Public-Private Cooperation.The Forum engages the foremost political,business and other leaders of society to shape global,regional and industry agendas.