《联合国开发计划署：2025年人类发展报告：选择至关重要：AI时代的人与可能性（英文版）（328页）.pdf》由会员分享，可在线阅读，更多相关《联合国开发计划署：2025年人类发展报告：选择至关重要：AI时代的人与可能性（英文版）（328页）.pdf（328页珍藏版）》请在三个皮匠报告上搜索。

1、REPORT 2025A matter of choice:People and possibilities in the age of AICopyright 2025 By the United Nations Development Programme 1 UN Plaza,New York,NY 10017 USAAll rights reserved.No part of this publication may be reproduced,stored in a retrieval system or transmitted,in any form or by means,elec

2、tronic,mechanical,photocopying,recording or otherwise,without prior permission.Sales no.:E.25.III.B.2 Print ISBN:9789211576092 PDF ISBN:9789211542639 Print ISSN:0969-4501 Online ISSN:2412-3129A catalogue record for this book is available from the British Library and Library of CongressGeneral discla

3、imers.The designations employed and the presentation of the material in this publication do not imply the expression of any opinion whatsoever on the part of the Human Development Report Office(HDRO)of the United Nations Development Programme(UNDP)concerning the legal status of any country,territory

4、,city or area or of its authorities,or concerning the delimitation of its frontiers or boundaries.Dotted and dashed lines on maps represent approximate border lines for which there may not yet be full agreement.The findings,analysis,and recommendations of this Report,as with previous Reports,do not

5、represent the official position of the UNDP or of any of the UN Member States that are part of its Executive Board.They are also not necessarily endorsed by those mentioned in the acknowledgments or cited.The mention of specific companies does not imply that they are endorsed or recommended by UNDP

6、in preference to others of a similar nature that are not mentioned.Some of the figures included in the analytical part of the report where indicated have been estimated by the HDRO or other contributors to the Report and are not necessarily the official statistics of the concerned country,area or te

7、rritory,which may use alternative methods.All the figures included in the Statistical Annex are from official sources.All reasonable precautions have been taken by the HDRO to verify the information contained in this publication.However,the published material is being distributed without warranty of

8、 any kind,either expressed or implied.The responsibility for the interpretation and use of the material lies with the reader.In no event shall the HDRO and UNDP be liable for damages arising from its use.The signed contributions in boxes and spotlights represent the opinions of the authors and are t

9、he product of independent research of their responsibility.They do not represent necessarily the position or opinions of the Human Development Report Office or UNDP.Any errors or omissions are the authors responsibility.They are presented in the report to stimulate debate and to encourage further di

10、alogue between researchers and decisionmakers.Printed in the USA,by AGS,an RR Donnelley Company,on Forest Stewardship Council certified and elemental chlorine-free papers.Printed using vegetable-based ink.The 2025 Human Development ReportThe cover and chapter images in the report feature portraits i

11、n the artistic styles of various historical periods and cultures,with subtle allusions to peoples use of technology.For example,the cover presents a modern woman with headphones,against a background with hints of technology in the style of prehistoric cave paintingsan echo of humanitys earliest atte

12、mpts to understand and shape the world.Combining history with symbols of modern technology,the images place humans at the centre and aim to bridge the past and futurepositioning todays breakthroughs in artificial intelligence(AI),and the media through which we interact with them,as part of humanitys

13、 unfolding and open-ended journey towards advancing human development.Working with AI,a graphic designer created the images by guiding the system with ideas and creative direction,prompting the AI to produce a range of visual outputs that the graphic designer then edited,developed and finalized.The

14、artworks themselves reflect how AI could reshape how we do things,unleashing new creative possibilities and augmenting what people can do.The cover and other images invite you to pause and reflectas we navigate the uncertainties and possibilities of a world with AI.REPORT 2025A matter of choice:Peop

15、le and possibilities in the age of AIA matter of choicePeople and possibilities in the age of AIHUMAN DEVELOPMENT REPORT 2025 iiiHUMAN DEVELOPMENT REPORT 2025TeamDirector and lead authorPedro ConceioResearch and statisticsJoseph Bak-Coleman,Nabamallika Dehingia,Nicholas Depsky,Pratibha Gautam,Moumit

16、a Ghorai,Divya Goyal,Yu-Chieh Hsu,Christina Lengfelder,Brian Lutz,Tasneem Mirza,Prachi Paliwal,Josefin Pasanen,Antonio Reyes Gonzlez,Som Kumar Shrestha,Ajita Singh,Heriberto Tapia,Yanchun Zhang and Zakaria ZoundiDigital,data and knowledge management,communications,operations,National Human Developme

17、nt ReportsNasantuya Chuluun,Seockhwan Bryce Hwang,Nicole Igloi,Admir Jahic,Fe Juarez Shanahan,Minji Kwag,Ana Porras,Qiamuddin Sabawoon,Stanislav Saling,Marium Soomro and Sajia WaisHUMAN DEVELOPMENT REPORT 2025The 2025 Human Development Report Advisory BoardCo-chairsLaura ChinchillaFormer President o

18、f Costa RicaA.Michael SpencePhilip H.Knight Professor Emeritus of Management,Graduate School of Business,Stanford UniversityMembersMasood AhmedPresident Emeritus,Center for Global DevelopmentDeemah AlYahyaSecretary-General,Digital Cooperation OrganizationKaushik Basu Professor of Economics and the C

19、arl Marks Professor of International Studies,Cornell UniversityHaroon BhoratProfessor of Economics and Director of the Development Policy Research Unit,University of Cape TownDiane Coyle Bennett Professor of Public Policy,University of Cambridge;Co-Director,Bennett Institute for Public Policy,Univer

20、sity of CambridgeGretchen C.DailyDirector,Natural Capital Project and Bing Professor of Environmental Science,Stanford UniversityMarc FleurbaeyResearch Director,CNRS;Professor,Paris School of Economics;Associate Professor,Ecole normale suprieure,ParisPaula IngabireMinister,ICT and Innovation,Republi

21、c of RwandaSheila JasanoffPforzheimer Professor of Science and Technology Studies,Harvard Kennedy SchoolRavi KanburT.H.Lee Professor of World Affairs,International Professor of Applied Economics and Management and Professor of Economics,Cornell UniversityLuis Felipe Lpez-CalvaGlobal Director,Poverty

22、 and Equity Global Practice,World Bank GroupJ.Nathan MatiasAssistant Professor,Department of Communication,Cornell UniversityArvind NarayananProfessor of Computer Science,Princeton University;Director,Center for Information Technology PolicyRapelang RabanaCo-CEO,Imagine WorldwideFrancesca RossiIBM F

23、ellow and the IBM AI Ethics Global Leader,TJ Watson Research CenterEmma Ruttkamp-BloemHead,Department of Philosophy and AI Ethics Lead,Center for AI Research,University of PretoriaZeynep TufekciHenry G.Bryant Professor of Sociology and Public Affairs,Princeton UniversityKrushil WatenePeter Kraus Ass

24、ociate Professor in Philosophy,University of Auckland Waipapa Taumata RauLinghan ZhangProfessor,Institute of Data Law,China University of Political Science and LawFOREWORDv ForewordArtificial intelligence(AI)is racing ahead at lightning speed.Yet as AI surges forward,human development stalls.Decades

25、 of progress,reflected in the Human De-velopment Index,have flatlined,with no clear recovery from the blows dealt by the Covid-19 pandemic and subsequent crises.We are at a crossroads:while AI promises to redefine our future,it also risks deepening the divides of a world already off balance.Are we o

26、n the verge of an AI-powered renaissanceor sleepwalking into a future ruled by inequality and eroded freedoms?Too often,headlines,policies and public debates fixate on what AI might achieve in some distant futureutopian or dystopian.These deterministic views are not only disempowering;they are profo

27、undly misleading.They obscure the fact that the future is being shaped now,by the choices we make today.The 2025 Human Develop-ment Report,A Matter of Choice:People and Possibilities in the Age of AI,reminds us that it is peoplenot ma-chineswho determine which technologies thrive,how they are used a

28、nd whom they serve.AIs impact will be defined not by what it can do but by the decisions we make in its design,development and deployment.Central to these decisions is how we view the role of people in an AI-driven world.Assuming that AI will inevitably sideline humanity overlooks the very force dri

29、ving its progress:us.AIs capacity to automate nonroutine tasks has stoked fears of human replace-mentbut this is only when we reduce people to mere task-performers.This Report challenges that view.It argues that humans,“the true wealth of nations,”are far more than the sum of the tasks we perform.Ra

30、ther than measuring AI by how closely it mimics us,the Report emphasizes how the differences between humans and machines can create powerful complementarities that expand human potential.This people-centred perspective becomes even more critical in a moment of overlapping global crises.It is temptin

31、g to believe that AI alone can solve our de-velopment challenges.But that belief invites compla-cency.It asks us to surrender responsibility and ignore the political,social and systemic barriers that have long impeded progress.The 2023/2024 Human Develop-ment Report,Breaking the Gridlock,made it cle

32、ar:our limitations are not technological but sociological.Many of the crises and inequalities we face persist not because solutions are lacking but because we have failed to act.With AI we must choose differentlyand we must choose now.We might resist the temptation to anthropomorphize AI,yet in many

33、 ways it acts like a mirrorreflecting and amplifying the values,structures and inequalities of the societies that shape it.AI does not act independently of us;it evolves through our decisions and our priori-ties.If we fail to address the injustices and divides that persist today,AI will only entrenc

34、h them further.But if we invest in human capabilities and commit to greater equity,AI can magnify the best of what humanity can achieve.Ultimately,the 2025 Human Development Re-port on AI is not about technologyit is about people,and our ability to reinvent ourselves in the face of pro-found change.

35、Achim Steiner Administrator United Nations Development ProgrammeviHUMAN DEVELOPMENT REPORT 2025 AcknowledgementsEvery Human Development Report is a voyage of discovery,exploring how the human development approach helps navi-gate pressing challenges and emerging opportunities.That navigation proved p

36、ar-ticularly challenging for this Report,given the rapidly changing context of artificial intelligence(AI).AI continues to astonish every day.It engenders a mix of hype and hope,along with fear and trepidation.It is attracting financial investment and human talent towards its continuing evolution,bu

37、t it is also becoming a source of geopoliti-cal tensions.There was really no roadmap helping us navigate what seemed like a new and constantly moving AI frontier.A technology that is in many ways just one more like many others that preceded it also felt at times different,in its ability to simulate

38、and replicate features that are so distinctively human.Therefore,this is a Report that captures the spirit of a particu-lar moment in time,with much uncertainty about what might follow in terms of both AI as a technology and its ultimate impact on peoples lives.Joining in this journey of exploration

39、 are the many individuals and organizations recognized here that con-tributed their expertise,wisdom and ex-pectations,as well as doubts,about what AI might mean for human development.The Advisory Board,always a crucial source of advice and guidance,was partic-ularly relevant this year and is recogn

40、ized next to the Report team not to implicate them in the findings but to show apprecia-tion for their fundamental contribution to the Reports framing and analysis.Complementing the advice from the Advisory Board,the Statistical Advi-sory Panel provided guidance on several methodological and data as

41、pects of the Reportparticularly those related to cal-culating its human development metrics.We are grateful to all the panel members:Ola Awad,Oliver Chinganya,Koen Decan-cq,Shatakshee Dhongde,Patrick Gerland,Aishath Hassan,Ivo Havinga,Richard Heys,Solomon Hsiang,Doho Latif Kane,Steven Kapsos,Milorad

42、 Kovacevic,Jaya Krishnakumar,Christoph Lakner,Steve Macfeely,Silvia Montoya,Anu Peltola,Iaki Permanyer,Andrew Rzepa,Michaela Saisana,Claudia Sanmartin,Hany Torky and Andrew Zolli.We are also thankful to colleagues who provided data assistance to the Statistical Annex,specifically,Jenny Cresswell,Ad-

43、olfo Gustavo Imhof,Vladimra Kantorov,Olivier Lab,Jong-Wha Lee,Stephan Lutter,Alasdair McWilliam,Eric Roland Me-treau,Oscar Milafu Onam,Damien Sass,Leo Tornarolii and Yanhong Zhang.Appreciation is also extended for all the data,written inputs and peer reviews of the Reports draft chapters,including t

44、hose by PB Anand,Paul Anand,Joel An-derson,Uur Ayta,Klaus Bruhn Jensen,Yi Bu,Leonardo Bursztyn,Miriam Car-rera Manzano,Maria-Louise Clausen,Nick Couldry,Andrew Crabtree,Fabien Curto Millet,Christiaan De Neubourg,Virginia Doellgast,Kevin Donovan,Pablo Egaa del Sol,Frank Esser,Adam Fejerskov,Rana Gaut

45、am,Anne Marie Goetz,David Hammond,Benajmin Handel,Tomasz Hollanek,Jeroen Hopster,Johannes Jae-ger,Rafael Jimenez Duran,Julia Karpati,Marie Kolling,Anton Korinek,Seth Lazar,Margauz Luflade,Michael Muthukrishna,Rose Mutiso,Kruakae Pothong,Stiene Praet,Carina Prunkl,Mitsy Barriga Ramos,Christoph Roth,A

46、nna Salomons,Stefka Schmid,Tobia Spampatti,Tara Thiagara-jan,Luis Hernn Vargas,Manuela Veloso,Juri Viehoff,Zi Wang,saWikforss,Kuan-song Victor Zhuang and David Zuluaga Martnez.We are especially thankful to our close collaborations with our partners:Mario Biggeri,Enrica Chiappero-Martinetti,Fla-vio C

47、omim,Carlos Alberto Garzon,Ann Mitchell and Kathy Rosenblum at the Human Development&Capability Orga-nization;Stefano Calcina,Valentina Caliri,Giuseppe Diglio,Gerardo Filippo,Marina Kodric,Fabio Marchetti,Bianca Mihalcea,Marco Presenti and Andrea Sironi at Gen-erali;Jon Clifton,Kiki Papachristoforou

48、 and Andrew Rzepa at Gallup;Suela Aksoy,Nancy Hey and Ed Morrow at the Lloyd Register Foundation;Antonio Corcoles,Ismael Faro,Zaira Nazario and Kush Varshney at IBM;David G.Blanchflower at Dartmouth College and Alexander Bryson from University College London;Beata Javorcik and Zoe Russo at the Europ

49、ean Bank for Reconstruction and Develop-ment;Nino Naderashvili and Charlie Zong at South-North Scholars;Juliana Alves Soares,Paul Anthony,Kimberley Blair Bolch,Nicholas Nam and Leslie J Yun at the World Bank;Sabina Alkire at the Oxford Poverty and Human Development Initiative;Stijn Broecke at the Or

50、ganisation for Economic Co-operation and Develop-ment;Lucas Chanel at the World Inequal-ity Lab;Ketan Patel at the Force for Good;Jonathan Richard Schwarz at the UK AI Security Institute/Thomson Reuters;Phil-lip Howard and Sebastian Valenzuela at the International Panel on the Information Environmen

51、t;Jos M.Tavares at the Nova School of Business and Economics;and Hannah Hess at the Climate Impact Lab.Our thanks are also extended to Olimpia Dubini,Olivia Lempa and Richard Steinert at the Nova School of Business and Eco-nomics working on the Capstone Project.Several consultations and seminars wit

52、h thematic and regional experts and nu-merous informal consultations with many individuals without a formal advisory role were held in the process of preparing this years Report.We are grateful for input in AcknowledgementsACKNOWLEDGEMENTSviithese consultations from Siri Aas Rustad,Tayma Abdalhadi,A

53、lexandra Abello Colak,Elena Abrusci,Adedji Adeniran,Fabrizio Andreuzzi,Anatola Araba,Vesa Arponen,Victoria Austin,Gifty Ayoka,Joon Baek,Maha Bahou,Onur Bakiner,Pallavi Bansal,Roxana Barrantes,Gustavo Bliz,Eliot Bendinelli,Cynthia Bennett,Rahul Bhar-gava,Nidal Bitar,Karl Blanchet,Joshua Blumenstock,J

54、oanna Bryson,Romina Ca-chia,Hailey Campbel,Maria Paz Canales,Michele Candotti,Michela Carlana,Dante Castillo,Han Sheng Chia,Zhang Chunfei,Paul Anthony Clare,Daniella Darlington,Erika Deserranno,Arkan El Seblani,Ethar Eltinay,Alberto Fernndez Gibaja,Elenore Fournier-Tombs,Victor Galaz,Helani Galpaya,

55、Daniela Garcia Villamil,Michael Gibson,Gabriel Gomes Couto,Piers Gooding,Andrea Guariso,Anita Gurumurthy,Jinhwa Ha,Jungpil Hahn,Hamza Hameed,Corinne Heckmann,Catherine Holloway,Marie Humeau,Ghis-lain Irakoze,Natalie Jabangwe,Parminder Jeet Singh,Yu Jianjun,Priscilla Ege Johnson,Seong Hwan Ju,Ma Jun,

56、Zubair Junjunia,Frederike Kaltheuner,Ozge Karadag,Mary Kawar,Harttgen Kenneth,Jungwook Kim,Niki Kim,Taeho Kim,Yoon Ko,Sengmeng Koo,Adithi Kumar,Nagesh Kumar,Protiva Kundu,Cheol Lee,Dong Hoon Lee,Hyun-kynung Lee,Emmanuel Letouze,Nicola Limodio,Bjrn-Ola Linner,Sonia Livingstone,Yu Lu,Jean Luc Mas-taki

57、,Ke Luoma,Lusa Franco Machado,Anu Madgavkar,Izhar Mahjoub,Joan Manda,Jenifer Mankoff,Audrin Mathe,Francesca Mazzi,Lena Menge,Saurabh Mishra,Hlne Molinier,Nusrat Molla,Amal Mowafy,Ava Nadir,Yushi Nagano,Daniel Naoujoks,Fabio Nascimbeni,Alain Ndayishimiye,Megan ONeill,Toby Ord,Gudrun stby,Nikolas Ott,

58、Nikhil Pahwa,Yuhun Park,Balaji Parthasarathy,Pratik Patil,Laurel Patterson,Jason Pielemeier,Fillippo Pierozzi,Carina Prinkl,Raphalle Rafin,Rebeca Robboy,Yurii Romashko,Ilana Ron Levy,Asma Rouabhia,Satyaki Roy,Tiffany Saade,Dong-Pyoung Sheen,Bahja Ali Shuriye,Rita Singh,Sebastian Smart,Sang Hyo Song,

59、Tong Song,Paul Spiegel,Serge Stinckwich,Jaimee Stuart,Inkyoung Sun,Yash Tadimalla,Zhou Taid-ong,Toshie Takahashi,Ma Tianyue,Jutta Treviranus,Chi-Chi Undie,Ott Velsberg,Stefaan Verhulst,Anna Walch,Skyler Wang,Zi Wang,Achim Wennmann,Olivia White,Isaac Wiafe,Kellee Wicker,Kebene Wodajo,Wang Xiaolin,Wan

60、 Xiaoyan,Yang Xingli,Nobuo Yoshida,Zhou Yu-Ya,Muhammad Zaman,Liang Zheng,Shen Zhou and Enrique Zuleta Puceiro.Further support was also extended by others too numerous to mention here.Consulta-tions are listed at https:/hdr.undp.org/towards-hdr-2025.Contributions,support and assistance from many coll

61、eagues across the UNfam-ily are gratefully acknowledged:the International Telecommunication Union,including Jin Cui,Fredrik Ericsson,Thierry Geiger,Youlia Lozanova,Jose Luis,Rosie McDonald,Martin Shaaper and Caroline Troein;the International Labour Orga-nization,including Janine Berg,David Bescond,E

62、kkehard Ernst,Andrea Mari-nucci,Uma Rani,Olga Streitska-Ilina and Dagmar Walter;the Office of the High Commissioner of Human Rights,including Scott Campell,Isabel Ebert,Peggy Hicks and Nathalie Stadelmann;the United Na-tions Office for South-South Cooperation,including Zanofer Ismalebbe and Nav-eeda

63、 Nazir;the United Nations Entity for Gender Equality and the Empowerment of Women,including Hlne Molinier and Raphalle Rafin;the United Nations Educational,Scientific and Cultural Orga-nization,including Priyadarshani Joshi,Iaroslava Kharkova,Irakli Khodeli,Karalyn Monteil,Claudia Roda and Prateek S

64、ibal;the United Nations University,including El-enore Fournier-Tombs,Tshilidzi Marwala,Serge Stinckwich and Shen Xiamomeng;the UN Secretary-Generals Special Envoy for Digital and Emerging Technologies Mehdi Snene;and the United Nations Industrial Development Organization Re-gional Office,including S

65、hraddha Srikant.Colleagues in at the United Nations Development Programme(UNDP)pro-vided advice and input and organized consultations.We are grateful to Tehmina Akhtar,Abdallah Al Dardari,Fabrizio An-dreuzzi,Iffat Anjum,Jacob Assa,Estefania Asturizaga,Marcos Athias Neto,Walid Badawi,Rodrigo Barraza,

66、Iram Batool,Fiona Bayat-Renoux,Yakup Beris,Robert Bernado,Benjamin Bertelsen,Jeremy Boy,Susan Brown,Camilla Bruckner,Mi-chele Candotti,Yu Ping Chan,Gary Chew,Hojin Chung,Enrique Crespo,Pauline Deneufbourg,Roqaya Dhaif,Violante di Canossa,Mirko Ebelshaeuser,Ahunna Eziakonwa,Almudena Fernandez,Kumiko

67、Fukagawa,Arvinn Gadgil,Victor Garrido,Herte Gebretsadik,Raymond Gilpin,Kiri Ginnerup,Carolina Given Sjlander,Carla Gomez,Janil Greenaway,George Gray Molina,El Hadji Fall,Joe Hooper,Caro-line Hopper-Box,Alexander Hradecky,Vito Intini,Ghida Ismail,Giulia Jacovella,Zulkarin Jahangir,Anne Juepner,Hurshi

68、d Kalandarov,Tomohiro Kawase,Antonin Kenens,Sujin Kim,Sharon Kinsley,Yuna Koh,Adithya Kumar,Alexis Laffittan,Julie Lee,Regina Lio,Jennifer Louie,Linda Maguire,Joan Manda,Michelle Muschett,Debashis Nag,Steliana Nedera,Liwen Ng,Keyzom Ngodup,Shoko Noda,Camila Olate,Robert Opp,Anna Ortubia,Hye-Jin Park

69、,Gayan Peiris,Isabella Rosso,Jelena Ruzicic,Pratyasha Saha,Sebnem Sahin,Turhan Saleh,Philip Schellekens,Anca Stoica,Helin Su Aslan,Hyunjee Sung,Ludmila Tiganu,Riccardo Trobbiani,Ramiz Uddin,Georges Van Montfort,Agi Veres,Kanni Wignaraja,Lesley Wright,Qu Xinyi,Haoliang Xu,Shinobu Yamaguchi,Weijing Ye

70、,Vitali Zakhozhyi and Ivana Zivkovic.We were fortunate to have the support of talented interns and fact checkers:Id-ris-Alaba Aderinto,Natalia Aguilar,Komla Amega,Raiyan Arshad,James Chabin,Andrea Davis,Jessica Karki,Danielle Mal-lon,Chiara Marcoccia,Nazifa Rafa,Yu-Ya Rong,Laura Sanzarello and Xiqin

71、g Zhang.The Human Development Report Office(HDRO)also extends its sincere gratitude to the governments of Japan and the Re-public of Korea for their financial contribu-tions.Their ongoing support is very much appreciated and remains essential.We are grateful for the highly pro-fessional work of our

72、editors and layout artists at Communications De-velopment Incorporatedled by Bruce Ross-Larson,with Joe Caponio,Meta de Coquereaumont,Mike Crumplar,Chris-topher Trott and Elaine Wilson.It was a communal experience learning together,especially with Bruce,about how regular viiiHUMAN DEVELOPMENT REPORT

73、 2025spoken language(natural language in the computer science jargon)is becom-ing a new interface to communicate with computational machines,as well as how artificial intelligence can support the preparation of these reports.That experi-ence extended to the collaboration with Therese Severinsen Marq

74、ues and the team at Studio Mnemonic in preparing the cover and images in the Report.Therese was given a difficult challengeto come up with options that centred artificial intel-ligence on the human and avoid clichs of robots or digital circuitsand she suc-ceeded in creating beautiful images with the

75、 help of artificial intelligence that met this brief.Over several years now the Human Development Report owes a deep debt of gratitude to UNDP Administrator Achim Steiner.This gratitude has ac-cumulated over the years because he has not only scrupulously preserved and protected HDROs editorial indep

76、endence but has always been generous with his time and wisdom.He has provided us with guidance and,on more than one occa-sion,challenged us to be more and more ambitious,so that we could make a differ-ence in advancing human development.We only hope to have been worthy of the trust and confidence th

77、at he has depos-ited in our team.Pedro ConceioDirectorHuman Development Report Office ContentsCONTENTS Foreward vAcknowledgements viOverview 2Terms and concepts 12CHAPTER 1Empowering people to make artificial intelligence work for human development 16Examining the demand side of AI 17Looking back a

78、digital transformation going from creator to destroyer?24Attention is all you need for tasks that AI may do well in the future 26Envisioning the human development opportunity of AI 34CHAPTER 2From tools to agents:Rewiring artificial intelligence to promote human development 46From doing what we do t

79、o choosing what we choose 47Entering a brave new(digital)world 49Embedding AI into our social fabric 54AI-infused social networks:What happens when AI makes choices for,between and among us?59Preserving and expanding human agency across scales 62CHAPTER 3Artificial intelligence across life stages:In

80、sights from a people-centred perspective 66Early childhood too little,too much,too risky 69School age access,regulation and ownership 72Adolescence smartphones,AI-powered apps and mental wellbeing,much ado about nothing?75Semi-autonomous adulthood with overlapping identities 79Older age trained,empo

81、wered and healthier?83Multistakeholder action for people-centred AI 86CHAPTER 4Framing narratives to reimagine artificial intelligence to advance human development 102Beyond techno-determinism:Technological change shapes and is shaped by society 103AIs potential for people with disabilities:Framing

82、a more nuanced narrative to expand human development 105Narratives about care technologies overlook the profoundly human and relational nature of care 109Narratives about gender digital divides paint an incomplete picture 113Technical solutions are not enough:Biases in AI are deeply intertwined with

83、 social norms and societal inequalities 119Framing a narrative on AI to advance human development 121CHAPTER 5Power,influence and choice in the Algorithmic Age 136Algorithms shape social choices and power 139Who has the power?Divides and dependencies are evolving amid furious AI races 146CHAPTER 6Re

84、imagining choices:Towards artificial intelligenceaugmented human development 162Building a complementarity economy to expand development frontiers 164Driving innovation with intent:Aligning socially and privately valuable AI research 172Investing in capabilities that count:Can AI enhance education a

85、nd health outcomes?179The road ahead:AIs promise to advance human development 185Notes 201References 221BOXES1.1 The many ways generative artificial intelligence differs from classical programming 271.2 The perils and affordances of artificial intelligence 30S1.2.1 Human intelligence is not defined

86、by that of a single human but of many:Could artificial intelligence get there?422.1 Artificial intelligence revolutionizing biomedicine 513.1 Artificial intelligence can violate childrens rights or protect them 713.2 Levelling the playing field for disadvantaged students 733.3 Artificial intelligenc

87、e on social media undermines agency and drives emotions but only for some young people so far 773.4 Harmful friends without benefits 81S3.1.1 Connected or disconnected?Exploring possible mechanisms between smartphones and mental wellbeing 904.1 Going beyond access:Womens disproportionate care respon

88、sibilities drive their lower digital skills 1164.2 As technologies advance,so do new ways of perpetrating violence against women 1185.1 Recommendations in digital platforms and human development:Artificial intelligence as part of the problem,part of the solution?1425.2 The UN Global Digital Compact

89、for addressing power imbalances and fostering inclusive artificial intelligence 1535.3 More subtle manifestations of power emerge in artificial intelligence models behaviour 1545.4 The potential for artificial intelligence audit protocols 1556.1 Assessing artificial intelligences productivity effect

90、s 1686.2 Smart systems,shared goals:The complementarity of artificial intelligence and digital public infrastructure 1706.3 Whos the boss?The rise of algorithmic management in the automobile manufacturing sector 1736.4 Bridging bytes and governments:Artificial intelligence ecosystems through partner

91、ships 178FIGURESO.1 About two-thirds of survey respondents in low,medium and high Human Development Index(HDI)countries expect to use artificial intelligence in education,health and work within one year 3O.2 Global progress in human development is losing steam,with the weakest and most vulnerable be

92、ing left farther behind 4O.3 The post-2020 slowdown in human development progress affects every region of the world 5O.4 People at each life stage use artificial intelligence(AI)for different purposes 7O.5 Young internet users are struggling everywhere 8O.6 Younger people expect to lose control over

93、 their lives due to artificial intelligence(AI)less than older people do 9O.7 Across occupations and Human Development Index levels,respondents expect that artificial intelligence will both automate and augment their work with higher expectations of augmentation 10O.8 ChatGPT answers are culturally

94、closer to those of humans in very high Human Development Index(HDI)countries 101.1 About two-thirds of survey respondents in low,medium and high Human Development Index countries expect to use artificial intelligence(AI)in education,health and work within one year 181.2 Low Human Development Index c

95、ountries are being left further behind 191.3 Most survey respondents are confident that artificial intelligence(AI)will make them more productive at work,and the more AI is used,the higher the share of respondents reporting feeling confident 201.4 The majority of monthly ChatGPT web traffic came fro

96、m middle-income countries by mid-2023 241.5 With classical programming,machines can execute routine tasks 251.6 The cost of computing declined by 12 orders of magnitude in the classical programming age 261.7 Beyond the routinenonroutine tasks dichotomy:What artificial intelligence(AI)can automate de

97、pends on the stakes and on the range of potential implications 291.8 The lower the level of skill and experience,the more workers benefit from artificial intelligence(AI)32S1.2.1 A human development interpretation of the evolution of computational machinesmore tasks helpful to humans with less effor

98、t 412.1 Sense of agency now and in an artificial intelligence(AI)defined future 472.2 Simpler forms of artificial intelligence(AI)may more easily promote human agency,whereas AI with high agenticity can have a broader range of more dramatic impacts 482.3 Interactions between and among humans and art

99、ificial intelligence 552.4 Cultural differences from the United States explain the use of ChatGPT 623.1 People at each life stage use artificial intelligence(AI)with varying frequency and for different purposes 683.2 Invest,inform and include for people-centred artificial intelligence(AI)693.3 Exces

100、sive screen time in early childhood is related to changes in the brain structure and to reduced language capacity and understanding 703.4 Mathematics achievement in the United States did not decline after calculators became available in the classroom 743.5 Pandemic-related stress is a complementary

101、explanation for adolescents mental illbeing 763.6 Multidimensionally poor people with little education lack access to the internet 793.7 Disentangling autonomy,authenticity and agency in the digital space 803.8 Automated systems may cut costs but distress customers 823.9 Very little internet use amo

102、ng older people 843.10 Stark variance in internet use among older people across countries with different Human Development Index levels 853.11 Across world regions older people who use the internet are less distressed than younger ones 863.12 Social,algorithmic and data-driven biases in older people

103、s healthcare 873.13 Harnessing artificial intelligence(AI)for human development invest,inform,include 87S3.1.1 Declining wellbeing,rising despair among young people in the United States 89S3.1.2 Increase in despair in the United States since 2010,especially among women 90S3.1.3 Young internet users

104、are struggling everywhere 91S3.1.4 The age at first smartphone ownership appears to matter for mental wellbeing 92S3.2.1 Respondents who prefer to live in a world without the platform 94S3.2.2 Consumer surplus across welfare measures 954.1 People with disabilities also face inequalities in internet

105、use 1064.2 Most patents for conventional assistive technology are filed in just a handful of countries 1074.3 as are most patents for emerging assistive technology 1084.4 Older people expect to have less choice and control over their lives as artificial intelligence technologies become more integrat

106、ed into daily life 1114.5 On average,only 35percent of graduates in science,technology,engineering and mathematics are women 1144.6 The share of graduates in science,technology,engineering and mathematics who are women has changed little since 20102011 1155.1 The market structure of the artificial i

107、ntelligence(AI)supply chain is concentrated 1385.2 Artificial intelligence transforming the way people retrieve information 1405.3 Recommender algorithms show how artificial intelligence is shaping social,economic and political processes 1415.4 Artificial intelligence(AI)outperforms human mediators

108、in finding common ground 1455.5 The majority of todays large-scale artificial intelligence models are developed by organizations based in the United States,followed by China and the United Kingdom 1505.6 Most global investment in artificial intelligence(AI)flowed to the United States in 2024 150CONT

109、ENTSxi5.7 Artificial intelligence(AI)talent has been flowing towards high-income countries 1515.8 India has the highest self-reported artificial intelligence(AI)skills penetration 1515.9 The artificial intelligence(AI)race today can be conceptualized as unfolding along a spectrum spanning innovation

110、 to arms 1526.1 Across Human Development Index(HDI)groups the largest share of jobs exposed to artificial intelligence(AI)falls into“a big unknown”1656.2 Men and people with greater levels of education report higher use of artificial intelligence(AI)for workacross all Human Development Index groups

111、1666.3 More respondents in low and medium Human Development Index(HDI)countries expect labour market changesthrough augmentation,automation and productivity boostswith artificial intelligence 1666.4 Across occupations and Human Development Index levels,respondents expect that artificial intelligence

112、 will both automate and augment their work with higher expectations of augmentation 1676.5 Across occupations respondents expect transformational change to their work 1686.6 Disruptive science and technological innovation was on a steady decline through 2010 1756.7 Artificial intelligence can inspir

113、e humans to reach new heights in creativity 1756.8 Educationconvergence in basic capabilities,divergence in enhanced capabilities 1796.9 Critical thinking mitigates students propensity towards extreme trust or distrust of online content 1806.10 The benefits of digital resources for learning critical

114、 thinking diminish with excessive use 1816.11 Mind the contextinitial conditions can compound development challenges 185S6.1.1 Computing performance has evolved at roughly the same pace as warming temperatures in recent decades 189SPOTLIGHTS1.1 Humans have agency,algorithms do not 361.2 A human deve

115、lopment perspective on the pursuit of artificial general intelligence 393.1 The decline in young peoples mental wellbeing in some parts of the world 883.2 The social media trap 943.3 Worker agency in the digital age 974.1 Narratives in economic decisionmaking 1254.2 Caring through digital platforms

116、1275.1 Threats to democratic reason in a high-choice information environment 1576.1 The promise and peril of leveraging artificialintelligence to address dangerous planetary change 1886.2 Universal and meaningful connectivity and artificial intelligence 1936.3 Global case studies of social dialogue

117、on artificialintelligence and algorithmic management 195TABLES1.1 Machine learning has extended the use of machines to many tasks that classical programming struggled with 232.1 Comparing characteristics of digital tools and artificial intelligence(AI)agents 50S3.2.1 The CARE framework 995.1 When do

118、 we confront high stakes?When“power over”is concentrated and impacts deeply or across many dimensions of peoples lives 1395.2 Gaps across country income groups based on popular artificial intelligence(AI)metrics 1495.3 Where there is a stronger case for international policy coordination on artificia

119、l intelligence 155STATISTICAL ANNEXReaders guide 273HUMAN DEVELOPMENT COMPOSITE INDICES 1 Human Development Index and its components 2782 Human Development Index trends,19902023 2833 Inequality-adjusted Human Development Index 2874 Gender Development Index 2925 Gender Inequality Index 2976 Multidime

120、nsional Poverty Index:developing countries 3027 Planetary pressures-adjusted Human Development Index 305Developing regions 310Statistical references 311A matter of choicePeople and possibilities in the age of AIOVERVIEW12HUMAN DEVELOPMENT REPORT 2025OVERVIEWA matter of choice:People and possibilitie

121、s in the age of AIOVERVIEW A MATTER OF CHOICE:PEOPLE AND POSSIBILITIES IN THE AGE OF AI3Artificial intelligence(AI)has broken into a dizzy-ing gallop.Each day seems to herald some new AI-powered algorithmic wonder.As a general-purpose technology,AI has been dubbed“the new electric-ity.”Regardless of

122、 whether the utopian,techno-solutionist1 visions of AIs most ardent advocates come to fruition or fizzle as snake oil(or worse),the world is pulsing with a powerful new technology,a new kind of dynamism or vitality,that differs from technologies of the past.Yet,the AI zeitgeist is awfully blinkered.

123、Headlines fixate on arms races,policymaking on risks.These are real.But they are not and should not be the whole story.We need to go beyond races and risks to possibili-ties for people,possibilities shaped by peoples choices.The choices that people have and can realize,with-in ever expanding freedom

124、s,are essential to human development,whose goal is for people to live lives they value and have reason to value.A world with AI is flush with choices the exercise of which is both a matter of human development and a means to ad-vance it.The future is always up for grabs,even more so now.Trying to pr

125、edict what will happen is self-defeating,privileging technology in a make-believe vacuum over the frictional realities and messier promises of peoples agency and their choices.From a human development perspective the relevant ques-tion instead is what choices can be made so AI works for people.This

126、years Human Development Report examines what distinguishes this new era of AI from previous digital transformations and what those differences could mean for human development(chapter 1),in-cluding how AI can enhance or subvert human agen-cy(chapter 2).2 People are already interacting with AI in dif

127、ferent ways at different stages of life,in ef-fect scoping out possibilities good and bad and un-derscoring how context and choices can make all the difference(chapter 3).Human agency is the price when people buy into AI hype,which can exacerbate Figure O.1 About two-thirds of survey respondents in

128、low,medium and high Human Development Index(HDI)countries expect to use artificial intelligence in education,health and work within one year14.423.619.066.168.945.9020406080Actual use of AIin the past monthExpected use of AIin one yearExpected increase in useHDI groupLow and mediumHighVery highShare

129、 of population(%)Note:Based on pooled data for 21 countries.For actual use in the past month,the following responses to the question,“In the past 30 days,have you ever interacted with artificial intelligence,such as chatbots,in any of the following ways?”were used to calculate the average use of AI

130、for education,health and work:“education”is based on the response“educational platforms of learning apps,”“health”is based on the response“health care services or applications”and“work”is based on the response“work-related tools or software.”For expected use in one year,the following responses to th

131、e question,“Over the next 12 months,how likely are you to use an artificial intelligence tool for the following?”were used to calculate the average use of AI for education,health and work:“education”is based on the response“for education and training,”“health”is based on the response“for medical adv

132、ice”and“work”is based on the response“for work tasks.”Expected increase in use is the difference between expected use in one year and actual use in the past month.Source:Human Development Report Office based on data from the United Nations Development Programme Survey on AI and Human Development.4HU

133、MAN DEVELOPMENT REPORT 2025exclusion(chapter 4)and harm sustainability.3 And,of course,who produces AI and for what matter a lot for everyone(chapter 5).Letting people take the reins makes good sense,because they expect AI to be a growing part of their lives.A global survey4 for this Report found th

134、at,at all levels of the Human Development Index(HDI),AI use is already substantial(for about 20percent of respondents)and is expected to shoot up fast.About two-thirds of respondents in low,medium and high HDI countries expect to use AI in education,health and work the three HDI dimensions within on

135、e year(figure O.1).Human development gaps are widening,and global progress may be losing steamFocusing on people can help many countries feel-ing caught in a human development pinch between Figure O.2 Global progress in human development is losing steam,with the weakest and most vulnerable being lef

136、t farther behind 0.805 0.785eu l av )IDH(xedn I tnempo l eveD namuH l abo lG 0.8252015201020052000.0 765 0.745 0.725 0.705 0.685 .0 665 0.64520232024Pre-2020 trend20212024 trendtrend(projected)(extrapolated trend)0202 4202 2030Threshold for very high HDI group:0.800 0.0000.0010.0020.0030.004Change i

137、n HDI value,19902024(excluding 20202022)0.0080.0070.0060.0052020201520102005200019951990Mean change(excluding 20202022)4.5 lowerthan the 19902024 mean change(excluding 20202022)2024(projected)Diference in HDI value betweenvery high and low HDI countries0.3800.4000.4200.4400.4602024202020152010200520

138、0019951992201620202024(projected)0.4100.4000.390Source:Human Development Report Office calculations based on data from Barro and Lee(2018),IMF(2024),UNDESA(2024),UNESCO Institute for Statistics(2024),United Nations Statistics Division(2025)and World Bank(2024).OVERVIEW A MATTER OF CHOICE:PEOPLE AND

139、POSSIBILITIES IN THE AGE OF AI5sky-high expectations for AI and sobering develop-ment realities,including ongoing violent conflicts and stresses on human security.Wounds from the 20202021 declines in global HDI value have not healed,and the rebound since may be losing steam.Just a few years ago we w

140、ere on course to live in a very high HDI world by 2030.5 That world was delayed by a few years based on the 20212024 trend.Now it is projected to be delayed by decades(top left panel of figure O.2).6While the global HDI value is projected to reach a record high in 2024,the increase would be the low-

141、est since records began 35 years ago(top right panel of figure O.2).Gaps between very high and low HDI countries,which for decades had been shrinking,have been widening over the past four years(bottom panel of figure O.2).The dramatic slowdown in HDI pro-gress cuts across all developing regions(figu

142、re O.3).Development pathways that have created jobs at scale and reduced poverty,thanks to expanded man-ufacturing and exports to international markets,are narrowing.7 A triple squeeze results from inadequate external financing,fewer opportunities in manufac-turing due in part to automation and trad

143、e tensions limiting export options.8Now enter AI,a development wildcard.9 If AI is seen simply as a supercharged extension of earlier digital technologies deployed to automate work,la-bour is condemned to cede the remaining ground to machines,further eroding development options.Is this what is in th

144、e cards?It is a matter of choices.Development depends less on what AI can do not on how human it appears and more on mobilizing peoples imaginations to re-shape economies and societies to make the most of it.Figure O.3 The post-2020 slowdown in human development progress affects every region of the

145、world0.6001999Arab States Regional Human Development Index value0.7500.7000.6502007201520230.6000.6500.7000.7500.8000.85019992007201520230.6000.6500.7000.7500.8000.8500.90019992007201520230.6900.7100.7300.7500.7700.79019992007201520230.5000.5500.6000.6500.70019992007201520230.4000.4500.5000.5500.600

146、1999200720152023East Asia and the PacificSouth AsiaSub-Saharan AfricaEurope and Central AsiaLatin America and the CaribbeanPre-2020 trendSource:Human Development Report Office calculations based on data from Barro and Lee(2018),IMF(2024),UNDESA(2024),UNESCO Institute for Statistics(2024),United Nati

147、ons Statistics Division(2025)and World Bank(2024).6HUMAN DEVELOPMENT REPORT 2025Making AI work for people is a matter of choicesAI does some things uniquely well,such as seeing pat-terns in huge datasets that are difficult or impossible for humans to discern.10 It does other things poorly,sometimes

148、making things up.11 It cannot frame prob-lems,as humans can do.Whatever new algorithmic feats are in store,there will always be spaces,howev-er in flux,where humans shine where humans do things that machines cannot do or are bad at,where societies value people rather than machines doing things and w

149、here people and machines go farther and faster together than separately.Evolving overlaps and complementarities between humans and AI-powered machines land societies at inflection points,after which trajectories will depend largely on two factors:what access societies have to AI and how they view an

150、d use it.These are choices,by the few or the many.Is the focus on overlaps,pit-ting what Daron Acemolu calls so-so AI against peo-ple,which could cut jobs without productivity gains?12 Or is it instead on complementarities and collabora-tion to envision new development pathways?13 Entire-ly new role

151、s,markets and industries could be in the offing.If anything,then,AI can be seen as adding hazy pages to the development playbook instead of stripping them away.Possible paths become wider,if less clear,given that much is yet unknown about what AI can do and how it will affect human decisions.“AI can

152、 be seen as adding hazy pages to the development playbook instead of stripping them away.Possible paths become wider,if less clear,given that much is yet unknown about what AI can do and how it will affect human decisionsPeople seem to expect as much:a cloudy glass half full.Nearly 4 in 10 responden

153、ts14 in the survey for this Report expect AI to automate and augment jobs.Overall expectations for augmentation(61percent)just edge those for automation(51percent).15 And the more that people use AI,the more confident they feel in its ability to increase productivity.Expectations in developing count

154、ries are particularly high.16 With so much promise and expectation,the bar for AI is higher than simply being useful or“doing good”;it is avoiding development disappointment.It is time to break the spell of technological inevita-bility:no path forward is about technology in isolation but rather how

155、it is deployed by whom,with whom,for whom and with what kind of accountability.Dif-ferent choices can help turn things around,and the lens of this years Human Development Report,focused on people and possibilities,identifies three areas of action for AI-augmented human development(chapter 6):1.Build

156、ing a complementarity economy,so people and AI find more opportunities to collaborate rather than compete.Rather than try to predict the future,policy makers should shape it,breaking away from trying to guess how humans will be replaced by AI,to see the poten-tial of what humans can do with AI.That

157、includes driving productivity gains through intelligence aug-mentation,leveraging the complementarities be-tween AI and people.Ensuring that AI is proworker,limiting curbs on agency and empowering workers to use AI to augment what they can do.Deploying AI in sectors where positive spillovers to othe

158、r sectors and across the economy can be leveraged,helping with economic diversification and job-creating structural transformation.Implementing fiscal measures and strengthening social dialogue that incentivize AI to safeguard decent work and supporting incumbent workers displaced by AI.2.Driving in

159、novation with intent,so opportunity for people is not an afterthought but a built-in integral part of AI design and deployment.AI should be harnessed to accelerate science through curiosity-driven basic research,as well as technological innovation not by automating creative processes but by augmenti

160、ng them.17 AI innovation can be steered through incentives that embed human agency in AI from design to deployment by aligning socially desirable and privately profitable innovation and supplementing existing AI benchmarks with new ones that capture AIs potential to advance human development.3.Inves

161、ting in capabilities that count,so people have the capabilities to make the most of AI in their lives and to thrive in a world with AI.AIs flexibility and adaptability should be lev-eraged to personalize education and healthcare OVERVIEW A MATTER OF CHOICE:PEOPLE AND POSSIBILITIES IN THE AGE OF AI7i

162、n different contexts,while attending to risks and concerns related to bias,privacy,affordability and equity.18 By tailoring learning or expanding health care,AI can also generate demand for complemen-tary human labour.19Together,the three areas invite policymakers at dif-ferent levels to shake off u

163、nhelpful narratives that swing between utopia and dystopia,to depart from disem-powering trends that sideline most people or put bull-seyes on their backs and instead to embolden people to reimagine their choices and expand their freedoms.Who,where,when and how?AIs possibilities depend on contextThe

164、 possibilities of AI depend on context:who,where,when,how?AI is more than just an opportuni-ty for peoples choices;it requires them.People of dif-ferent ages use AI for different purposes(figureO.4).AI has shown promise for helping students by provid-ing study assistance when educators or parents ha

165、ve time or resource constraints20 or by improving per-sonalized,adaptive learning.21 AI could bridge gaps in the light of constrained education resources and help level the field for disadvantaged students.22 This is in addition to not in lieu of teachers,who uniquely provide,among other things,nece

166、ssary social inter-actions critical to students overall development.Until recently,one of the most well-established em-pirical regularities across countries was that subjec-tive measures of wellbeing(such as life satisfaction)followed a U-shaped pattern with age:younger and older people reported hig

167、her wellbeing than those in middle age(late 40s to early 50s).23 About 1015 years ago that began to change in some countries.Despair among young people shot up,and life satisfaction tanked.24 Young women fare worse than young men.25What explains the dramatic declines among young people?The picture i

168、s complex and evolving.That the trend is most evident in some very high HDI countries and parallels the broader diffusion of smartphones has implicated digital technologies.In a global survey of people with access to the internet,the typical U-shape curve is completely absent.In its place is essenti

169、ally a diagonal line,with young peo-ples mental wellbeing at the bottom(figure O.5).26The opportunities for and risks to young peo-ple from digital technologies,including AI,are Figure O.4 People at each life stage use artificial intelligence(AI)for different purposesPurpose of AI use by occupation

170、groupStudentNonworkforceWorkforceRetired0102030506040Share of survey respondents(%)EducationWorkHealthEntertainmentNote:Based on pooled data for 21 countries.For purpose of AI use,the follow-ing responses to the question,“In the past 30 days,have you ever interacted with artificial intelligence,such

171、 as chatbots,in any of the following ways?”were used to calculate the average use of AI for work,education,entertainment and health:“work”is based on the response“work-related tools or software,”“ed-ucation”is based on the response“educational platforms of learning apps,”“entertainment”is based on t

172、he response“entertainment(e.g.streaming serv-ices/gaming)”and“health”is based on the response“health care services or applications.”For occupation group the following responses to the question“What best describes you?Are you?”were used:“working”includes self-identified full-and part-time employees a

173、nd self-employed respondents,and“not working”includes homemakers and unemployed respondents.Source:Human Development Report Office based on data from the United Nations Development Programme Survey on AI and Human Development.8HUMAN DEVELOPMENT REPORT 2025particularly relevant for many lower HDI cou

174、ntries,where age structures skew young and digital pene-tration has farther to go.That is itself an opportuni-ty to chart a path informed by lessons elsewhere.The age structures of many higher HDI countries lean the other way,towards the old.Although patterns dif-fer across countries,the world as a

175、whole is greying quickly,with 1.4billion people age 60 or older ex-pected by 2030.27 At the same time younger people expect to lose control over their lives due to AI less than older people do(figure O.6).AI has enabled pathbreaking innovations in as-sistive and accessible technologies that can expa

176、nd choices and opportunities for people with disabili-ties,technologies such as live captioning,image de-scriptions and translation of sign language into voice or text.28 But achieving the full reach and potential of these and other applications depends on more than technology alone.Social choices a

177、nd contexts matter,too,29 including,at the most fundamental level,whether these applications are accessible and affordable.Likewise,gender inequalities permeate both the production and consumption of AI.The sur-vey for this Report finds that irrespective of educa-tion qualifications,men are more lik

178、ely than women to use generative AI for work.30Building a complementarity economySeemingly every day,a new AI model exceeds human scores on a narrowly defined benchmark,often bear-ing apocalyptic sobriquets such as Humanitys Last Exam.From this supply-side view humans are framed as one-dimensional b

179、enchmarks in a zero-sum com-petition for finite spots in our future economy an economy of human replacement.Yet incorporating the demand side reveals how policy choices and strategies can promote a complementarity economy,where AI could augment and extend existing human labour,31 yield a more inclus

180、ive labour market32 and lead to new industries,jobs and tasks.33AI can automate tasks that have long remained resistant nonroutine tasks that cannot be accom-plished by some industrial machine.Yet rarely do jobs comprise solely what can be readily delegated to machines.Consider radiologists,who were

181、 viewed a decade ago as at risk of no longer being needed fol-lowing the success of AI in interpreting radiological imagery.Today,demand for radiologists remains as high as ever.34 AI diagnosis is a far cry from deploy-ing medical knowledge in a clinical setting which,even if it were feasible,patien

182、ts might reject.35 A decade on,the story of AI in radiology is one of complementarity improving diagnostics through AI that augments rather than replaces radiologists.36AIs capacity for augmenting human abilities can likewise serve as a vital onramp for economic inclu-sion.For example,AI tends to im

183、prove the perfor-mance of newly hired call centre workers but has lesser effects for seasoned veterans.37 Similar results have been documented in writing tasks,38 software development39 and management consultancy,40 among others.41 Firms are adopting AI for product in-novation more than for process

184、automation and see-ing higher sales,revenue and employment through better outputs.42Figure O.5 Young internet users are struggling everywhere02040608010012014075 andolder657455644554354425341824Average Mental Health Quotient scoreMiddle East and North AfricaWestern EuropeNorth AmericaAge group(years

185、)South AsiaSub-Saharan AfricaLatin AmericaOceaniaNote:Data are from the Global Mind Project at Sapien Labs.The Mental Health Quotient score is a tool that encompasses 47 aspects of mental function as-sessed on a life impact scale that span the dimensions of Mood&Outlook,the Social Self(or relational

186、 aspects),Adaptability&Resilience,Drive&Motiva-tion,Cognition and Mind-Body Connection.The higher the score,the better perceived mental wellbeing.The survey was conducted during 20202024.Source:Thiagarajan,Newson and Swaminathan 2025.OVERVIEW A MATTER OF CHOICE:PEOPLE AND POSSIBILITIES IN THE AGE OF

187、 AI9As AI systems are integrated into jobs,working effectively alongside AI understanding its limita-tions,interpreting its outputs and applying human judgement will be critical.New kinds of tasks and related expertise will be needed at the nexus of peo-ple and machines.Some envision three new roles

188、:ex-plainer,trainer and sustainer.43Yet AI can disrupt and displace work.Robust social protection systems alongside adaptive skills building aligned with emerging needs can improve employ-ment prospects,44 while on-the-job training may sup-port those whose jobs and tasks are reshaped by AI.45 AI sys

189、tems rely heavily on human labour throughout the supply chain,from development and design to data labelling and annotation.46 As an AI-enabled economy expands,social dialogue and collective bargaining are key for new meaningful decent work opportunities.Labour augmentation opportunities,despite thei

190、r big potential,are not inevitable.The digital divide persists,such that access and relevant skills are lim-iting factors for using technology more broadly,and these challenges apply equally to AI in the workplace.Starting nearly a generation ago,digital technolo-gies began suffusing high-income cou

191、ntries,whose workforces today typically enjoy widespread access to digital devices and have extensive experience using them.47 Elsewhere the persistent digital divide is likely to be a major barrier to realizing the positive effects of AI on jobs and beyond.48Looking ahead,people expect AI to both a

192、utomate and augment their work,but they expect the balance to tilt towards augmentation(figure O.7).Whether the expectations for augmentation will be met depends on policies and incentives to catalyse complementary between people and AI.Getting this wrong will lead to development disappointment in t

193、he short term and possibly wider economic divergence in the coming decades.One possibility is averting hasty worker replacement caused by deployment of so-so AI that destroys jobs without generating productivity gains and instead promoting fiscal policies that en-courage augmentation.49Driving innov

194、ation with intentAI can accelerate discovery and innovation and trig-ger new frontiers of creativity,50 potentially becom-ing a method of invention.51 That is,a new tool to Figure O.6 Younger people expect to lose control over their lives due to artificial intelligence(AI)less than older people do20

195、151050510152425343544455960 and olderChange(%)Age groupHigh HDILow and medium HDIVery high HDINote:Based on pooled data for 21 countries.Data show,for each age group,the change in perceived agency as measured by the difference in the percentage of respondents who feel they have a high level of contr

196、ol over their lives today and the percentage who expect to feel a high level of control five years from now,as AI becomes more integrated into everyday life.Source:Human Development Report Office based on data from the UNDP Survey on AI and Human Development.10HUMAN DEVELOPMENT REPORT 2025empower pe

197、ople to fulfil the deeply human aspira-tions to understand and create.Rather than auto-mating tasks in creative processes associated with scientific and technological innovation,the key is augmenting human intelligence52 by leveraging the complementary capabilities of AI and humans to ac-celerate in

198、novation53 and creativity more broadly.54The direction of AI innovation could be steered in ways that align with socially desirable and privately profitable outcomes.55 AI benchmarks have become fundamental tools for evaluating the performance,capabilities and safety of AI models.56 Supplement-ing t

199、he current lot with new standards that assess AIs contribution to human development could help steer AI innovation in that direction.57The complex intersection of different country pri-orities with global and local constellations of tech firms is fuelling a geopolitical innovation race that risks le

200、aving many countries and people behind.58 The mismatch between suppliers and users matters for many reasons.One is cultural.AI models reflect the cultures where they were developed.ChatGPT responses are closer culturally to those of humans in very high HDI countries and most distant from those in lo

201、w HDI countries(figure O.8).Combatting cultural and linguistic bias is one reason many countries desire to be part of the AI supply chain.AI supply depends on three key inputs computing power,data and talent some of which are highly concentrated,posing unique challenges to many lower HDI countries.O

202、nly a handful of voic-es wield power over and through AI.Few of us have much direct say over it.What choices trickle down to us may seem atomizing and binary:buy the latest gadget or not,accept the cookies or not.Take-it-or-leave-it terms of service agreements can boil down to granting powerful firm

203、s carte blanche access to our daily lives or to being excluded from digital plat-forms,where for better or worse ever more of our lives,interactions and relationships take place.Figure O.7 Across occupations and Human Development Index levels,respondents expect that artificial intelligence will both

204、 automate and augment their work with higher expectations of augmentationExpected automation(%)080100Expected augmentation(%)1008020 40 606040200Note:Based on pooled data for 21 countries.Each dot represents the per-centages of respondents in an occupation group in a country who expect au-tomation a

205、nd augmentation from AI to affect their occupation.The following occupational groups are used:professional/higher administrative,skilled,un-skilled/semi-skilled,services,clerical,farm and other.The shaded area repre-sents a higher share of respondents expecting augmentation than automation.Source:Hu

206、man Development Report Office based on data from the United Nations Development Programme Survey on AI and Human Development.Figure O.8 ChatGPT answers are culturally closer to those of humans in very high Human Development Index(HDI)countriesHDI groupLowMediumHighVery highCorrelation between ChatGP

207、T answers andhuman responses0.900.850.800.750.700.650.60Note:Higher values on the vertical axis indicate greater cultural and values similarity between ChatGPT and respondents in a given country(indicated by a dot).Source:Based on data from Atari and others(2025),who compared results across 65 count

208、ries from the World Values Survey.OVERVIEW A MATTER OF CHOICE:PEOPLE AND POSSIBILITIES IN THE AGE OF AI11Narratives that focus on and reinforce only zero-sum thinking crowd out opportunities where coop-eration could add a lot of value.At the global level opportunities for international cooperation o

209、n AI exist,not necessarily on everything but certainly in some specific and important areas.The rationale is especially compelling in computer-provided over-sight,content provenance and model evaluations.59 Indeed,important work across many internation-al institutions and fora are well under way.The

210、 UN Global Digital Compact,which encourages cross-jurisdiction and science-informed dialogue can ena-ble countries to learn from each other and fine-tune regulatory approaches,as well as level the playing field so all countries can meaningfully participate in and benefit from AIs potential.Investing

211、 in capabilities that countTo prepare young people to strive with AI,education needs to focus on learning outcomes,as well as criti-cal,creative and relational thinking,moving beyond simply increasing years of schooling.When integrat-ing AI in education,avoid using AI as a crutch,by teachers or stud

212、ents,and treat it as a companion to unleash new ways of learning.This involves deploy-ing AI to scale interventions known to enhance edu-cation outcomes,such as customized learning,rather than deploying it for its own sake.In healthcare AI should be deployed to comple-ment expertise,particularly whe

213、n it is scarce,as in lower-income countries and settings,empower-ing healthcare workers to do more in resource-and expertise-constrained contexts.60 Healthcare sys-tems and organizations should safely and trans-parently integrate AI technologies strengthening both institutional and frontline provide

214、r capacity to use these systems,while clearly communicating to patients how the systems are employed in clinical decisionmaking to build trust.Because the unintend-ed side effects of AI in health services may change over time,monitoring AI biases and health inequali-ties needs to be seen as continuo

215、us.61New horizons for human developmentScientific and technological progress propel develop-ment.62 Waves of technological innovation have made us healthier,wealthier and more knowledgeable,while shifting patterns of economic opportunity and redraw-ing inequalities.63 Not because of inherent feature

216、s of the technologies,but because of active decisions by people,firms and governments and the incentives shaped by newly created institutions.As AI moves from a niche technology to a cornerstone of peoples lives across multiple domains,its potential to advance human development has to be seized.That

217、 depends on more than algorithms;it depends on our choices.The potential everywhere is big,including in lower HDI countries,whose narrowing development path-ways feel more and more like a development tightrope over a widening chasm.AI can act as a bridge to other advanced technologies that can facil

218、itate industrial up-grading,64 to greater diversification and integration up and down global value chains,65 to better markets for self-employed workers such as freight drivers66 and to new knowledge,skills and ideas that can help every-one,from farmers67 to small business owners.68Of course,that de

219、pends on access not just to“the new electricity”AI but also to the old.Yet tapping AIs potential goes well beyond access,however im-portant it may be.In a world of AI,divides will also spin along another axis:which societies can make the most of a game-changing technology,focusing on how AI compleme

220、nts and augments what people do,and which societies cannot,by either mistaking for it su-percharged extensions of earlier computing technolo-gies or deploying it in ways that compete withpeople.“The future is in our hands.By building a complementarity economy,driving innovation with intent and inves

221、ting in capabilities that count,societies can use AI to expand peoples choices and possibilities.The future is in our hands.Technology is about people,not just things.Beneath the razzle-dazzle of invention lurk important choices,by the few or the many,whose consequences will reverberate across gener

222、ations.By building a complementarity econo-my,driving innovation with intent and investing in capabilities that count,societies can use AI to expand peoples choices and possibilities.In doing so,new development pathways for all countries will dot the horizon,helping everyone have a shot at thriving

223、in a world with AI.12HUMAN DEVELOPMENT REPORT 2025Terms and concepts Agency(human):Peoples ability to hold values,set goals and make commitments that may,or may not,advance their wellbeing.1Agent(AI):An artificial intelligence(AI)system that can autonomously process information,makes deci-sions and

224、complete tasks.2Agenticity(AI):The degree to which an AI agent can autonomously and proactively execute tasks and act as an agent(see above)over extended periods of time.3Algorithmic bias:Systematic errors in AI decisionmaking,often discussed in the context of er-rors that lead to inequitable outcom

225、es,exacerbate dis-parities or reinforce existing patterns of discrimination.4Algorithms:A specified process or set of steps that accomplishes a task,with roots in early mathematics but often used to describe sets of formal instructions provided to a computer.5Alignment:The degree to which an AI syst

226、em ex-hibits consistency with human values,ethics and in-tended outcomes.6Artificial general intelligence:A catchall term for hypothetical AI that exhibits intelligence that gener-alizes across a wide range of contexts.7 However,defi-nitions,feasibility and coherence of the concept itself remain a s

227、ubject of scientific debate.8Artificial intelligence:Software developed to ac-complish things typically associated with human intelligence,from simple rules-based systems to modern generative AI and large language models.9Benchmarks(AI):Quantitative assessments of AI to enable evaluation of its perf

228、ormance,efficiency,capabilities,safety,bias,impacts and other features.10 Chatbots:AI designed to have conversations,ranging from early approaches that relied on explicit rules to more modern large language models and generative AI.Computational machines:Devices that perform mathematical operations

229、ranging from simple tabu-lation and physical computation to advanced modern forms of AI.Computer vision:Techniques,ranging from clas-sical computing to machine learning,for enabling computers to accomplish image-based tasks.11Fine-tuning:Taking an existing model and provid-ing additional training to

230、 adjust,extend or improve its performance.12Frontier models:Although not well defined,often used to refer to cutting-edge,recently developed,ex-citing or particularly capable AI models.13Generative artificial intelligence(including large language models):AI specifically designed to generate informat

231、ion and content such as text,imag-es,videos and protein structures.14Generative pretrained transformers:An ap-proach to developing AI that relies on a pretraining step on large,unlabelled datasets(such as text from the internet)to train a family of models known as transformers.After the initial pret

232、raining,the model is subsequently refined on labelled data.15Hallucination:A term used to describe the possibility of AI generating false information,generating factual-ly correct outputs that are irrelevant to what the user is asking for or generating statements that contradict each other.In genera

233、l,it refers to making statements without regard to the truth.16 For example,AI may create a false fact and trace it to a reference that does not exist.(Human)intelligence augmentation:An approach to developing or using AI that improves humans abili-ty to leverage their own cognitive capabilities.17L

234、abelling:Detecting and tagging training data with additional information to facilitate machine learning.18OVERVIEW A MATTER OF CHOICE:PEOPLE AND POSSIBILITIES IN THE AGE OF AI13Large language model:Forms of AI trained on very large datasets of human-generated text.19Machine learning:An approach to d

235、eveloping AI in which the systems behaviour is not a result of ex-plicit instructions but instead is learned from data or experience.20Model collapse:A phenomenon that occurs when AI is recursively trained on AI-generated data,even-tually resulting in degradation or outright failure of the models pe

236、rformance.21Multimodal(AI):Forms of AI that can process or generate information across multiple modalities,such as audio,text and images.22Neural networks:An approach to machine learning in which computers interact with networks of individ-ual units(neurons)that learn by altering their con-nections

237、to one another over time.23Open source,open data:Software(or perhaps data)for which the code is made publicly available under a copyright licence that enables others to use,study and change the code for any purpose.Parameters:The variables that a machine learning AI model adjusts throughout the cour

238、se of training.Prompt:Instructions provided to generative AI to shape or determine its output.Prompt engineering:The process of developing more complex prompts that better enable AI to pro-duce a desired response.Reasoning or chain-of-thought(AI):A technique for developing large reasoning models tha

239、t,rather than simply generating output,are trained to gener-ate a series of intermediate steps between the task specification and final output.This approach im-proves performance on some benchmark,but debate lingers as to whether these systems are engaging in true reasoning or merely mimicking or ha

240、llucinating the process of reasoning.24Reinforcement learning:A method of training in which various decisions the system(here,AI)makes are associated with different levels of reward.Learn-ing is achieved by adjustments that enable larger re-ward in subsequent steps.Retrieval augmented generation:A t

241、echnique for improving AI responses that enables it to re-trieve information from elsewhere(such as the in-ternet or a dataset)in the process of generating its response.Small models:AI models that are smaller in terms of parameter counts or complexity,often cheaper to train,modify and use.Training d

242、ata:Images,text,video or any other type of data used for machine learning and AI.Turing machine:An abstract model of a computa-tional system proposed by Alan Turing that applies rules to stored information such that it can imple-ment any possible algorithm.NOTES1.UNDP 2024.2.Mukherjee and Chang 2025

243、.3.Mukherjee and Chang 2025.4.Kordzadeh and Ghasemaghaei 2022.5.Chabert and Barbin 1999.6.Ji and others 2023.7.Goertzel 2014.8.Mitchell 2024.9.McCarthy and others 2006.10 Raji and others 2021.11.Ballard and Brown 1982.12.Ding and others 2023.13.Cottier and others 2024.14.Banh and Strobel 2023.15.Ach

244、iam and others 2023.16.Hicks,Humphries and Slater 2024.17.Jarrahi,Lutz and Newlands 2022.18.Kotsiantis,Zaharakis and Pintelas 2006.19.Naveed and others 2023.20.Jordan and Mitchell 2015.21.Shumailov and others 2024.22.Zhang and others 2020.23.McCulloch and Pitts 1943.24.Mitchell 2025.15Empowering peo

245、ple to make artificial intelligence work for human developmentCHAPTER116HUMAN DEVELOPMENT REPORT 2025As artificial intelligence(AI)races ahead,this chapter turns the focus to peoplenot just to those who build AI but to how people everywhere can use it to improve their lives.This is the most relevant

246、 question from a human development perspective.Used in the right way,AI offers an opportunity to expand human capabilities.The chapter challenges unhelpful myths about AI replicating humans and calls for reimagining the relationship between people and this powerful new technology.Despite all the thi

247、ngs that AI can do,it cannot replace human judgement.Thinking beyond replacing humans reveals opportunities for AI to augment human development and enhance the unique contributions of human intelligence,including expanding human scientific and expressive creativity.CHAPTER 1Empowering people to make

248、 artificial intelligence work for human developmentCHAPTER 1 EMPOWERING PEOPLE TO MAKE ARTIFICIAL INTELLIGENCE WORK FOR HUMAN DEVELOPMENT17“Both the technologies developed and the manner in which they are used for exploitation or emancipation,for broadening prosperity or concentrating wealth are det

249、ermined foremost not by the technologies themselves but by the incentives and institutions in which they are created and deployed.”National Academies of Sciences and Medicine 2024,p.84As artificial intelligence(AI)reaches ever more stunning abilities,how will it shape our work,our re-lationships,our

250、 lives?With AI appearing to“reason,”1 will it come after our jobs?Could artificial general intelligence,the pursuit of which is one of humani-tys most ambitious technological endeavours,make people worse off?2 Should we fear that something like artificial superintelligence might wipe out human civil

251、ization?3Rather than try to answer these questions by pre-dicting what will happen,this Report asks what choic-es can make AI work for people.It proposes a human development framework to see how AI differs from previous digital technologies and to navigate the fu-ture of this rapidly changing techno

252、logy,wherever it may go.4 Instead of looking to the future through a foggy fear of the unknown,this chapter invites us to shape that future by knowing more about what AI can and cannot do now and what might be possible as AI evolves.5Examining the demand side of AIMuch policy and media attention foc

253、uses on the supply side of AI which firms and countries will get ahead in the AI race6 and how to ensure that the production and deployment of AI are free from ac-cidents,misuse or systemic negative social impacts7 and grounded in human rights.8 Supplementing these crucial considerations,the main fo

254、cus here is on the demand side of AI,its use across society,examining how it can either enhance or subvert human agency(chapter 2),9 how it is already changing people at dif-ferent life stages,often in harmful ways(chapter 3),and how succumbing to AI hype can exacerbate ex-clusion(chapter 4).The key

255、 reason to consider the user side of AI is that historically the impact of technological innovation on improving productivity and increasing living stand-ards has depended on complementary changes in the organization of economic activity,not simply replac-ing older technologies with newer ones.The c

256、hang-es in the organization of economic production during the transition from steam power to electricity are a well-studied example that has been invoked to ex-plain the lag between the adoption of digital technol-ogies and productivity gains.10 Moreover,only a small fraction of the social value of

257、innovation has been appropriated by the innovators.11 By one estimate digital entrepreneurs of the late 1990s appropriated only about 7percent of the additional value created by new digital firms in the United States alone.12 Ac-counting for the value of digital goods in 13 countries added$2.5trilli

258、on in consumer welfare(or 6percent of their combined GDP),with larger welfare gains accruing to lower income countries and individuals within countries.13Another reason is that people expect AI to be a growing part of their lives.A global survey for this Re-port found that AI use is already substant

259、ial for about 20percent of respondents at all Human Develop-ment Index(HDI)levels.14 But even more stunning,at least two-thirds of respondents in low,medium and high HDI countries expect to use AI in education,health and work the three HDI dimensions within one year(figure 1.1).15The chapter argues

260、that AI represents a technolog-ical inflection point beyond simply having more pow-erful digital tools.AI invites new ways of exploring how economies at all income levels can harness its potential to advance human development.16 But the task is particularly urgent for low-income and many middle-inco

261、me countries,given that the pathways that created jobs at scale and reduced poverty over the past two to three decades,based on expanding manufacturing industries and exporting to interna-tional markets,are narrowing.17 Low HDI countries continue to diverge from very high HDI countries(figure 1.2),w

262、ith many skipping the kinds of structur-al transformation that run through manufacturing,by having employment move straight from agriculture to services rather than shifting to manufacturing in between.18 The narrowing of pathways for low-and middle-income countries is related in part to the automat

263、ion bias of the ongoing digital transforma-tions,but AI offers new options if opportunities to 18HUMAN DEVELOPMENT REPORT 2025complement rather than replace work are explored.19 AI on its own is not a panacea.20 Its impact will de-pend ultimately on whether people,firms and gov-ernments adjust and r

264、eorganize to make the most of it.That includes accelerating the transition to low-carbon economies and supporting the multiple trans-formations historically associated with development(from rural to urban,from home production to mar-ket,from informal to formal,from self-employment to wage work).21Th

265、e chapters three key messages:The value of AI for human development lies not in whether computational machines(machines,for short)are intelligent but in the ways they can augment human intelligence.22AI does some things very well,things that no machine or human has ever done before.But one must avoi

266、d anthropomorphic generalizations that could mislead people into thinking that AI can do everything more capably.23 Some things are best left either to humans or to other pre-AI digital tools.24Comparisons of human and artificial intelligence are fraught with fear,uncertainty and false hope(spotligh

267、t 1.1).25 Whether machines are close to being humanlike(writing a poem)distracts from identifying how to use AI to augment what humans wish to do(helping with poetic expression).26 AI is better than any human at chess,but people still play against each other and are getting better at it with AI.27 A

268、I algorithms have increased music streaming,which has stimulated demand for live performanc-es.28 This suggests that the authenticity of human connections and the need to identify with other hu-mans will remain important,even if machines can Figure 1.1 About two-thirds of survey respondents in low,m

269、edium and high Human Development Index countries expect to use artificial intelligence(AI)in education,health and work within one year14.423.619.066.168.945.9020406080Actual use of AIin the past monthExpected use of AIin one yearExpected increase in useHDI groupLow and mediumHighVery highShare of po

270、pulation(%)Note:Based on pooled data for 21 countries.For actual use in the past month,the following responses to the question,“In the past 30 days,have you ever interacted with artificial intelligence,such as chatbots,in any of the following ways?”were used to calculate the average use of AI for ed

271、ucation,health and work:“education”is based on the response“educational platforms of learning apps,”“health”is based on the response“health care services or applications”and“work”is based on the response“work-related tools or software.”For expected use in one year,the following responses to the ques

272、tion,“Over the next 12 months,how likely are you to use an artificial intelligence tool for the following?”were used to calculate the average use of AI for education,health and work:“education”is based on the response“for education and training,”“health”is based on the response“for medical advice”an

273、d“work”is based on the response“for work tasks.”Expected increase in use is the difference between expected use in one year and actual use in the past month.Source:Human Development Report Office based on data from the United Nations Development Programme Survey on AI and Human Development.CHAPTER 1

274、 EMPOWERING PEOPLE TO MAKE ARTIFICIAL INTELLIGENCE WORK FOR HUMAN DEVELOPMENT19surpass humans in some tasks.29 In fact,it has been argued that the value of the real,the authentic,may increase as AI is more widely deployed.30 Harnessing the human-augmenting power of AI to em-power people requires que

275、stioning misleading narratives that AI can replicate and replace human intelligence.AI goes beyond what earlier digital tools can do.Pre-AI digital tools faithfully executed sequenc-es of steps to automate routines but struggled with things such as recognizing a cat in an image,which AI can now do.A

276、s a result,the scope for potential automation expanded.31 But focusing on automation sells short the potential of humans and machines alike.32 It can lead to deploying what Daron Acemolu called so-so AI33 for things people already do very well,with few if any productivity benefits34 but with job los

277、ses35 and other downsides of AI,including exploitative labour practices in data labelling36 and environmentally stressing en-ergy and material requirements.37More generally,focusing exclusively on automa-tion ignores humans complex multifaceted roles.Passing a medical test,which AI can now do,is far

278、 different from applying medical knowledge in a clinical setting,where contextual awareness and subjective human interactions are critical.38Even if some automation takes hold,AI is also creating new tasks for people,given,for example,its potential to personalize services,as in med-icine.39 AIs wide

279、 availability makes advanced expertise more accessible,40 and open-source AI allows customizing AI to varied local contexts.41 Seeing AI as a new way for humans to take advan-tage of the knowledge others have accumulated over generations42 opens windows for people anywhere to solve problems and purs

280、ue new ven-tures.43 At the same time it creates new challenges,ranging from intellectual property management44 and the compensation of creative workers that gen-erate content used to train AI models45 to concerns over privacy and human rights,which may be made vulnerable in new ways.46 Despite the m

281、any ways AI is useful,its inability to bear responsibility leaves it unable to fulfil many roles in society,creating further demand for AI-augmented human roles.AI can be very good at seeing data patterns that are hard for humans to discern,47 but it is not an oracle that can predict the future.48 I

282、n a courtroom even seemingly accurate AI tools for deciding who should receive bail cannot know whether a given individual truly poses a flight risk.49 Assuming that AI knows that can lead to excessive deference to AI,risking ceding human agency(chapter 2).50Another key reason AI cannot replace huma

283、ns in many contexts is that it bears no responsibility for its actions.51 Knowing that some decisions affecting our lives are made by a real person who is accountable is an irreplaceable feature of so-cial arrangements and one reason people react against automated enforcement of government regulatio

284、ns.52Thinking beyond replacing humans reveals op-portunities for AI to augment the unique contribu-tions of human intelligence,including expanding human scientific and expressive creativity.Human evaluation of AI outputs is often required,particu-larly in high-stakes situations,further expanding the

285、 scope of AI augmentation.For example,in legal and medical applications,given that AI can hallu-cinate(including by producing plausible sounding Figure 1.2 Low Human Development Index countries are being left further behindDiference in HDI value betweenvery high and low HDI countries0.3800.4000.4200

286、.4400.46020242020201520102005200019951992201620202024(projected)0.4100.4000.390Source:Human Development Report Office calculations based on data from Barro and Lee(2018),IMF(2024),UNDESA(2024),UNESCO Institute for Sta-tistics(2024),United Nations Statistics Division(2025)and World Bank(2024).20HUMAN

287、 DEVELOPMENT REPORT 2025but factually wrong statements or generating statements that contradict each other).53 Moreover,having humans interact with AI using regular spo-ken language may introduce ambiguity in what people are trying to achieve.54 What is high stakes(elaborated in chapter 5)is a matte

288、r of individual and social choice,so there is much scope to expand AI augmentation as a result of the need for human evaluation of AI outputs in many situations.In sum,both humans and AI are sold short by no-tions of replacing humans simply because AI can automate some tasks.Instead,AIs potential is

289、 best leveraged to augment human strengths,such as intel-ligence and agency.Automation and augmentation are twin features of the relationship between humans and AI that will determine AIs impact on human de-velopment.In the world of work,the net effect on em-ployment will depend on how the two force

290、s balance out in the short term,on what new tasks are created on longer time scales and on how demand for more efficiently produced goods and services evolves all uncertain but the result of deliberate policy,firm and individual choices.55 The role of choices represents opportunities to make AI work

291、 for people.This is par-ticularly important because most survey respondents are confident that AI will make them more productive at work,and this confidence increases as AI use rises(figure 1.3).An alien intelligence is becoming part of our livesThe novel capabilities of AI particularly generative A

292、I,which showcases remarkable advances in content generation and creative tasks require recognizing that something new has entered peoples lives.That Figure 1.3 Most survey respondents are confident that artificial intelligence(AI)will make them more productive at work,and the more AI is used,the hig

293、her the share of respondents reporting feeling confidentIndex of use of AI for Human Development(last month)Share of population that is confident that AIwill increase their work productivity(%)4050607080080204060100Age group152425343544455960 and olderNote:Based on pooled data for 21 countries.For a

294、ctual use in the past month,the following responses to the question,“In the past 30 days,have you ever in-teracted with artificial intelligence,such as chatbots,in any of the following ways?”were used to calculate the average use of AI for education,health and work:“education”is based on the respons

295、e“educational platforms of learning apps,”“health”is based on the response“health care services or applications”and“work”is based on the response“work-related tools or software.”Confidence that AI will increase productivity is based on respondents who answered“likely”or“very likely”to the question,“

296、You believe AI will increase your productivity at work.”Source:Human Development Report Office based on data from the United Nations Development Programme Survey on AI and Human Development.CHAPTER 1 EMPOWERING PEOPLE TO MAKE ARTIFICIAL INTELLIGENCE WORK FOR HUMAN DEVELOPMENT21something is raising f

297、resh questions because so much about it is unknown and perhaps unknowable.Neu-roscientist Terrence Sejnowski described the appear-ance of large language models such as ChatGPT,a kind of generative AI,56 in this way:A threshold was reached,as if a space alien sudden-ly appeared that could communicate

298、 with us in an eerily human way.Only one thing is clear LLMs large language models are not human.Some aspects of their behaviour appear to be intelligent,but if not human intelligence,what is the nature of their intelligence?57In the near future,and perhaps forever,we will have to grapple with Sejno

299、wskis question.Scientists,philosophers and people in general continue to de-bate whether AI is approaching,or has even already achieved,some degree of human understanding.58 In Sejnowskis framing it seems only right to mix con-cern and optimism for sharing the planet with arte-facts that exhibit int

300、elligence once squarely in our purview.How will AI change us as individuals?As so-cieties and cultures?As a planet?“There are many opportunities for AI to advance innovation and creativity and many options to explore new complementarities between AI and humans without having machines replace humansT

301、here are many opportunities for AI to advance innovation and creativity and many options to ex-plore new complementarities between AI and hu-mans without having machines replace humans.59 AI has the potential to generate demand for new exper-tise and new tasks.60 But using AI may imply difficult tra

302、deoffs.61 For example,how much does society gain from improved scientific output from individual sci-entists using AI compared with the potential loss of variation across these outputs?62 What moral and eth-ical frames do we need to consider if machines can act as moral proxies?63 The interactions b

303、etween AI and humans will play out differently in different cul-tural contexts,64 but large language model responses converge towards particular cultural frames,often those first and fastest across the digital divide.65Amid the myriad ways AI might affect our world mundane,absurd or extreme it can b

304、e easy to feel adrift in the possibilities.Yet Sejnowski firm-ly anchors us:large language models,and AI more broadly,are not human,not even living organisms(spotlight 1.1).From a human development perspec-tive choices should be guided by how to combine uniquely human characteristics with AIs unique

305、 complementary abilities.This will not be effortless.Building and maintaining an augmentative relation-ship with AI are hard.66 Augmentative relationships require moving beyond easy applications that lev-erage AI as a crutch,undermining human intellect rather than augmenting it.67 The rest of this c

306、hapter explores how to do this.AI is better at helping people than replacing themThe vocabulary around AI often misleads starting with the term“intelligence.”While useful for de-scribing AI abilities,intelligence should not imply that machines are acquiring human traits.68 AI is not able to frame pr

307、oblems or act on its own behalf(spotlight 1.1).Because AI can do some things so well,some people assume that humans will not be needed to do those things.It was predicted in 2016 that with-in a decade advances in AI medical imaging would lead to the disappearance of radiologists.69 Extrap-olations a

308、long the same lines continue to posit that artificial general intelligence will leave no work for people.70AI deployment need not replace humansA decade later the prediction about radiologists has been proven wrong.71 By contrast,demand for radi-ologists is growing,with a global shortage at the time

309、 of writing.72 Using AI in a task(reading and classify-ing medical images)did not mean that AI replaced radiologists for many reasons,three of which merit close consideration.73 First,even though AI could exe-cute one task of radiologists,it was useless for several others,including those that are in

310、herently social and require interacting with people74 and those that are constrained by the institutional and organizational features of radiologists work context.75 Second,in-troducing AI to help read medical images created tasks that did not exist before,requiring new skills such as the ability to

311、 understand and interpret the 22HUMAN DEVELOPMENT REPORT 2025recommendations from AI.76 So,using a machine to execute a task can replace but also create tasks.77 Third,having AI classify medical images liberated radiologists time to devote more attention to other tasks,making them more efficient and

312、 effective.78 AI not only failed to replace radiologists;it also failed to reduce the value of their work.79 In the future AI may replace tasks and even occupations digital technolo-gies have reshaped the world of work by doing exact-ly that,and automation tends to reduce employment and wages for in

313、cumbent workers even when the economy as a whole is better off,as we will see later.80Who gets to decide how AI is deployed?AI technical affordances alone do not determine wheth-er AI will be deployed;there must be an organizational reason as well and for firms,a business reason.For example,a recent

314、 study found that while 36percent of US private sector jobs were exposed to automation through AI advances in computer vision capabilities,the economic case made sense for only 8percent.81 But new forms of generative AI are much more accessible and provide greater opportunities for use in a more de-

315、centralized way.For example,even though only 18per-cent of US school districts provide any guidance on AI,60percent of principals and 40percent of teachers used AI in the 2023/2024 school year.82 Among work-ers in 27 countries,almost half used AI every day in 2024,up from about 30percent in 2023.83A

316、I could thus be accessible to the many self-employed workers in low-and middle-income countries.84“The ladder of generality describes the evolution of computational machines as the pursuit of machines that can execute an ever-wider range of tasks(their generality)with less and less human input,direc

317、tion or intervention(human effort)While workers may now have more agency in using generative AI,firms seeking to increase revenue and decrease costs will play a central role in how AI is de-ployed.Deploying technological innovation to reduce labour costs tends to worsen wages and employment for incu

318、mbent workers,even when overall employ-ment and labour productivity rise.85 AI can be de-ployed to automate tasks,much like previous digital technologies,but the economic impact of AI at the firm level appears to come more from greater product inno-vation than lower production costs.86 Perhaps that

319、is why a recent survey found that about a quarter of US firms using AI did so in part to replace worker tasks but two-thirds were not pursuing task replacement.87However,firms might still deploy AI to reduce op-erating costs,including labour costs,particularly if prevailing narratives focus on the b

320、etter-than-human abilities of AI and if AI-producing firms emphasize the benefits of replacing people.88 Seizing on AIs po-tential to augment rather than replace people will not be automatic.89 It will require deliberate choices to re-shape incentives and provide information on what AI can and canno

321、t do.We are on a road to nowhere;come on inside:Taking that ride to intelligence augmentationThe case of AI and radiologists shows that AI has reduced the human effort needed to get a machine to execute a task.At the same time the underlying AI that enhances medical image reading has many other appl

322、ications,such as recording of vehicle li-cense plates and automation of industrial and agri-cultural processes.AI expands the range of tasks that machines can execute.This borrows from Arvind Narayanan and Sayash Kapoors ladder of general-ity,a description of the evolution of computational machines

323、as the pursuit of machines that can execute an ever-wider range of tasks(their generality)with less and less human input,direction or intervention(human effort).90 But where are we now?And what comes next in the evolution of computational ma-chines?We briefly describe four stages,each marked by high

324、er generality and lower human effort than the preceding one(spotlight 1.2):1.Machines with hardware designed for one task(such as digital cameras)Each task requires separate hardware.Low generality(machine designed for one task only)and high human effort(build and operate hardware for each task).2.G

325、eneral-purpose hardware(classical programming)91 One general-purpose computer can handle mul-tiple tasks thanks to software.CHAPTER 1 EMPOWERING PEOPLE TO MAKE ARTIFICIAL INTELLIGENCE WORK FOR HUMAN DEVELOPMENT23 Generality increases substantially but still re-quires writing explicit instructions fo

326、r each task or domain of tasks;human effort to have the machine execute tasks is reduced to the need to operate the software.3.Machine learning(pregenerative AI)Instead of coding tasks in full detail,feed the machine data from which it can learn a task,or let the machine learn from known rules by in

327、ter-acting with itself.Generality expands further to tasks that are hard to specify with instructions;human effort declines because of the greatly reduced need to operate software.4.Generative AI Leverages large datasets spanning text,video,images and sound.Generality is so broad that it spans draft

328、ing texts,writing computer code,composing music and translating languages;human effort is lower be-cause minimal user direction using regular writ-ten or spoken language is required for the task to be executed.92Humans have long imagined computational ma-chines.Talos,an automated guardian robot was

329、ide-alized in Greek mythology more than 2,500 years ago.93 We began to bring such science fictions to life at the dawn of the electric age in the 19th cen-tury,enabling automation of once uniquely human information-processing tasks by constructing com-putational machines,such as the Hollerith tabula

330、-tion machine that helped process the 1890 US census(spotlight 1.2).94 That machine was characteristic of the first stage:computational devices built with spe-cific hardware from scratch to execute a single task.Generality is low,and the corresponding human ef-fort to automate a given task high,beca

331、use hardware needs to be built for each task.Such hardware is still with us digital cameras,automated teller machines,many medical devices and internet switches.Todays programmable computers,in which a com-puter(one piece of hardware)can be preprogrammed to execute many different tasks,correspond to

332、 classi-cal programming,the second stage(spotlight 1.2).95 This vastly increased the generality of tasks that a machine can execute and reduced the human effort required to do so.With AI the nature of effort to offload tasks to a ma-chine has changed yet again,reaching a third stage,extending genera

333、lity further to tasks difficult for clas-sical programming to execute.Rather than relying on written code,systems learn their functionality from a corpus of data(think of data as examples that train the machine):this is the basic idea of machine learning,which has yielded multiple applications(table 1.1).The most recent stage is the availability of large language models and other forms of generati