《喬治梅森大學&ampamp史汀生中心：2023全球人工智能基礎設施（AI Infra）報告（英文版）（49頁）.pdf》由會員分享，可在線閱讀，更多相關《喬治梅森大學&ampamp史汀生中心：2023全球人工智能基礎設施（AI Infra）報告（英文版）（49頁）.pdf（49頁珍藏版）》請在三個皮匠報告上搜索。

1、GEORGE MASON UNIVERSITY AI STRATEGIES TEAM&THE STIMSON CENTERPRESENT2023 GLOBAL ARTIFICIAL INTELLIGENCEINFRASTRUCTURES REPORTAuthors:J.P.Singh Amarda ShehuCaroline WessonManpriya Dua With a Foreword from David A.Bray2023 Global Artificial Intelligence InfrastructuresReportJ.P.Singh1,Amarda Shehu2,Ca

2、roline Wesson1and Manpriya Dua21Schar School of Policy and Government,George Mason University2Department of Computer Science,George Mason UniversityForeword from David Bray33Distinguished Fellow,The Stimson CenterSuggested citation:Singh,J.P.,Amarda Shehu,Caroline Wesson,and Man-priya Dua.The 2023 G

3、lobal Artificial Intelligence Infrastructures Report.Witha Foreword from David Bray.AI Strategies Team and the Institute for DigitalInnovation,George Mason University,and the Stimson Center,Washington DC.August 2023.This research is supported by a$1.389 million grant from the Minerva Research Initia

4、tive.Please send queries on this report to aipolicygmu.edu1Contents1The AI Strategies Team at George Mason42Foreword:Ensuring We Build the Right Foundation to Evalu-ate Trust in AI and Societies53Summary&Recommendations104Introduction134.1National Landscapes.134.2Stages in the Development of AI Stra

5、tegies.144.3Existing Narratives in AI Infrastructures.144.4AI Wardrobes.165Methods and Data-Set175.1Our Dataset:Introducing the National AI Policies.196The Empirical Findings206.1Comparing National AI Infrastructures:Policy Priorities Revealed 206.2Comparing Intra-national AI Strategies:Policy Depth

6、.266.3Document Level Analysis:Analyzing the 5 Most Important Top-ics in Our Intra-national Documents.337Conclusion428Team Biographies43Bibliography462List of Tables1Clusters,Topics,Top 10 Words,and Countries.232Top 10 Countries for Number of AI Documents.263Topics for Leading AI Countries and EU.294

7、Main Topics Across Documents.34List of Figures1The AI Wardrobe.162A sample word cloud of a topic based on education and talent.The font size of each word follows the probability assigned by thealgorithm to that word in this particular topic(as determined byanalysis of the documents in the corpus).18

8、3National AI Infrastructures:Heat-map relating country-topicdistributions.214National AI Infrastructures Topic Word Clouds Topic 1 does nothold much significance in the corpus and is omitted.225Intra-National AI Infrastructures:Heat-map relating country-topic distributions.276Intra-National AI Infra

9、structures Topic Word Clouds.287Intra-national Policy Documents:Topic Probabilities Per Country 308Country Counts Per Topic.309Top Panel:Datasets and Governance Topic Word Cloud(topic1 out of 5 topics learned by the model over intra-national doc-uments).Bottom Panel:Datasets and Governance DocumentC

10、loud relating the probabilities with which this topic emergesover countries(national and intra-national documents),utilizinglarger font to visibly relate a document where the topic is mostprominent(probabilistically).3510Top Panel:Education and Training Topic Word Cloud.BottomPanel:Education and Tra

11、ining Document Cloud.3611Top Panel:Economy Topic Word Cloud.Bottom Panel:Econ-omy Document Cloud.3712Top Panel:Contracts Topic Word Cloud.Bottom Panel:Con-tracts Document Cloud.3813Top Panel:Transport Topic Word Cloud.Bottom Panel:Trans-port Document Cloud.3914Education versus Talent Approaches.4031

12、The AI Strategies Team at George MasonThe AI Strategies team at George Mason provides evidence for how culturalvalues and institutional priorities shape artificial intelligence(AI)infrastructuresin national and global contexts,in order to better understand the effects ofcomparative AI contexts for s

13、ecurity.AI Strategies is funded by a three-year,$1.39 million grant that was awardedto George Mason University to study the economic and cultural determinants forglobal artificial intelligence(AI)infrastructuresand describe their implicationsfor national and international security.The team began wor

14、k on the project onApril 15th 2022.This is the first yearly AI infrastructure report.The grant was awarded by the Department of Defenses reputed MinervaResearch Initiative,a joint program of the Office of Basic Research and the Of-fice of Policy that supports social science research focused on expan

15、ding basicunderstanding of security.Team MembersCurrently the team is comprised of three social scientists,two computer scien-tists,a computer science PhD student,and three political science,public policy,and philosophy doctoral and Masters students.Given this diverse team,theresearch reflects uniqu

16、e analytical creativity:the team members have workedtogether for over a year on this research,and learned to build off one anothersstrengths to understand the landscape of national AI infrastructures and howto apply NLP methodologies to empirically base their comparisons and con-textualize the subje

17、ct matter and country expertise.Read our research teamsbiographies in Section 8.42Foreword:Ensuring We Build the Right Foun-dation to Evaluate Trust in AI and SocietiesDear Readers,As a field,Artificial Intelligence(AI)has been around since the mid-20th cen-tury.In 1955,U.S.computer scientist John M

18、cCarthy coined the term.Laterin 1959,McCarthy collaborated with Marvin Minsky to establish the ArtificialIntelligence Project,nowadays MITs CSAIL(Computer Science and ArtificialIntelligence Laboratory).In parallel to McCarthy and Minsky,U.S.politicalscientist Herbert Simon completed a PhD in 1943 ex

19、ploring decision-makingin administrative organizations and pursued research that later influenced thefields of computer science,economics,and cognitive psychology.In 1957,Simonpartnered with Allen Newell to develop a General Problem Solver separatinginformation about a problem from the strategy requ

20、ired to solve it.All four individuals McCarthy,Minsky,Simon,and Newell would go onto receive the ACM A.M.Turing Award during their respective careers.In thealmost six and a half decades that followed,AI research developed several flavorsof systemic approaches to include:Logical Reasoning and Problem

21、-Solving Al-gorithms,Expert Systems,Statistical Inferences and Reasoning,Decision Sup-port Systems,Cognitive Simulations,Natural Language Processing,MachineLearning,Neural Networks,and more.Though AI has many subcategories and has had many flavors of approachessince the 1950s,within the last few yea

22、rs,a subset of Neural Networks built onthe transformer architecture have revolutionized natural language processingand given rise to what are now known as Large Language Models(LLMs).Justin the last year,LLMs such as ChatGPT and variants,have activated significantpublic interest,excitement,and anxie

23、ty with regards to the future of AI.Whilethe full extent of the public,business,community,and individual value of LLMsremains to be seen,the ability of these models to provide responses to effectiveengineered prompts regarding the generation of predictive text,synthesized im-ages,as well as the full

24、 gamut of multimedia audio and even video outputs hascaptured the public zeitgeist.I.A valuable compass reading as to where different nationshave decided to steer approaches to AIPundits globally have indicated both excitement and concerns about whethermachines may be able to perform work previously

25、 thought only performable byhumans as well as whether they may be able to produce content and interactionsthat appear human.It is precisely at this moment that this 2023 Global Artifi-cial Intelligence Infrastructures Report by J.P.Singh,Amarda Shehu,and theirdoctoral students is so prescient.By bri

26、dging together multiple fields,includingthe best of computer science,economics,political science,and public policy,ina collaborative manner akin to the best work of Herbert Simon Singh and5Shehu have produced a valuable compass reading pointing to where differentnations have decided to steer their a

27、pproaches to AI for the future ahead.Theirreport presents both rigorous and much needed insights that demystify some ofthe current fervor around the future AI and societies,namely:First,the report shares convincing evidence that humanitys AI-associatedfuture will not be set by just the United States

28、 and China alone there existdifferent AI strategies being pursued by multiple nations beyond just thesetwo large nations,with different objectives and proposed paths outlined inthese national AI policies.Second,while there is no singular grand strategy across the fifty-four nationalAI plans analyzed

29、 in this report Singh and Shehu find similar choice elementsin the national strategies analyzed.The researchers dub these similar choiceelements a collective AI Wardrobe,a term coined by Caroline Wesson,one ofthe doctoral students in the team,to relate the various choices each countrycan make in ass

30、embling a tailored AI national strategy outfit.Third country clusters are apparent among the different national strategiesthat were analyzed for this report to include the European Union,East Asia,Spain leading a Latin America cluster,and the United Kingdom leading aBritish influence cluster.Whether

31、 or not these clusters will result in closer AI-related business interactions,nation-to-nation civil relations,and geopoliticalties amongst the countries more closely aligned with regards to their nationalAI strategies represents a crucial area to watch both now and in future.II.Building the necessa

32、ry foundation an interdisciplinarymix of fields to tackle Trust,AI,and SocietiesJuxtaposed against the global zeitgeist regarding AI,this important 2023 reportexists amid a deeper milieu of important questions regarding trust within andacross nations.In October 2017,the Pew Research Center found tha

33、t lessthan forty-five percent of residents living in the United States under the age oftwenty-five years old thought capitalism was in their opinion a good force insociety.Contemporary studies at the time also found declining levels of trustamong a similar age demographic in the essentialness of liv

34、ing in a democracy not just for the United States but also for Sweden,Australia,the Netherlands,New Zealand,and the United Kingdom.Together these global trends blend tocreate a central question namely:Can nations invoke strategies that result in Trust in AI and societies?and a corollary:Can nations

35、encourage Trust in AI and societies,while facinggrowing distrust in their economic and political systems?Readers should note that trust can be defined as the willingness to be vul-nerable to the actions of an actor not directly controlled by you a definitionthat works for both human and AI actors.Mu

36、ltiple studies have establishedthat the antecedents of trust include the perceptions of benevolence,compe-tence,and integrity of the actor to an individual.If perceptions of these three6antecedents are positive,then trust is more likely.If perceptions are negativeor absent,then trust is less likely.

37、The clustering of similar choice elements in this report,specifically the setof elements that comprise the reports described“AI Wardrobe”,represent animportant tool for leaders in the public and private sectors to assess if a na-tional AI strategy has the requisite elements to address challenging qu

38、estions ofimproving Trust in AI and societies.Cumulatively,this question of Trust in AI and societies represents an essen-tial one for nations AI strategies with regards to their expressed objectives.Interms of expressed objectives,though LLMs and their outputs have capturedthe current public consci

39、ousness of 2023,there are so many more outcomes forwhich AI can be employed by nations,communities,and networked groups ofpeople working to shared outcomes beyond just generative content.Readersare invited,after seeing the analysis and results in the report,to consider moreexpansive objectives for A

40、I and societies,to include exploring how:Can AI improve human understanding of decisions we need to make now?Can AI help improve understanding the impact of our decisions(or lackthereof)on possible local and global futures?Can AI help improve human collaborations across sectors and geographies,poten

41、tially tipping and cueing humans that there are other humans withsimilar projects underway?Can AI help improve identification and reconciliation of misaligned goals andincentives be they community,regional,or global for important peace-keeping activities?Can AI help improve public safety,internation

42、al security,and global pre-paredness for disruptions both natural and human-caused in the world?Can AI help improve the operations and resilience of networked,digital tech-nologies for both organizational and public benefit especially in an era ofincreasing internet devices?Can AI help improve the“e

43、ssential fabrics”of open societies to include free-dom of speech,freedom to think differently,and the need for an educatedpublic to help inform pluralistic discussions all amid a digital tsunami ofdata?Can AI help improve education,focus,and entrepreneurial activities to tacklebig,thorny,“hairball”i

44、ssues like climate change,immediate&long-term foodsecurity,natural resources,and future sustainability for a planet of 9+billionpeople?These important questions represent a few of the important,shared out-comes to be explored and achieved through AI strategies that bring togetherhuman communities.Wh

45、ile this report does not answer them all,it does in-dicate the different objectives being pursued by different nations with regardsto their AI strategy as well as their performative declarations meant for thebroader international community.Furthermore,this report both embodies anddemonstrates the im

46、portance of interdisciplinary teams for AI research and AI7education.Working across multiple disciplines is essential for both research andeducation especially as policymakers,business leaders,and students alike learnto explore and advance the necessary AI technical,commercial,civil,and ethicalconce

47、pts required for a more positive future ahead.III.For Trust,AI,and Societies,what if the Turing Test isthe wrong test for AI?This report represents a vanguard assembly of an interdisciplinary mix of fieldsto include the best of computer science,economics,political science,and publicpolicy.Ultimately

48、 for AI to succeed in benefiting nations,communities,andnetworked groups of people,we must understand human nature more.Wehumans are products of natural selection pressures.Darwinian evolution isakin to a“blind watchmaker”and as a result evolution has not preparedus to encounter the true alienness o

49、f AI.It is risky for humans to think AI isaligned to the same things we want and value,especially when the alignmentproblem of an AI to specific outcomes remains an unsolved challenge for severalneural network approaches.In addition,we humans anthropomorphize lotsof things including animals,weather,

50、inanimate objects,as well as machinesand now AI even if those things do not act,think,or behave at all like ushumans.Furthermore,training an AI depends heavily on the datasets employed,meaning both extant human datasets as well as our human choices regardingAI may amplify some of the more socially b

51、eneficial or detrimental elementsof human nature.These elements include the considerable number of knownhuman biases that each of us possess,to include confirmation bias,sunk costbias,“in vs.out group”biases(aka,xenophobia),and many more biases though,fortunately,these biases can be mitigated some b

52、y education and experiences.By both providing a valuable compass reading as to where different nations havedecided to steer their approaches to AI for the future ahead,and building thenecessary foundation for bringing together an interdisciplinary mix of fields tostudy national AI strategies Singh,S

53、hehu,and their students enable readersto ask what I professionally consider to be the crucial question of the 2020s,specifically:what if the Turing Test is the wrong test for AI?It is important to remember the original Turing test designed by computerscience pioneer Alan Turing himself involved Comp

54、uter A and Person B,withB attempting to convince an interrogator Person C that they were human,andthat A was not.Meanwhile Computer A was trying to convince Person C thatthey were human.In reading the findings and conclusions of this 2023 report,I invite readers to consider what if this test of a co

55、mputer“fooling us”is thewrong test for the type of AI that our society needs,especially if we are toimprove extant levels of trust among humans and machines collectively?After all,consider the current state of 2023 LLMs where benevolence ofthe machine is indeterminate,competence is questionable as e

56、xisting LLMs arenot fact-checking and can provide misinformation with apparent confidence andeloquence,and integrity is absent as the LLMs can with some variability changetheir stance if user prompts ask them to do as such.These crucial questions8regarding the antecedents of trust associated with AI

57、 should not fall upon thesedigital innovations alone.First,these are systems designed and trained byhumans.Second,ostensibly the 2023 iteration of generative AI models willimprove in the future ahead.Third,and perhaps most importantly,readerswho care about the national AI strategies present in 2023

58、around the world alsoshould carefully consider the other“obscured boxes”present in human societies,such as decision making in organizations,community associations,governments,oversight boards,and professional settings.All of which brings us back,in conclusion to the earlier corollary to thecentral q

59、uestion of Trust in AI and societies,namely:Can nations encourageTrust in AI and societies,while facing growing distrust in their economic andpolitical systems?It could be that for the near future,both members of thepublic and representative leaders both in the public and private sectors needto take

60、 actions that remedy the perceptions of benevolence,competence,andintegrity namely Trust both in AI and societies(sans AI)simultaneously.As mapping positive,deliberative paths forward to improve the state of Trustin AI and Societies is important,this important 2023 report delivers a prescientview of

61、 the current expressed state of fifty-four different national AI strategiesto help us understand the present and consider the next steps necessary for thefuture ahead.David BrayDavid BrayDr.David A.Bray is a DistinguishedFellow with the Stimson Center and theBusiness Executives for National Security

62、(BENS).He also is a CEO for different“under the radar”tech and data ventures and has served in a variety of leadershiproles in turbulent environments,includingbioterrorism preparedness and response,Ex-ecutive Director for a bipartisan NationalCommission on R&D,non-partisan leader-ship twice global C

63、IO 100 award-winner,work with the U.S.Navy and Marines,and advisor to the U.S.Special Operation Command on the challenges of coun-tering disinformation online.He has received the Joint Civilian Service Com-mendation Award and the National Intelligence Exceptional Achievement Medal.David served as Ex

64、ecutive Director for the People-Centered Internet coalitionChaired by Internet co-originator Vint Cerf and is a Senior Fellow with the In-stitute for Human-Machine Cognition.Business Insider named him one of thetop“24 Americans Who Are Changing the World”and he was named a YoungGlobal Leader by the

65、World Economic Forum.He previously gave the AI WorldSociety Distinguished Lecture to the United Nations on UN Charter Day.93Summary&RecommendationsIn 2016,the United States published its National Artificial Intelligence Researchand Development Strategic Plan,usually understood in policy communities

66、asthe first statement of its AI infrastructure strategy(Select Committee on Ar-tificial Intelligence,2016).Since then over 60 countries have announced theirnational or sectoral AI policies.This report employs computer science techniques to analyze the publishednational AI plans of 54 countries.In ot

67、her words,we employ AI to analyze AIstrategies.The report includes an analysis of 213 documents on AI strategies.Apart from national plans,the set includes reports and publications from variousgovernment departments,ministries,nation commissions,bodies appointed toforward recommendations for specifi

68、c issues and sectors.Our computer science methodology,specifically Latent Dirichlet Analysis(LDA)(Blei,Ng and Jordan,2003),is calibrated to recognize embedded orlatent topics that each document contains.It does so through providing prob-abilities of words that are most likely to occur together in ea

69、ch document.Alldocuments are analyzed together for a pre-specified number of topics,ascer-tained through rigorous methodological criteria.The choice of the number oftopics reflects fulfillment of various methodological LDA criteria for model sta-bility(consistency)and topic stability(coherence).A do

70、cument may featurea dominant topic,or a document may contain two or more topics.Further,we employ a technique known ensemble-LDA(e-LDA)to provide stable resultsassessed over multiple model specifications.Collectively we present the most detailed and comprehensive empirical anal-ysis undertaken of na

71、tional AI infrastructures to date.This analysis providescomparisons and contrasts across 54 national strategies and a granular lookat what these strategies contain.We note the priorities that are containedin documents,but our analysis also points out the policy depth for particularcountries.Policy d

72、epth refers to the extent to which countries have coveredthe entire gamut of issues that comprise an infrastructure,and the institutional10and financial resources they have committed to these issues.For example,AIpolicies from leading powers such as United States and China contain depthfor basic res

73、earch capabilities in science and mathematics,while the EuropeanUnion policies contain the most depth for data governance and ethics.For ex-ample,one of the strategic objectives stated in the Chinese AI strategy states:“by 2025,China will achieve major breakthroughs in basic theories for AI,suchthat

74、 some technologies and applications achieve a world-leading level and AIbecomes the main driving force for Chinas industrial upgrading and economictransformation”(State Council,2017).We make three major claims:There is no grand strategy or conclusion that applies to all AI infrastructures.Countries

75、and clusters of countries feature different objectives and how toachieve them.Countries are pursuing a variable mix of similar elements in their nationalstrategies.We propose and utilize the concept of AI Wardrobes to show thevarious elements available for putting together an AI infrastructure and t

76、hevariable ways in which countries are putting together these wardrobes.Clusters of countries pursuing similar strategies are identifiable.Our machinelearning algorithms are able to point out some obvious clusters from theEuropean Union,Latin America,and East Asia.But there are also surprises.United

77、 Kingdom leads a British influence cluster.Spain is prominent in theLatin American cluster.Our three major claims are made at three different levels:We analyze 54 plans that are taken to be national.These are often perfor-mative.They are as much about national priorities as they are declarationsmean

78、t for the international community.But they reveal the broad trajectoryand differences among national strategies.11 We analyze 213 documents including the national plans that national gov-ernments,commissions and departments have published on their AI infras-tructures.Unlike,the performativity and di

79、fferences among national plans,the intra-national plan reveal fewer national differences but a few countrieshave more policy depth than others.We notice countries that are at the earlystages of policies regarding their AI infrastructures,versus those that havedetailed regulatory and sub-sector polic

80、ies.We also analyze the 213 documents,regardless of country labels,and here wesee the broad topics that stand out in country plans.These include trans-portation,education,data ethics,and regulation.Looking at the documentswe can then understand the countries that dominate these topics and alsosome b

81、road differences among them.Based on our analysis we present three policy recommendations:Comparative analyses like ours provide countries sign posts and guidelinesfor their ambitions.There is no one size fits all for designing national AIinfrastructures.Different countries have different capabiliti

82、es and priorities.Regulating AI will depend on country preparedness and political systems.Grand pronouncements such as fears about sacrificing our human rights orprivacy to machine-led systems in our media about AI need a reality check.Several countries,generally with democratic systems,are putting

83、together orstruggling to put together systems of accountability,while others barely fea-ture any such concerns.This provides room to think about governance issues,rather than ceding this authority to machines(or corporations)prematurely.AI policies have many good stories to tell about service provis

84、ion.These in-clude AI applications for health,education and research,and transportation.124IntroductionNational governments regularly publish strategies for economic development,healthcare,transportation,and other areas of importance for improving gov-ernance and the lives of citizens within the cou

85、ntry.As artificial intelligence(AI)has become an increasingly important and revolutionary technology coun-tries have started to draft and publish national AI strategies to articulate theirdirection and vision for the technology and its application to different sectorsof society.These strategies also

86、 assist governments with goal setting,resourceallocation,organizational coordination,and international cooperation in the de-velopment of AI infrastructures and technologies.Our 2023 AI Infrastructures Report identifies which values,priorities,goals,and policy mechanisms have influenced the developm

87、ent of national AI infras-tructures,as represented through national AI strategies.We utilize the term“AI infrastructure”to refer to machine learning(ML),natural language pro-cessing(NLP),and associated technologies that bring together data and algo-rithms which touch all aspects of human life includ

88、ing ordering goods,reviewingemployment and housing applicants,diagnosing diseases,regulating traffic pat-terns,and instituting drones and robots as elements of warfare.National AIstrategies give us insight into the complex narratives of technological diffusion,security concerns,and the implications

89、AI will have for prosperity.First,this policy report will provide an overview of the landscape of thesepolicies which includes both national policies,intra-national policies,and insome cases international policies and agreements.Second,the report brieflydescribes the research methodology.Our researc

90、h employs a natural languageprocessing methodology,Latent Dirichlet Allocation(LDA),which produces setsof topics that capture the key words with probability distributions that occurtogether for each document in a designated set of documents.This methodologygoes further than the typically used text-m

91、ining techniques by employing AIbased techniques that consider probability and context of words rather thanjust word frequency.Next,the majority of this policy report will provide anoverview of our results on the 54 national AI strategies at three levels:thenational level accounts for the top-level

92、policy document,the intra-nationallevel analyzes several policy documents in each country,and a third level thatanalyzes all documents collectively to see which topics stand out globally andthe main country documents that address these topics.Although we include theEuropean Union in our analysis of

93、documents,we use the shorthand countriesfor our analysis.4.1National LandscapesIn 2017,the first set of countries published national artificial intelligence(AI)infrastructure plans,the first published came from Canada and China(StateCouncil,2017;CIFAR,2017).Meanwhile,the United States government pub

94、-lished a key document in 2016,the National Artificial Intelligence Research andDevelopment Plan,which later helped to inform the 2019 and 2023 updates of13this document and several other additional plans authored by the White House,Department of Defense,and other departments and offices(Select Comm

95、itteeon Artificial Intelligence,2016;”Select Committee on Artificial Intelligence”,2023).Other countries which were first movers in publishing their AI strate-gies include the European Union(2018),Germany(2018),and India(2018)(NitiAayog,2018;German Federal Government,December 2020;European Commis-si

96、on,2018).Currently there are over 50 countries which have published nationalAI strategies.These countries are diverse geographically,by income,and tech-nological capabilities,but there are only two countries from Sub-Saharan Africa Mauritius and Uganda.These national AI strategies articulate the pre

97、vious,current,and future efforts of the authoring nation to shape and support thedevelopment of AI.Many countries have also published intra-national strategy documents,whichare strategy documents from directorates,departments,and commissions.Thesedocuments are typically specific to the authoring org

98、anizations area of focussuch as education,start-ups and business-expansion,cybersecurity,data pro-tection,health,and transportation.Some examples of such documents includeAustralias“Artificial Intelligence and Emerging Technologies in School”strat-egy published by the Australian Department of Educat

99、ion and Training(2019)or Chinas“Regulations for the Promotion of the Development of the ArtificialIntelligence Industry in Shanghai Municipality”which was published by TheStanding Committee of the 15th Shanghai Municipal Peoples Congress(2022).There are varying levels of overlap between these strate

100、gies and greater nationalAI strategies.Ultimately though,these strategies also have important implica-tions for the development of AI infrastructures within a country,and specificallyfor the infrastructures directly related to the targeted domain area.4.2Stages in the Development of AI StrategiesGiv

101、en the recent advent of national AI strategies,some countries have more de-veloped articulated strategies than others.Some national strategies are focusedon essential infrastructures needed to develop and deploy AI systems,whereasother strategies articulate elements of how specific ministries and of

102、fices willcreate the next step in AI infrastructures,technologies,and regulatory envi-ronments.In many cases,AI infrastructures in one country are influenced bysimilar steps in other countries.Our results,presented later in this report,haveidentified this empirically.4.3Existing Narratives in AI Inf

103、rastructuresThere are many overarching narratives that surround the development of AI,national AI plans,and the different ways in which AI can be employed.Thesenarratives have made their rounds through the media,politicians speeches,po-litical platforms,debates,and in some cases,policy.This research

104、 seeks tounderstand what narratives exist in AI policies around the world.Our method-ology for doing this is detailed later in this report.It is important though to14address some of the dominant narratives in this space,as they directly relate tothe narratives which emerge from this research.Many na

105、tional AI plans tout the uniqueness of AI as a revolutionary technol-ogy which we have yet to see before.This statement is said in a matter-of-factmanner,without much concern for the historical progression of technology.Wehave had revolutionary technologies before the printing press,electricity,thei

106、nternet,and others what makes AI so different?This question is not typicallyanswered,but instead the argument is made by pointing out the many differentaspects of life which will be changed by the roll out and continuous innovationof AI.To push back against this narrative is not to reject the massiv

107、e impactsAI will have on life as we know it,but it is important to realize that other tech-nologies have been introduced into society that have completely changed theway in which we live.We have thus adapted to these technologies and learnedhow to use existing institutions and governance structures

108、to manage how thesetechnologies impact life.It will ultimately be up to a variety of societal andgovernance institutions to determine if AI is a job replacer or augmenter,a toolfor social good,or a tool for a dystopian future.Many of these national AIplans seek to deal with shaping the future of how

109、 AI impacts everyday life but it is important to note that what narratives they employ determine theirultimate policy trajectory.Another major narrative of AI involves the potential for the technology tobe utilized by governments as a tool for oppression.A concern often articulatedin various media i

110、s that in authoritarian states,governments will utilize AIto surveil,police,censor,suppress,and manipulate their populations.Thereis empirical proof that this has occurred in a few countries already.What isimportant to note though is that these tools will be available to all governments,authoritaria

111、n or democratic.Empirically what we demonstrate later in ouranalysis of the national AI plans is that pluralist states are more likely to openthe process for developing systems to govern and support the development androll out of AI with consultation from civil society.This is less true of non-plura

112、list or authoritarian governments.We do not refute the latter narrative,but more research needs to be done on the mechanisms which lead to the use ofAI tools by governments for these purposes.Regardless,there is overlap betweenhow pluralist and non-pluralist countries seek to develop AI and implemen

113、t itfor the purposes of economic development,security,and other broader uses.Our machine learning based analysis of the national AI plans around theworld better understands what narratives are influencing national strategies todevelop and deploy AI.Utilizing our methodology we identify the ways in w

114、hichspecific narratives might lead to different mechanisms to shaping AI infrastruc-tures.It is important to note that many stories about AI in popular discoursetake a negative tone.This report takes an evidence-based approach to veer awayfrom both negativity or overly optimistic AI scenarios.Many o

115、f the national AIplans emphasize the importance of inclusion,diversity,and transparency as thetechnology continues to develop and be deployed.154.4AI WardrobesFigure 1:The AI WardrobeTo best understand the way in which countries build their national AI policies,we have created the concept of AI ward

116、robes.Using this concept,we argue thatthe universe of national AI strategies can be conceived as a global wardrobe anda countrys national AI plan is an outfit.How countries present their national AIinfrastructure narrative or story is an outfit which they selected from the shareduniverse of options

117、within the AI wardrobe.Therefore,within the wardrobe fornational AI strategies there are concepts such as“trustworthy AI,”“inclusion,”“data privacy,”“R&D strategies,”and“international collaborations”.Theseconcepts are different articles of clothing and just as in reality,clothing cancome in differen

118、t patterns,shapes,and with different embellishments.Countriescan select which policy elements they want to utilize.In some cases,countriesmight simply pick a“shirt”such as emphasizing“workforce training”and wearit.In others,a country might adapt it to their needs,stylizing it with a pin or abutton t

119、o represent their own distinct flair and style.China places a great dealof emphasis on“talent”in its education strategies.Some clothing pieces in thewardrobe also might be“hand-me-downs”or elements which countries directlycopy or duplicate from others.Many EU countries adopted the language on16inclu

120、sion and diversity from the EUs own plan.Meanwhile,other pieces ofclothing might be newly produced,or original ideas.Indias start-up policies orGermanys emphasis on mittelstand or SME policies are examples(Niti Aayog,2018;German Federal Government,December 2020).Ultimately,countries getto decide wha

121、t elements of AI policy to put into their national AI strategy,andthey get to determine how to put these pieces together to create their own outfitin their own distinct style.It is our goal with this research to ground the narratives surrounding thedevelopment and deployment of AI in empirical analy

122、sis.To do this we havedeveloped a methodology which utilizes computer science techniques,specificallyLDA and eLDA.This is detailed in the next section of this report.5Methods and Data-SetThere is existing research which compares AI infrastructures,but much ofit relies on more traditional methodologi

123、es,which include close text readings(Bareis and Katzenbach,2022)and keyword frequency analyses in combinationwith other qualitative methods(Robinson,2020;Wilson,2022;Foffano,Scant-amburlo and Cort es,2022).Some uses of natural language processing(NLP)methods have been used(Hine and Floridi,2022),but

124、 the corpus of input textsis only a subset of national AI plans rather than the universe of national AIplans,which our research covers.Additionally,much of the work in this spaceis focused on two country comparisons with a heavy emphasis on the U.S.andChina.In March 2023,a report entitled Building T

125、rust in AI:A Landscape Analysisof Government AI Programs by Susan Aaronson was published by the Centrefor International Governance Innovation(Aaronson,2023).This report lever-ages the OECD AI policy website and database(discussed in this section)andasks important questions on government efforts to d

126、evelop AI capabilities andtrustworthy AI,evaluate their own efforts in developing AI,and the develop-ment of best practices on AI and trustworthy AI at the OECD.This researchis comprehensive and provides an overview of different policy approaches to AIaround the globe,but it has a different purpose

127、than the research of our team.The CIGI research is more focused on broad policy mechanisms,evaluation,and the development of best practices for policy rather than understanding theimportant values and narratives which motivate the development and implemen-tation of AI policies at a micro level.Our r

128、esearch uncovers the motivationsand narratives which exist in national AI policies and compares these narrativesacross countries,which yield data that can help us better understand the globaldiffusion of values in AI infrastructures.We do this using a computer sciencemethodology,LDA(Blei,Ng and Jord

129、an,2003).Our LDA methodology employed in this research goes beyond existing method-ologies in analyzing national AI policies in two important ways.Firstly,we haveanalyzed the universe of existing national AI policies.Secondly,we are utiliz-ing an ensemble LDA methodology(eLDA)which builds on the exi

130、sting and17pre-trained LDA model.Latent Dirichlet Allocation or LDA is an unsuper-vised NLP algorithm,which we can run through on a set of documents,andthe algorithm outputs the main topics which exist in the document set or inspecific documents within the set.As a conceptual aid,one can think of a

131、topicas being the“main idea”or“key theme.”LDA identifies the main topics in aset of documents by analyzing patterns among the words within the documentsspecified.This is done through a process known as Gibbs sampling,assign-ing probabilities to both words and topics until a converge is established;t

132、healgorithm effectively“tells”us which words are associated with which topics,and which topics are associated with which documents,each weighted accordingto its“prominence.”As previously mentioned,LDA is an unsupervised model,meaning that we as researchers do not tell the algorithm what topics we wa

133、ntit to search for;the model determines this through analyzing the probabilitiesof words in specific contexts throughout each document and ultimately the setof documents being utilized.It is important to note that LDA also does notsimply consider the frequency of words within a document but consider

134、s thecontext of the word through analyzing the words used around each word andthus determining specific meanings of words.Ultimately the LDA algorithmis applied to our set of documents,and after several iterations it determines aset of topics which are likely to have generated the final collection o

135、f words ineach policy document.Each document in our dataset is modeled by a mixtureof topics produced by the LDA algorithm.Subject matter experts can thenanalyze the set of words which inform a topic to determine what a specific topicmeans.See Figure 2 for an example.Figure 2:A sample word cloud of

136、a topic based on education and talent.Thefont size of each word follows the probability assigned by the algorithm to thatword in this particular topic(as determined by analysis of the documents in thecorpus).Our methodology goes even a step further,as we are utilizing ensembleLDA(eLDA).With traditio

137、nal LDA,there may be differences in the resultingtopics each time one runs the algorithm.This happens because the way in18which the model calculates the probabilities which determine the topics.eLDAstabilizes LDA results by building instead an ensemble of models and then dis-tilling the topics in co

138、mmon among the various ensembles,therefore stabilizingthe methodology.The shared,stable topics are then utilized to offer a stablemeta-model.The ensemble approach eliminates the impact of model sensitiv-ity on topics,thus improving the methodologys accuracy and reliability.In ourmethodology,we build

139、 several ensembles and compare the resulting meta-modelsby analysis of the coherence of the topics;we utilize various state-of-the-art topiccoherence metrics to essentially measure measure how interpretable topics areto humans.The meta-model with the highest topic coherence is utilized to ob-tain re

140、sults for interpretation.Ultimately,the output of this methodology is aset of topics which represent our universe of national AI policies.These topicsmake the narratives,goals,and mechanisms(and even motivations)a countryis utilizing to shape the development of their AI infrastructures more explicit

141、.Our methodology emphasizes a narrative based approach to understandingAI infrastructures in comparative contexts.This approach gives a far more com-prehensive and detailed analysis of the“AI wardrobe”and the different“outfits”which countries have selected from it.Existing comparative studies of nat

142、ionalAI strategies have tried captured the broad narratives which exist in these strate-gies through broad interpretations or focus on a few key words.We build onthese works in a meaningful way by bringing an empirically based approach,one that accounts for far more granularity,to analyzing these do

143、cuments.5.1Our Dataset:Introducing the National AI PoliciesOur dataset is made up of 54 countries who have published national AI poli-cies.These countries span different regions and are diverse in levels of na-tional income,technological development,and type of political system.Wehave identified 54

144、national strategy documents for each of these countries and167 intra-national documents.Therefore,a total of 221 text documents havebeen included in our eLDA model which inform the resulting topics output fromthe model.These documents thus cover several different wardrobe aspects,orillustratively we

145、 could say,these documents are several different blouses,pants,skirts,and jackets in a variety of colors and patterns.Practically,given our methodology our team needed to collect each of theexisting national AI strategies and as many intra-national policies available.Fortunately,there are existing r

146、esources which have compiled many of thesedocuments.The most complete resource is OECDs AI Policy Observatory.This database features a“country dashboard”for 70 countries.Each countrydashboard is a page which hosts all of the national and intra-national AI policies,as well as other policies relevant

147、to AI,for a specific country that are known byOECD.In some cases,countries may have intra-national documents but not anational level policy.Our team started the document collection process by using the OECD database.Throughout this process though we identified some problems with the database,making

148、it impossible to simply download all of the resources and input them19into our eLDA model.To arrive at our final dataset our team visited each of thecountry dashboards and carefully investigated each document hosted on eachdashboard to determine if the resource was(1)an official government policy do

149、c-ument,(2)relevant for AI policy,and(3)not a stub.Some documents were pub-lished in non-English languages,and we utilized Google translate and other NLPtools for the purpose of translating these documents to run them through oureLDA model.Our team also found documents outside of the OECD database.T

150、hese countries include China,Qatar,and Russia.In some cases,our teamalso located documents not hosted on the OECD website through web searches.These searches were necessary due to many resources on the OECD databasebeing reported as web-pages with dead URLs.We learned through our datacollection proc

151、ess that it is fairly common for AI documents that are hostedon web-pages to be removed or hosted elsewhere as government administrationschange,websites are overhauled,or policies are changed/updated.For our analysis we also needed to designate which policies were the nationalplans and which were in

152、tra-national plans.We have built our methodology tohave a few different levels of results.One level of analysis is just national plans,another analyzes both national plans and intra-national plans.We designatednational plans utilizing our own system rather than relying on the existing classi-ficatio

153、ns from the OECD website.Our system for categorizing these documentsrelies on a governments designation of their national plan.In some cases,thereare multiple documents designated as a countrys national plan;this is becausethe national government identifies multiple documents as such.6The Empirical

154、FindingsWe analyzed a total of 54 national plans and 167 intra-national plans,makinga total of 221 documents in our final model.Our national plans reveal broadnational policy priorities,while intra-national plans show levels of policy depth.We also performed document specific analyses which show sha

155、red policy issues(as identified by topics)across countries.We have organized our results be-low,emphasizing the topics resulting from the eLDA model,and which countrydocuments are best summarized by which topics.For ease of understanding wehave provided word clouds,which illustrate the key words tha

156、t make up eachtopic thus allowing us to summarize the meaning captured in each topic.6.1Comparing National AI Infrastructures:Policy Pri-orities RevealedFrom our 54 national plan documents the following results were produced.It isimportant to note that in some cases countries had two or three offici

157、al docu-ments that made up their“national plan”due to updated versions of the policiesbeing published.This is the case for the national plans for the United States,Canada,China,Estonia,Finland,France,Germany,India,and Japan.Thenumber of topics selected for our final eLDA model was 15,meaning that th

158、e20model sorted the dominating themes into 15 different topics,and then calculatedthe probability of each national plan featuring that topic in its contents.Thechoice of 15 fulfills our methodological criteria for model stability(consistency)and topic stability(coherence).The eLDA model produces a h

159、elpful heat-map(Figure 3)containing theseresults.This heat-map features topics on the horizontal axis(labeled 1-8)andcountries(as represented by their national AI plans)on the vertical axis.Theshading in the heat-map indicates topics which have a higher probability of pres-ence in a countrys nationa

160、l AI plan and are shaded darker.You will also noticeon this heat-map that there are dendograms on the left and top.The dendo-gram on the top of the heat-map shows how topics relate to one another whilethe dendogram on the left-hand side shows how different countries relate to oneanother.These relati

161、onships are determined by the Hellinger distance(Nikulin,2001)and agglomerative hierarchical clustering.Figure 3:National AI Infrastructures:Heat-map relating country-topicdistributions.21Figure 4:National AI Infrastructures Topic Word CloudsTopic 1 does not hold much significance in the corpus and

162、is omitted.At this level,we can report several major findings across clusters:There are clusters and there are distinct leaders that do not cluster with manystates.The major clusters are:EU,Latin America,historic British influence.The major leaders are China,Germany,Japan,and the United States.Topic

163、s feature divergences but there are many convergences.At a macrolevel,we see correlation matrices among topics.That of Germany is relatedto the EU,for example.At a micro level,we see fine distinctions.Germanyemphasizes standards,while the United States emphasizes benchmarks,butthey are both about te

164、chnical thresholds.Country plans can contain dominant but multiple topics.The U.S.plan isunique with the dominance of topic 8.However,the Chinese plan has domi-nant topic 8 but also topic 3,which is dominant for Germany and Switzerland.We are able to identify seven distinct clusters in this heat-map

165、.These clus-ters have largely broken down in by geographic organizations which make sensedue to overarching geopolitical associations and historical relationships.Table 1 presents the different clusters,their associated topic produced bythe eLDA model,and the member countries.Note that some countrie

166、s are only22partially included in a cluster,this is because they may have high probabilitiesfor more than one topic,which is possible when using the eLDA methodologyand more accurately captures the reality that some plans may have been in-fluenced by more than one pre-existing plan or competing inte

167、rests/prioritiesdomestically.Cluster NameTopic#&Top 10 WordsAssociated CountriesGreece Cluster2Electronic,mediumterm,procedure,register,digitization,provision,up-grade,intervention,utiliza-tion,tourismGreece,CyprusGerman-Swiss Clus-ter3 Federal,centre,mobil-ity,programme,digitaliza-tion,dialogue,Eur

168、ope,in-stance,shape,SMEGermany,SwitzerlandEconomic growth&development4 Personnel,agriculture,phase,figure,entrepreneur,analyze,master,food,agri-cultural,robotThailand,SouthKorea,Ukraine,MauritiusCommonwealth-British Influence5 discussion paper,de-ployment,startup,consider-ation,centre,intervention,d

169、ataset,paper,solve,figureIndia,Singapore,Malta,UnitedKingdom,Ireland,Uganda,United Arab Emi-rates,Australia,Qatar,Canada,(Partially includesSaudi Arabia)EU Cluster6 programme,economic af-fairs,publication,organiza-tion,competence,digitaliza-tion,centre,actor,final re-port,utilizationEuropean Union(P

170、artiallyincludes Spain)LatinAmerica-Spain influence7propose,productive,axis,seek,actor,OECD,re-lation,guarantee,agenda,analyzeSpain,Argentina,Uruguay,Colombia,Chile,Mexico,Peru,Brazil,(Partially in-cludes Russia)Science&Technol-ogy First Movers8federal,workforce,dataset,domain,hardware,benchmark,eng

171、ineering,robot,cybersecurity,tech-niqueUnited States,China,JapanTable 1:Clusters,Topics,Top 10 Words,and CountriesNote:Cluster 1 omitted because no major countries were included in itThe following analysis of topics is not chronological:leads with first movers,pace-setters(Germany),then EU,UK/Common

172、wealth,then Latin America/Spain.23Topic 8Top 10 words:federal,workforce,dataset,domain,hardware,bench-mark,engineering,robot,cybersecurity,techniqueTopic 8 contains themes that have a high probability to be found in thenational plans of the United States,China,and Japan.All of these countriesare lea

173、ders already in technology and innovation therefore it makes sensethat they might have well articulated and formulated plans towards develop-ing and deploying AI domestically.This topic emphasizes datasets,hardware,the workforce and benchmarks(rather than standards)and basic science ca-pabilities.Th

174、us indicating some level of understanding of the needs for the AIindustry,and a desire to be a part in providing the proper environment for theseelements of AI infrastructures to be developed and utilized.The emphasis alsoon basic science and engineering is important to note as these lend themselves

175、to strategies that prioritize research and the cutting edge.Additionally,we seethe use of the term“cybersecurity”which has yet to show up in other topics indicating that these countries have concerns about elements of security.Thisis in line with the existing geopolitical conflicts between the U.S.a

176、nd China.Another primary feature of this topic are clear statements about aspirations tobe global leaders in the AI industry.Topic 3Top 10 words:Federal,centre,mobility,programme,digitalization,dialogue,Europe,instance,shape,SMETopic 3 is closely related with the strategies of Germany and Switzerlan

177、d.These plans emphasize the importance of federal action in the AI space andrecognize the importance of dialogue and consultation.Switzerlands strategyis a broader plan,“Digital Switzerland Strategy”which puts an emphasis on AIand its importance for the country to be a leader in the technology.Meanw

178、hileGermanys AI strategy is solely focused on the involvement in the developmentand applications of AI.Both the German and Swiss plans heavily discuss theimportance for dialogue and coordination within the European region and otherinternational organizations.Topic 6Top 10 words:programme,economic af

179、fairs,publication,organiza-tion,competence,digitalization,centre,actor,final report,utilizationTopic 6 is closely related to many of the countries within the European Union.The top words that define this cluster are closely related to specific EU programsand issue spaces of specific interest to the

180、countries in the union.These wordssuggest a large amount of strategic action proposed in these national plans.Itis likely that these plans suggest to development of new programs and organiza-tions,or reorientation of existing ones,to support the development of AI.Thereis also an element here of asse

181、ssment,in the references to publications and finalreports.This reflects upon the processes which typically occur within the EUto coordinate member actions in specific areas.Specific countries included hereare:Denmark,Hungary,and Estonia among others.24Topic 5Top 10 words:discussion paper,deployment,

182、startup,consideration,centre,intervention,dataset,paper,solve,figureThis topic,like topic 4,is focused heavily on economic growth and development.Yet there is a difference in approach.This topic is more focused on two things,first strategizing which we see through words such as“discussion paper,”“de

183、-ployment,”“consideration,”“centre,”“intervention,”and“paper.”These planshave a high probability of emphasizing different tasks which need to be done bya variety of actors to support the development of AI.It appears there is a wholeof government type approach.Aside from that,there is a focus on star

184、tups anddatasets,which are important for developing a robust domestic AI industry.Many of the countries which have a high probability of their plans containingthis topic have ties to British colonialism or influences during history.Thesestates are full of ambition and seek to manage the development

185、of AI througha variety of different policy levers.Topic 7Top 10 words:propose,productive,axis,seek,actor,OECD,rela-tion,guarantee,agenda,analyzeTopic 7 is has a high probability of appearing in the national strategies of manyLatin American countries(including Argentina,Uruguay,Colombia,Chile andMexi

186、co)but also countries such as Russia,Turkey,and Spain.Some of thesecountries clustering together may inherently make sense due to historical andcultural legacies.Meanwhile others are more surprising.The topic seems to em-phasize broad processes and objectives(the term axis is a machine translationof

187、 Spanish word ejes or objective).Topic 7 speaks to country plans which areformulating different government tasks to undertake to develop AI(“propose,”“actor,”“guarantee,”and“agenda”).These plans also seem to be interested inproductivity and working to catalyze productivity with AI and effective plan

188、s toharness the technology for economic benefits.Finally,the reference to OECDindicates that these countries might either be interested in collaborations,ordesire to learn from other leaders in the OECD(which has done work on ana-lyzing AI policies around the globe as referenced earlier in this pape

189、r).Topic 4Top 10 words:Personnel,agriculture,phase,figure,entrepreneur,analyze,master,food,agricultural,robotTopic 4 is characterized as a topic which encapsulates plans focused primar-ily on economic growth and development.There is a heavy emphasis on theagriculture industry in this topic,but also

190、on entrepreneurship and personnel.The top ten words in this topic are indicative of strategies which have a stronglikelihood of utilizing AI for the purpose of catalyzing economic development.These countries include Thailand,Ukraine,South Korea,and Mauritius.Eachof these plans makes multiple mention

191、s to the importance of applying AI to the25agriculture sector,with three of the four plans having designated sections tothe topic.Additionally,all emphasize the importance of entrepreneurship andinnovation as key elements for developing the technology and using AI to sparkfurther company and device

192、creation.Ultimately these plans appear to thata holistic approach,not limited to each countrys strength but touching on avariety of important industries and aspects to develop and deploy AI into foreconomic development.Topic 2Top 10 words:Electronic,medium term,procedure,register,digiti-zation,provi

193、sion,upgrade,intervention,utilization,tourismIn this level of analysis topic two was most closely associated with the nationalAI strategies of Greece and to a slightly smaller probability,Cyprus.Thismeans that there is a high probability that the word distributions associatedwith topic two will appe

194、ar in Greece and Cyprus national AI strategies.Theseplans appear to be highly associated with applying AI and other technologiestowards medium term goals,likely with particular interest in specific industriesof importance,such as tourism.There appears to be recognition of a need toupgrade and digiti

195、ze to see forward additional AI aspirations.6.2Comparing Intra-national AI Strategies:Policy DepthFor the analysis of our intra-national policy documents we included both the54 national AI policy documents and the 167 intra-national documents.Inother words,the results developed from running the eLDA

196、 model on our totaldataset of 221 text policy documents.As mentioned earlier,intra-national pol-icy documents are documents relevant to AI from various government agencies,departments,commissions,or institutions.We have provided a breakdown ofthe top ten countries which have published intra-national

197、 AI documents in Ta-ble 2.Examples of these documents include autonomous vehicle policies,AIeducation policies,and policies for data protections as they relate to AI.Country/RegionNumber of DocumentsEuropean Union28United Kingdom17China16Japan16United States15Colombia15India6Germany5France5Table 2:T

198、op 10 Countries for Number of AI Documents26These plans have important implications for how AI is developed,used invarious industries,but also for how a variety of spheres of society are preparedfor an AI enabled future.The analysis produced from this model shows thedifferent policy themes which exi

199、st amongst the featured countries and theirwhole of government approaches to AI policy.In Figure 5 we have provided a heat-map which shows what countrys com-bined national plans and intra-national plans with topics created from the eLDAmodel.This diagram shows the different countries included in the

200、 analysis onthe vertical axis and the different topics on the horizontal axis.The darker bluethe spaces are colored,the higher the probability is that that specific topic iswithin the set of documents associated with that country.In Figure 6 we havea breakdown of the contents of each of the topics g

201、enerated from our eLDAmodel.Figure 5:Intra-National AI Infrastructures:Heat-map relating country-topicdistributions.27Figure 6:Intra-National AI Infrastructures Topic Word CloudsHere are a few broad themes at the intra-national level:The eLDA methodology resulted in eight different topics.Before thi

202、s analysiswas completed,the expectation was that an LDA with all international docu-ments included would result in more focused,sharper topics.The results were,in fact,different with a large amount of coherence around three dominanttopics(topics 7,4,and 5).These topics capture the necessary fundamen

203、talsa country needs to develop an AI infrastructure.The countries which areassociated with topics are different from the four countries and one regionwe have identified as global leaders in AI research and/or ethics:Germany,China,the European Union,Japan,and the United States.Their associated topics

204、 are listed in Table 3.28Country/RegionTopic#&Top 10 Key WordsJapan1 operator,trained model,vendor,uti-lization,paragraph,handle,attack,item,assume,plantUnited States2federal,explanation,enforcement,patent,comment,FDA,NIST,algorith-mic,disability,accuracyEuropean Union3 liability,member state,damage

205、 cause,fault,parliament,compensation,stress,GDPR,obligation,trustworthyChina6 trustworthy,attack,figure,scenario,theory,municipal,municipality,chip,col-lege,reformGermany8 federal,algorithmic,ethics,obligation,GDPR,harm,consent,believe,operator,discriminationTable 3:Topics for Leading AI Countries a

206、nd EU Figure7 illustrates different topics and their probabilities to have relatedword distributions appear in national plans of the U.S.,Russia,Germany,EU,China,Canada,and India.We can conclude from this figure that somecountries contain multiple topics while others contain only one.We arguethat th

207、is is likely because some countries have elements of AI policies whichare originally derived while others borrow elements from other pre-existingpolicies,thus illustrating some level of policy diffusion.Another reason forthis is AI policies which are in their earlier stages of development in termsof

208、 articulating their goals and mechanisms for developing and utilizing AIsystems.Figure 8 illustrates the number of countries which have some levelsof all the topics in their word distributions,and this is modeled by numbersof countries with each topic.Of these topics,7&4 are largely associatedwith f

209、undamentals of any AI policy,and 5&8 are associated with basics ofdata governance and regulatory issues(see topic breakdown in Figure 6).29Figure 7:Intra-national Policy Documents:Topic Probabilities Per CountryFigure 8:Country Counts Per Topic30Topic 1Top 10 words:operator,trained model,vendor,util

210、ization,para-graph,handle,attack,item,assume,plantTopic 1 is closely associated with documents from Japan.Many words thatare related to this topic are highly technical including“trained model”and“op-erator,”indicating that some intra-national documents by the Japanese govern-ment are likely sufficie

211、ntly detailed.Additionally,we see words such as“attack”indicating concerns about security,which relates to Japans national plan be-ing associated with the topic containing a reference to“cybersecurity.”TheJapanese government has published several intra-national documents including“Act of Protection

212、of Personal Information,”“AI Utilization Guidelines,”“Ma-chine Learning Quality Guidelines,”and“Social Principles of Human-CentricAI.”These documents are highly specific,and some quite technical.This showsa depth to the governments approach to AI policy.There are multiple sets ofguidelines published

213、,which also lend themselves to regulatory perspectives anda high degree of government guidance to development of technology.Topic 2Top 10 words:federal,explanation,enforcement,patent,comment,FDA,NIST,algorithmic,disability,accuracyThis topic is closely associated with the United States.In fact,the U

214、.S.features only this topic it its topic composition,therefore it is incredibly im-portant for understanding the U.S.national and intra-national documents.Wecan note that some of the terms in this topic are specifically related to the U.S.since they are agencies in the American Government FDA and NI

215、ST.Otherwords include“federal”which indicates that the federal government does havea role and plans related to AI.Additionally,there are several words related togovernance “enforcement”and“patent”specifically.There is also an elementof AI policy being open for shaping by the public through the word“

216、comment.”Some intra-national documents published by the U.S.include documents fromthe U.S.Patent and Trade Office,the“NIST Principles for Explainable AI,”the“Plan for Federal Engagement in Developing Technical Standards and RelatedTools,”and the”Artificial Intelligence and Algorithmic Fairness Initi

217、ative.”Topic 3Top 10 words:liability,member state,damage cause,fault,parlia-ment,compensation,stress,GDPR,obligation,trustworthyTopic 3 is closely associated with the European Union,but not with spe-cific EU countries.This is likely due to the nature of EU approaches versusthe approaches of member s

218、tates.The EU has published several intra-national(intra-regional in this case)documents.One of these includes the document ref-erenced as“GDPR”which is the EUs data protection law.Documents from theEU seem to emphasize trustworthy and fair AI systems.This is evident throughwords such as“trustworthy,

219、”“obligation,”“liability,”and“compensation.”Italso appears that there might be some allusion to different concerns related tothe roll out of AI.Some plans published by the EU include“Ethics Guidelines31for Trustworthy AI,”“Open Data Directive,”the“Framework of ethical aspectsof artificial intelligen

220、ce,robotics,and related technologies,”and the“SectoralConsiderations on the Policy and Investment Recommendations for Trustwor-thy Artificial Intelligence.”These different documents show the massive rangein publication topics but also the emphasis on ethics,dialogue,and trust.Topic 4Top 10 words:Org

221、anization,bias,regulator,footnote,algorithmicdecision make,evidence,harm,algorithmic,outcome,barrierTopic four contains several countries.This topic captures many elementsthat are critical for the regulation of AI.Terms that indicate efforts or planstowards regulating AI include“regulator”and“eviden

222、ce.”This topic also em-phasizes some new elements such as“harm”and concern over“bias”therefore,there are likely elements of concern about AI in these documents,but also likelyplans for regulation to mitigate the impacts of these potentially negative im-pacts of AI.It is important to note that most o

223、f the countries which are alsoin topics 7 and 5 also have overlap with this topic.This is due to the fact thattopics 7 and 5 are also topics which relate to basic aspects of AI infrastructures.Topic 5Top 10 words:subsection,organization,registrar,figure,mission,imperative,cent,paragraph,outcome,clas

224、sroomTopic 5 is another topic which contains a variety of countries and overlapswith topics 4 and 7.This topic primarily emphasizes economic organization,featuring words such as mission,imperative,outcome,firm,manufacture,andGDP.This topic being present indicates plans that may focus on the uses of

225、AIfor economic development and growth.This is important in the view of manycountries.There are also aspects of this topic which allude to education suchas“classroom”which connects the importance of education to the developmentof AI but also the consideration of the implications of AI for education.T

226、wocountries which heavily feature topic 5 include Australia and Canada.Topic 6Top 10 words:trustworthy,attack,figure,scenario,theory,munici-pal,municipality,chip,college,reformTopic 6 is heavily featured in the intra-national sets of AI strategies pub-lished by China and South Korea.This topic empha

227、sizes a variety of distinctthemes in AI strategies.There is an emphasis on security but also on trustwor-thiness thus focusing on both outside threats to AI systems and needs of thosedomestically to trust AI systems sufficiently.Another key term is“chip”whichis fitting given both China and South Kor

228、eas involvement in semiconductormanufacturing,which both countries likely hope to grow and improve in future.There are also references to universities emphasizing likely both education andresearch,which are key elements of strategies for workforce development.32Topic 7Top 10 words:actor,table,exploi

229、tation,analyze,variable,sandbox,institutional,digitization,productive,decreeTopic 7 should be considered the most basic organizer of AI infrastructures.This term is one we are using to describe the basic elements needed for AI infras-tructures to be created and support the development of the technol

230、ogy.Of the54 countries included in this analysis,42 of them feature some elements of thistopic.Important themes within this topic are:actors,institutions,execution,economics(commerce,blockchain,productive),processes(dialogue,sandbox),and capabilities(digitization).These terms are all important piece

231、s of designingthe right environment for innovation in AI,but also for continuing to evaluateplans and effectiveness as well.Countries which have a large portion of this topicinclude most of the Latin American states,many of the European Union states(but not Germany,Ireland,or the EU plan),and develo

232、ping countries.Thistopic overall captures the 3Cs of AI infrastructures:competencies,concerns,and capabilities.Topic 8Top 10 words:federal,algorithmic,ethics,obligation,GDPR,harm,consent,believe,operator,discriminationTopic 8 is associated closely with Germany and shows a high degree of eth-ical and

233、 societal concerns.This is seen through the prevalence of words suchas“ethics,”“GDPR,”“harm,”“consent,”and“discrimination.”It is also in-teresting to see the word“believe”which may indicate an element of normsbeing important in these intra-national plans,and these norms may also relateto ethics and

234、fairness in the AI infrastructure space.Some documents publishedby the German government that were included in this analysis include the“In-terim Report One Year Strategy,”the“Opinion of the Data Commission,”and“Ethical Guidelines for Self-Driving Cars.”The emphasis on data protectionand safety in s

235、elf-driving cars aligns with the themes of this topic as well.6.3Document Level Analysis:Analyzing the 5 Most Im-portant Topics in Our Intra-national DocumentsTo round out our analysis of our dataset of AI policy documents we ran oureLDA model on our set of 167 intra-national policy documents.We opt

236、ed torun this analysis at the document level rather than the country level.We didthis in order to establish the most popular topics throughout these documents.Once we obtained these top topics,we were able to then connect the documentsassociated with each topic to their publishing country.For this a

237、nalysis doc-uments from 37 countries were included,as some of the countries which havenational plans do not have intra-national plans.Additionally,we have addedKenya to this analysis,a country which has intra-national documents but not anational level AI plan.We have broken down and classified the f

238、ive top topics inthese documents and presented this information in Table 4.Here is our macroanalysis of the intra-national documents:33Topic NumberTopic Name Descriptor1Data&Governance2Education&Training3Economy4Contracts&Liability5TransportTable 4:Main Topics Across Documents Our results show that

239、27 of the 38 included countries have published docu-ments that feature elements of all five of the top topics that eLDA identified.Meanwhile,there are 11 countries that do not address all 5 of these topics.These countries are:Egypt,Estonia,Greece,Iceland,Kenya,Malta,Norway,Singapore,Thailand,Tunisia

240、,and Vietnam.Our eLDA model determined that topics 2 and 3 are the most similar,showinga connection between education,training,and the economy.This is likely dueto the manner in which these topics are discussed as education and trainingare seen as important for developing an AI-ready workforce.Topic

241、s 1 and 4also have overlap,primarily due to the expectation that data and governanceissues are related to contracts and property.Now we present 5 word-clouds which show the universe of documents in ourintra-national dataset(figures 9,10,11,12 and 13).In these word-clouds the font sizes indicate the

242、documents which have ahigher probability of possessing word distributions that constitute this topic.This means that these documents have a higher probability of containing thisspecific topic as a major theme.Topic 1:Data and governanceTop 10 words:component,dataset,algo-rithm,attack,external,sectio

243、n,internal,input,check,stageThe word-cloud presented for topic 1 has a large amount of diversity for countriespublishing documents associated with data and governance.This is aligned withthe results we found from our intra-national policy depth analysis findings.Dataand governance are important basi

244、c foundations for AI infrastructures.Somedocuments which have high probabilities of containing word distributions as-sociated with this topic include:Japans Machine Learning Quality Guideline,Chinas Security Specification and Assessment Methods for ML Algorithms,EUs Independent High Level Expert Gro

245、up on AI,the United States NISTPrinciples for Explainable AI,and the EUs Robustness and Explainability ofAI JRC Technical Report.Words associated with this topic include the follow-ing:algorithm,component,dataset,indicator,attack,check,fairness,reliability,and accuracy.34Figure 9:Top Panel:Datasets

246、and Governance Topic Word Cloud(topic 1 outof 5 topics learned by the model over intra-national documents).Bottom Panel:Datasets and Governance Document Cloud relating the probabilities with whichthis topic emerges over countries(national and intra-national documents),uti-lizing larger font to visib

247、ly relate a document where the topic is most prominent(probabilistically).35Figure 10:Top Panel:Education and Training Topic Word Cloud.Bottom Panel:Education and Training Document Cloud.36Figure 11:Top Panel:Economy Topic Word Cloud.Bottom Panel:Economy Document Cloud.37Figure 12:Top Panel:Contract

248、s Topic Word Cloud.Bottom Panel:Contracts Document Cloud.38Figure 13:Top Panel:Transport Topic Word Cloud.Bottom Panel:Transport Document Cloud.39Topic 2:Education and trainingTop 10 words:talent,university,enter-prise,strengthen,major,algorithm,construction,basic,smart,schoolIn this word-cloud it i

249、s immediately apparent that documents from China havedistributions of words that are associated with education and training.Keydocuments which are heavily associated with this topic are:Chinas AI Tal-ent Training Report,Chinas AI Industry Talent Development Report(2019-2020),Chinas AI Innovation Act

250、ion Plan for Institutions of Higher Education,Chinas Establishment of Ministry of Education AI Technology Innovation Ex-pert Group,and Chinas Guidelines for Construction of National New Genera-tion AI Standards System.In the associated topic word-cloud we can immedi-ately notice words that tell us t

251、his topic is speaking to education and training talent(specifically pertinent for China due to the 1000 Talents Program),university,enterprise,strengthen,school,and industrial.In iterations our mod-els prior to eLDA analysis,the Chinese approach to education encapsulated inthe key word“talent”stood

252、out as distinct from the other countries educationstrategies.We have reproduced word clouds here from those earlier iterations.Figure 14:Education versus Talent Approaches40Topic 3:EconomyTop 10 words:economy,infrastructure,job,growth,in-dustrial,ministry,university,centre,fund,programmeIn this topi

253、c there is a large range of countries publishing documents.This isnoticeable due to the relatively similar size of many of the documents presentedin the word-cloud.Some documents which have high probabilities of containingword distributions associated with the economy include:the United KingdomsCoun

254、cil AI Road-map,the United Kingdoms Industrial Strategy Building ABritain Fit for Future,Kenyas Digital Economy Blueprint,Australias Pros-perity Through Innovation,Singapores A Guide to Job Redesign in Age of AI,and Slovakias Action Plan for Digital Transformation 2019 to 2022.The word-cloud associa

255、ted with this topic thematically includes words such as:economy,growth,infrastructure,productivity,industrial,and job.Topic 4:Contracts and liabilityTop 10 words:contract,party,article,provider,vendor,operator,section,paragraph,agreement,utilizationThis topic is relatively diverse in terms of countr

256、ies which are publishing docu-ments containing word distributions associated with contracts and liability.Itshould be noted though that three of the largest documents associated withthis topic are from Japan.They are Japans Protection of Personal InformationAct,Japans Contact Guidelines on Utilizati

257、on AI and Data Version 1,andJapans Practical Guidebook on Providing Data for Employee Development inAI and Data Science.Other countries with documents with high probabilitiesof word distributions associated with this topic include:Icelands Act on DataProtection and Processing of Personal Data,Austra

258、lias Intellectual PropertyLaws Amendment,and Colombias Law for Promotion of AI Technology andEntrepreneurship.Some of the words associated with this topic are shown in thecorresponding word-cloud.Words that are particularly important are:party,article,contract,vendor,operator,provider,obligation,usa

259、ge,and agreement.Topic 5:TransportTop 10 words:liability,damage,article,damage cause,vehicle,fault,compensation,thing,liable,strict liabilityAt first glance,one will notice that there are several European Union docu-ments included in topic 5.This means that several EU documents contain worddistribut

260、ions that have a high probability of being related to the topic of trans-port.Some of the European Union documents include:EUs Civil LiabilityRegime for AI,EUs Report on Safety and Liability Implications of AI,Internetof Things and Robotics,and EUs Resolution on AI Questions of Interpretationand App

261、lication of International Law.Some other countries publishing docu-ments related to this topic include Australia with their National EnforcementGuidelines for Automated Vehicles,Lithuania with their Law on Self DrivingCars,and Austria with their Code of Practice Testing of Automated Driving onPublic

262、 Roads.Words associated with this topic include:liability,vehicle,fault,article,liable,damage,driver,damage cause,and compensation.417ConclusionThis policy report has provided a detailed analysis of 213 artificial intelligencepolicies,ranging from national AI plans to intra-national plans.Given thee

263、xpanse of this analysis much can be said about the universe of AI policiesand documents and thus how countries are developing AI and implementingAI technologies into a variety of government and societal systems.This reportillustrates that there are convergences and divergences in the values and goal

264、sfor national AI infrastructures around the world.The research illustrates thatthere is a wardrobe(or set)of policy mechanisms which are commonly used inAI infrastructures and that countries choose their own combinations of thesemechanisms to best fit their needs.This is why we see distinct clusters

265、 ofcountries when it comes to the values embodied in their AI policies.To wrapup this report we provide three major takeaways from this research,which willbe expanded upon in the coming years.Three major takeaways:There are common elements among all national AI infrastructures,and over-laps can be s

266、een across regional boundaries.These common elements illustratethat there are different levels of AI infrastructure development and aspectsof diffusion where countries may have learned from the experiences of othercountries who have developed AI infrastructures before them.Despite this overlap,there

267、 are key differences among national AI plans there are those which heavily emphasize matters of privacy,transparency,and accountability while others many not articulate the importance forthese protections and considerations.Some of these differences are due to thestage at which a country is in devel

268、oping their AI infrastructure,while othersmay relate to political systems or aspirations.Despite this though,there isa need to focus on policy depths and regulatory or governance capabilitiesrather than unnecessary fear mongering about the state of AI around theworld.Finally,while studying these nat

269、ional AI infrastructures it is helpful to utilizea comparative approach which allows for a deeper understanding of the uni-verse in which AI technologies are being developed and deployed.Differentapproaches have influences on future iterations of AI policies and the use ofAI in service provision.The

270、refore,utilizing methodologies which thoroughlyanalyze the universe of AI infrastructures and the values,goals,and imple-mentation of them provide a more complete and realistic view of the state ofAI as it is being innovated and harnessed in a variety of country contexts.428Team BiographiesJ.P.Singh

271、(Principal Investigator)J.P.is a Distinguished University Professor andCo-Director of the Center for Advancing Human-Machine Partnership at George Mason University.He is a Loomis Council member with the StimsonCenter,and a Richard von Weizs acker Fellow withthe Robert Bosch Academy,Berlin.He works a

272、tthe intersection of technology,culture and polit-ical economy in global contexts examining trans-formative impacts from provision of telephone ser-vice in poor countries,to the use of AI in globalvalue chains in cutting-edge industries.J.P.hasconsulted or advised international organizations such as

273、 the British Council,UNESCO,the World Bank,and the World Trade Organization,and conductedfield research in 36 countries.His current book project explores AI and innova-tion in Germany,India,and the United States.A winner of numerous researchawards and fellowships,he has written 10 books and over 100

274、 scholarly articles.Amarda Shehu(Co-Principal Investigator)Amarda is a Professor of Computer Science,Associate Dean for AI Innovation,and AssociateVice President of Research for Institute of Digi-tal InnovAtion.During 2019-2022,Amarda co-directed the Trans-disciplinary Center for Ad-vancing Human-Ma

275、chine Partnerships and servedas an NSF Program Director in the CISE Di-rectorate.Amarda considers herself an interdis-ciplinary scientist.Her research record includesfoundational advances in AI,Machine Learning,and Algorithmics,and purposeful research thatpushes the barriers of our understanding of

276、physical world.Amarda has a gal-vanizing view of computing and relentless energy and advocacy for advancementof knowledge across scientific disciplines and improvement of human condition.Jesse L KirkpatrickJesse is a Research Associate Professor andthe Acting Director of the Institute for Philoso-ph

277、y and Public Policy at George Mason Univer-sity.He is also an International Security Fellowat New America and serves as a consultant fornumerous organizations.His research is interdis-ciplinary,cutting across such fields as Philosophy,Political Science,Public Policy,and the Life andComputer Sciences

278、.At its core,it aims to ex-plore two central,interrelated themes:(1)how asuite of technologies,singularly and in convergence,impact peace and security,and(2)what the ethical,social,and policy implications of these impacts may be.43Michael HunzekerMichael is an associate professor at GeorgeMason Univ

279、ersitys Schar School of Policy andGovernment,the associate director of the ScharSchools Center for Security Policy Studies,anda Senior Non-Resident Fellow at the Center forStrategic and Budgetary Assessments.A for-mer Marine Corps officer,he works on conven-tional deterrence,defense reform,and organ

280、iza-tional learning.Antonios AnastasopoulosAntonios is an assistant professor at GeorgeMason Universitys Computer Science Depart-ment.He is a part of the George Mason Nat-ural Language Processing Group and is inter-ested in various aspects of multilingual Natu-ral Language Processing and Machine Lea

281、rn-ing,with a main focus on Machine Trans-lation and Speech Recognition for endangeredlanguagesandlow-resourcesettingsingen-eral.Caroline Wesson(Project Manager)Caroline is a PhD candidate and Presiden-tial Scholar studying Political Science at GeorgeMason University.She is currently working onher d

282、issertation which focuses on science and in-novation policy,with a focus on innovation clus-ters in different global contexts.Caroline holdsa B.S.and M.S.in International Affairs from theGeorgia Institute of Technology and has workedin research positions at RAND,the Center forStrategic and Internati

283、onal Studies(CSIS),andthe World Bank.Manpriya DuaManpriya is a PhD student in the ComputerScience Department at George Mason University.Her research centers Machine Learning,focusingon Deep Learning architectures for Natural Lan-guage Processing(NLP).She holds a Masters inScience degree in Computer

284、Science from GeorgeMason University,and has worked with Palo AltoResearch Center on a DARPA project investigat-ing use of NLP methods for military defense strat-egy.44Vasilii NosovVasilii is a PhD in Public Policy studentat Schar School and a Graduate Research As-sistant at the Center for Advancing

285、Human-Machine Partnership.His research interestslie in the intersection of economics,interna-tional relations,and institutional and techno-logical transfer.Vasilii holds a Master of Sci-ence degree in Economics from Tufts Univer-sity.William“Webby”ApplegateWebby is an MA student in Philosophy atGeor

286、ge Mason University and a Graduate Re-search Assistant for the Institute for Philosophyand Public Policy.His research is mainly focusedaround two questions:how to make good deci-sions;and how to run a country well.Webbyholds a BA in Philosophy and a minor in Eco-nomics from Wittenberg University.In

287、additionto his research position,Webby has experience inmultiple startup companies and is currently work-ing on his thesis on philosophy and uncertainty.45ReferencesAaronson,Susan Ariel.2023.“Building trust in AI:a landscape analysis ofgovernment AI programs.”.Bareis,Jascha and Christian Katzenbach.

288、2022.“Talking AI into being:Thenarratives and imaginaries of national AI strategies and their performativepolitics.”Science,Technology,&Human Values 47(5):855881.Blei,David M.,Andrew Y.Ng and Michael I.Jordan.2003.“Latent DirichletAllocation.”J.Mach.Learn.Res.3(null):9931022.CIFAR.2017.“CIFARpan-Can

289、adianArtificialIn-telligenceStrategy2017-20.”http:/www.cifar.ca/ai/pan-canadian-artificial-intelligence-strategy.European Commission.2018.Artificial Intelligence for Europe.Communicationfrom the Commission to the European Parliament,the Council,the EuropeanEconomic and Social Committee and the Commi

290、ttee of the Regions,EuropeanCommission,COM(2018)237 final.Foffano,Francesca,Teresa Scantamburlo and Atia Cort es.2022.“Investing inAI for social good:an analysis of European national strategies.”AI&societypp.122.German Federal Government.December 2020.“Artificial Intelligence Strategyof the German F

291、ederal Government.”.Hine,Emmie and Luciano Floridi.2022.“Artificial intelligence with Americanvalues and Chinese characteristics:a comparative analysis of American andChinese governmental AI policies.”AI&Society”pp.122.Nikulin,M.S.2001.Hellinger distance.Encyclopedia of Mathematics EMSPress.Niti Aay

292、og.2018.Discussion Paper,National Strategy for Artificial Intelligence#aiforall.Discussion Paper,Niti Aayog,Government of India.Robinson,Stephen Cory.2020.“Trust,transparency,and openness:How in-clusion of cultural values shapes Nordic national public policy strategies forartificial intelligence(AI)

293、.”Technology in Society 63:101421.Select Committee on Artificial Intelligence.2016.The National Artificial In-telligence Research and Development Strategic Plan.National Science andTechnology Council,Executive Office of the President of the United States.”Select Committee on Artificial Intelligence”

294、.2023.The National ArtificialIntelligence Research and Development Strategic Plan.Select Committee onArtificial Intelligence,National Science and Technology Council,ExecutiveOffice of the President of the United States,2023 Update.46State Council.2017.State Council Notice on the Issuance of the Next

295、 Genera-tion Artificial Intelligence Development Plan(Trans.Rogier Creemers,Gra-ham Webster,Paul Tsai,Paul Triolo,and Elsa Kania).State Council,China.Wilson,Christopher.2022.“Public engagement and AI:A values analysis ofnational strategies.”Government Information Quarterly 39(1):101652.47A PRODUCT O

296、F THE AI STRATEGIES TEAM AT GEORGE MASON UNIVERSITY HTTPS:/WWW.AISTRATEGIES.GMU.EDU/THIS RESEARCH IS SUPPORTED BY A$1.389 MILLION GRANTFROM THE MINERVA RESEARCH INITIATIVE.PLEASE SENDQUERIES ON THIS REPORT TO AIPOLICYGMU.EDU.Suggested citation:Suggested citation:Singh,J.P.,Amarda Shehu,Caroline Wesson,and Manpriya Dua.The 2023 Global ArtificialIntelligence Infrastructures Report.With a Foreword from David Bray.AI Strategies Team and the Institute forDigital Innovation,George Mason University,and the Stimson Center,Washington DC.August 2023.