世界经济论坛-组织中AI公平教育的整体指南（英）-2021.9-31正式版.ppt

《世界经济论坛-组织中AI公平教育的整体指南（英）-2021.9-31正式版.ppt》由会员分享，可在线阅读，更多相关《世界经济论坛-组织中AI公平教育的整体指南（英）-2021.9-31正式版.ppt（33页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、A Holistic Guide toApproaching AI FairnessEducation in OrganizationsW H I T E P A P E RS E P T E M B E R 2 0 2 1每日免每日免费获费获取取报告报告1、每日微信群内分享7+最新重磅报告；2、每日分享当日华华尔尔街日街日报报、金融时报；3、每周分享经济经济学人学人4、行研报告均为公开版，权利归原作者所有，起点财经仅分发做内部学习。扫一扫二维码关注公号回复：研究报告加入“起点财经”微信群。Cover:GettyImages/metamorworksInside:GettyImages/met

2、amorworks;GettyImages/Drazen;GettyImages/AndreyPopov;GettyImages/pixelfit;GettyImages/alvarez;GettyImages/anandaBGD;GettyImages/IgorKutyaevContents345ForewordExecutive summary1 Introduction6681.1 A brief introduction to AI fairness1.2 Why fairness?1.3 Approaching fairness education in a corporation1

3、11316182022242627292 Senior leadership3 Chief AI ethics officers4 Managers5 Build teams6 Business teams7 Policy teams8 ConclusionCase study AI and youthContributorsEndnotes 2021 World Economic Forum.All rightsreserved.No part of this publication may bereproduced or transmitted in any form or by anym

4、eans,including photocopying and recording,orby any information storage and retrieval system.DisclaimerThe Network of Global Future Councils is an invitation-only community that serves as a braintrust for the World Economic Forum and the world at large.This paper has been written by theWorld Economic

5、 Forum Global Future Council on AI for Humanity.The findings,interpretationsand conclusions expressed herein are a result of a collaborative process facilitated and endorsedby the World Economic Forum,but whose results do not necessarily represent the views of theForum,nor the entirety of its Member

6、s,Partners or other stakeholders,nor the individual GlobalFuture Council Members listed as contributors,or their organizations.A Holistic Guide to Approaching AI Fairness Education in Organizations2September 2021A Holistic Guide to ApproachingAI Fairness Education in OrganizationsForewordEmily RattK

7、ay Firth-ButterfieldHead of AI and Machine Learning,Project Specialist,AI andMachine LearningMember of the Executive CommitteeWorld Economic ForumManager,Global FutureCouncil on AI for HumanityWorld Economic ForumOver the past few years,many technologyThe GFC comprises 24 experts from aroundthe worl

8、d,who are making advancements inthis space.Representing many professional andcultural backgrounds,sectors and industries,the group recognized,and early on came to aconsensus on,the multidimensional nature ofAI fairness which precludes any one definitionof what“fair”AI looks like.For this reason,thef

9、ollowing report outlines a holistic approachto addressing AI fairness education in ancompanies and,frankly,many organizations thathad never thought of themselves as inherentlytechnological have begun to recognize theimportance of ethical and responsible development,design,deployment and use of artif

10、icial intelligence(AI).This can be attributed to the increased use ofAI in many commonplace organizational functions,such as marketing platforms,talent managementtools and search engines;and,by extension,theincreasingly frequent ethical dilemmas emerging fromthe way in which these AI applications ha

11、ve beencreated or used.Many organizations have developedprinciples regarding AI ethics,such as transparency,organization,which can be adapted to differentsector and industry contexts as necessary.Thereport draws on several collective values includingaccess,equality,equity and transparency.explainabi

12、lity,privacy,robustness and fairness.1Not every organization will have the resourcesnecessary to hire a dedicated team of AI ethicists,let alone experts focused on fairness specifically.Developing curricular materials to educate employeeson the potential implications of biased or unfair AI,as well a

13、s methodologies,tools and practices toaddress these implications,will require even greaterfinancial commitment and more resources.Withthis holistic look at the role and impact of differentmembers of a business in addressing AI fairness,we hope to provide options and a North Star forany organization

14、open to improving its practicesor creating products with a positive and equitableimpact on the larger population of the world.The Global Future Council(GFC)on AI for Humanitywas convened under the mandate of findingsolutions to critical issues of AI fairness.Somereaders may wonder why this council o

15、f expertsis primarily focused on fairness when there aremany other issues within the burgeoning field of“AI ethics”worthy of being further addressed.Withthis council convening in 2020 at a time marked bydeep economic and social unrest and injustice,theWorld Economic Forum hoped to shine a spotlighto

16、n fairness as an essential part of any future inwhich AI continues to be developed,deployed andused at scale.A Holistic Guide to Approaching AI Fairness Education in Organizations3Executive summaryAs organizations automate or augment theirdecision-making with AI,there is a high risk thatthe resultan

17、t decisions either create or reinforceunfair bias.The negative impact of bias andunfairness in AI does not affect individual victimsalone.Organizations that design,develop anddeploy AI can face serious repercussions such asbrand/reputational damage,negative sentimentamong employees,potential lawsuit

18、s or regulatorypenalties,and loss of trust from all stakeholders,including customers and the general public.This paper outlines six functions within an organizationthat have a particular role in operationalizing AIfairness:senior leadership,chief AI ethics officers,managers,build teams(data scientis

19、ts,developers,product designers,engineering teams etc.),businessteams(customer-facing teams such as sales,marketing and consulting)and policy teams.Eachsection of the report delineates the responsibilities ofthe team in contributing to AI fairness outcomes,aswell as the competencies and training tha

20、t should bemeasured and provided to the team to enable themto carry out their responsibilities.This report aims to address one important part oforganizations approach to AI fairness:educatingdifferent teams about the role they play in advancingAI fairness.Holistic learning and education onAI fairnes

21、s across an organization can driveemployees to understand the important role theyplay in contributing to better,more equitable andmore ethical use of AI.The report also includes recommendations on furtherefforts needed to improve AI fairness outcomesbeyond education,including defining an organizatio

22、nsfairness objectives,creating a supportive corporateculture and hiring diverse teams at all levels and partsof an organization.It ends with a case study on achild-centred approach to AI fairness,to help readersto contextualize the information presented throughout.A Holistic Guide to Approaching AI

23、Fairness Education in Organizations4Introduction1The Fourth Industrial Revolution is blurring theboundaries between the physical,digital andbiological worlds.AI is driving this revolution.A Holistic Guide to Approaching AI Fairness Education in Organizations5Artificial intelligence(AI)has been in ex

24、istence sincethe mid-1950s,when John McCarthy,computerand cognitive scientist,coined the term.AI refersto technologies that employ data and algorithmsto perform complex tasks that would otherwise torequire human decision-making.AI is consideredcolloquially to be the simulation of humanintelligence i

25、n machines.However,AI systems todaymay perform with high accuracy on a given taskor dataset,they do not have“general intelligence”,or the ability to autonomously comprehend andrespond to any decision,particularly in socialcontexts,in the same way that humans do as theyconstantly make decisions throu

26、ghout their day.1.1 A brief introduction to AI fairnessAs a technology that can be used for a number ofends,including increasing efficiency and automatingrote tasks,AI is being deployed to some degreein most large-scale organizations.The extent ofAI implementation has become a differentiator formany

27、 businesses.According to KPMG,84%offinancial services companies state that AI adoptionorientation or disability and do not discriminateamong various people or groups of people.For example,an AI-based hiring system mayrecommend candidates who are more outgoing orextroverted because many extroverted c

28、andidateswere hired in the past.However,this decisiondoes not take into account whether introvertedmannerisms could be a result of cultural differences.This could be an unfair outcome of a technicallyaccurate AI system.accelerated during the COVID-19 pandemic.2However,as AI picks up momentum in busi

29、nessapplications across the globe,the question of AIfairness looms large.Because an AI system touches many teams withinan organization before it is used by a customer orstakeholder design,data science,developers,marketing,sales etc.AI fairness cannot be madethe responsibility of one team alone.For e

30、xample,design teams,in their excitement to build a naturallanguage processing model that functions in manylinguistic contexts,may overlook issues of accessfor users with hearing impairments.Similarly,salesteams focused on AIs power to create value mightneglect the ethical implications of a sale to a

31、nauthoritarian government.AI fairness is a pillar of the larger field of AI ethics,which aims to maximize its positive impact whilemitigating the risks of AI to benefit humans andthe environment.AI ethics studies the design,development and deployment of AI systems inaccordance with agreed-upon value

32、s and principlessuch as data responsibility,privacy,inclusion,transparency,accountability,security,robustnessand fairness.While these principles and valuesmay appear in an organizations code of conduct,AI ethics and fairness should be approachedholistically,beyond compliance,as a continuousprocess t

33、o improve products and services tobetter serve both customers and broader society,ensuring AIs life cycle protects human rights andThis report aims to address one important part oforganizations approach to AI fairness:educatingdifferent teams about the role they play in advancingAI fairness.Holistic

34、 learning and educationwell-being.3regarding AI fairness in an organizations ecosystemcan drive employees to understand the importantrole they play in contributing to better,moreequitable and more ethical use of AI.Decisions made by AI systems are said to befair if they are objective with regard to

35、protectedindicators such as gender,ethnicity,sexual1.2 Why fairness?As organizations automate or augmentAI systems.This is especially important since AIsystems,largely designed by homogeneous groups only 26%of positions in data and AI are held bydecision-making with AI,there is a high risk thatdecis

36、ions either create or reinforce unfair bias.Theproblem of bias is not unique to AI.Creating fairand equitable systems is still a work in progressfor all societies.This stems from core social issuesunrelated to technology,such as structures ofpower and economy,and the lack of inclusion ofheterogeneou

37、s perspectives in decision-making.Just as we take steps to address discriminationwomen,5and around 67%of AI professors are white6 may amplify the biases of developers,therebyharming exponentially more people.The negative impact of bias and unfairnessin AI does not affect individual victims alone.Org

38、anizations that design,develop and deploy AIcan face serious repercussions including brand/reputational damage,negative sentiment amongemployees,potential lawsuits or regulatorythrough education,public policy and regulations,we need to take steps to mitigate against unintendedand inappropriate discr

39、imination embedded in4A Holistic Guide to Approaching AI Fairness Education in Organizations6It is against thisbackdrop that wedeliver concertedattention to AIfairness within thelarger realm of AIethics.penalties,and loss of trust from all stakeholdersincluding customers and the general public.work

40、as well for certain groups.A typical example of“quality-of-service harms”11 is the varying accuracyin face recognition for different ethnic groups,which has a wide-ranging impact.A biased AIsystem could also reinforce discrimination againstcertain groups by perpetuating stereotypes.12A noteworthy ex

41、ample of this is the use of theCOMPAS algorithm,which disproportionately ratedblack defendants at higher risk of recidivism whencompared with their white counterparts.13Just a few years ago,discussions about AIfairness were still mostly conducted in academicand research circles.In recent years,as gr

42、eaterattention has been paid to real-life scenariosand use cases in which individuals have beenharmed by algorithmic decision-making,manyorganizations are working to apply AI fairnessresearch to improve their workforce developmentand business management processes.Thismay include using ethical and in

43、clusive designpractices in the initial stages of product design,orimplementing gatekeeping processes so that AIproducts are not deployed before they have beenchecked for fairness.Indeed,it is notable that the extent of the impact ofunethical behaviour by human beings tends to belimited by an individ

44、uals relationships or position ina community or organization.The effect of biasedor unfair AI systems,however,can be far greater,as software created in one corner of the world caneasily be shared,sold and scaled around the world.According to a recent PWC market study,47%of organizations test for bia

45、s in data,models andAI fairness could go undetected unlessattention is paid to it.As humans are inherentlybiased in many ways,based on factors such asupbringing,education,environment and more,they may unconsciously bake biases into the AIsystems they build.Without awareness by theteams building AI m

46、odels of the importance of AIfairness,if the AI system is deployed,the harmscaused may go undetected.Take,for example,the use of AI in resource allocation unless twoindividuals from different groups compare whatthey received,the discriminated individual willnot realize that they received less than t

47、he otherindividual.Quality-of-service harms could similarlygo undetected for a long time,especially if adiscriminated individual subconsciously acceptsa lower standard of service because it alignswith other similarly discriminatory experiences.Harms from AI systems that perpetuate commonstereotypes

48、are even more insidious as they maygo unrecognized in non-diverse companies andcommunities.Fundamentally,all types of harmcaused by AI bias require a deliberate choice byan organization or stakeholder to investigate andmitigate them.human use of algorithms.But,according to a 20217BCG study,executive

49、s are broadly overestimatinghow responsible theyre actually being and arentappropriately measuring their use of AI againstpractical guidance frameworks.According to8BCGs assessment of organizations responsibleAI(RAI),of the companies it surveyed,14%werelagging,34%were developing,31%were advancedand

50、21%were leading.BCG also found that organizations are seriouslyoverestimating their RAI progress.When BCGasked executives how they would define theirorganizations progress on its RAI journey,resultsindicated:no progress(2%of respondents),had defined RAI principles(11%),had partiallyimplemented RAI(5