1、 BIS Working Papers No 1179 The impact of artificial intelligence on output and inflation by Iaki Aldasoro,Sebastian Doerr,Leonardo Gambacorta and Daniel Rees Monetary and Economic Department April 2024 JEL classification:E31,J24,O33,O40.Keywords:artificial intelligence,generative AI,inflation,outpu

2、t,productivity,monetary policy.BIS Working Papers are written by members of the Monetary and Economic Department of the Bank for International Settlements,and from time to time by other economists,and are published by the Bank.The papers are on subjects of topical interest and are technical in chara

3、cter.The views expressed in them are those of their authors and not necessarily the views of the BIS.This publication is available on the BIS website(www.bis.org).Bank for International Settlements 2024.All rights reserved.Brief excerpts may be reproduced or translated provided the source is stated.

4、ISSN 1020-0959(print)ISSN 1682-7678(online)The impact of artificial intelligence onoutput and inflationI AldasoroBISS DoerrBIS&CEPRL GambacortaBIS&CEPRD ReesBISApril 11,2024AbstractThis paper studies the effects of artificial intelligence(AI)on sectoral and aggregateemployment,output and inflation i

5、n both the short and long run.We construct anindex of industry exposure to AI to calibrate a macroeconomic multi-sector model.Building on studies that find significant increases in workers output from AI,wemodel AI as a permanent increase in productivity that differs by sector.We findthat AI signifi

6、cantly raises output,consumption and investment in the short andlong run.The inflation response depends crucially on households and firms an-ticipation of the impact of AI.If they do not anticipate higher future productivity,AI adoption is initially disinflationary.Over time,general equilibrium forc

7、es leadto moderate inflation through demand effects.In contrast,when households andfirms anticipate higher future productivity,inflation rises immediately.Inspectingindividual sectors and performing counterfactual exercises we find that a sectorsinitial exposure to AI has little correlation with its

8、 long-term increase in output.However,output grows by twice as much for the same increase in aggregate pro-ductivity when AI affects sectors producing consumption rather than investmentgoods,thanks to second round effects through sectoral linkages.We discuss howpublic policy should foster AI adoptio

9、n and implications for central banks.JEL Codes:E31,J24,O33,O40.Keywords:artificial intelligence,generative AI,inflation,output,productivity,monetarypolicy.We thank seminar participants at the BIS for helpful comments and suggestions.Contact:Aldasoro(inaki.aldasorobis.org),Doerr(sebastian.doerrbis.or

10、g),Gambacorta(leonardo.gambacortabis.org),and Rees(daniel.reesbis.org).The views expressed here are those of the authors only and not necessarilythose of the Bank for International Settlements.1IntroductionRecent advances in artificial intelligence(AI)have raised hopes of a boost to economicgrowth.M

11、any scholars believe that AI has the potential to be“the most importantgeneral-purpose technology of our era”(Brynjolfsson et al.,2023).The recent inroads ofgenerative AI in everyday applications in particular promise widespread efficiency gains.1Unlike automation through robots,which can accomplish

12、 only explicitly understood(i.e.,routine)tasks,AI can infer tacit relationships that are not fully specified by underlyingsoftware(Autor,2022).By transforming occupational tasks,altering corporate strate-gies,and affecting production efficiency,AI may have significant consequences for labourmarkets,

13、firms,and whole industries(Agrawal et al.,2019).A key channel through which AI affects economic growth is through improvements inproductivity(Acemoglu and Restrepo,2018;Aghion et al.,2018).Micro-economic stud-ies find that generative AI can make workers tremendously more productive,especially inoccu

14、pations that require cognitive work(Brynjolfsson et al.,2023;Noy and Zhang,2023).AI also boosts firm growth and innovation(Babina et al.,2024).At the macro-economiclevel,analyses suggest that AI could raise annual productivity growth by around 1 per-centage point(pp)per annum over the next decade(Ba

15、ily et al.,2023;Goldman Sachs,2023).The adoption of AI can hence be thought of as an increase in productivity thatexpands an economys output capacity.Compared to information technology(IT),whoseimpact took years to be reflected in aggregate productivity numbers(Fernald and Wang,2015),AI is considera

16、bly easier to use and implement in processes as it is a general-purpose technology that does not require the deployment of new hardware,deep userknow-how,or a substantial reconfiguration of business practices.As a consequence,theimpact of AI on productivity will likely be felt in the coming years al

17、ready(Brynjolfssonet al.,2018;Furman and Seamans,2019).2In this paper,we investigate the effects of AI on aggregate output and inflation,as1Generative AI generally refers to algorithms that can be used to create new content such as textor images,based on patterns detected in large training datasets.

18、The most well-known application isChatGPT,but there are numerous others.2Think of the advent of IT:firms needed to replace their paper-based systems with computers,familiarise themselves with the new concept of“software”,and train their staff.Meanwhile,publicinfrastructure,for example in the form of

19、 broadband,took years if not decades to provide sufficientcoverage.AI,on the other hand,can be used with the now near-ubiquitous smartphones and computers.Rather than requiring workers to learn how to use a fundamentally new system(think of an Excelspreadsheet versus a rolodex),AI can be used throug

20、h an intuitive language-based interface.1well as on output and employment in different sectors.We do so by first constructing ameasure of exposure to AI at the industry level.We then embed this exposure measureinto a macroeconomic multi-sector model,calibrated to the US economy using input-output ta

21、bles.We also use the model to perform counterfactual exercises.We start by constructing an industry-level measure of exposure to AI(AIIE)at the2-digit NAICS level.Building on the indicator developed in Felten et al.(2021),themeasure combines survey evidence on the extent to which AI applications can

22、 be usedin different workplace abilities with information on the importance of various abilitiesin different occupations and industries.A higher AIIE score indicates that an industryhas many occupations in which AI applications can be used.The industry with thehighest AIIE score is finance and insur

23、ance,followed by management of companies andenterprises.3Those with the lowest score are agriculture,forestry,fishing and huntingas well as transportation and warehousing.Note that our measure does not capturewhether AI complements or substitutes for specific occupations,which requires strongassumpt

24、ions to categorise each specific task(Pizzinelli et al.,2023).We then calibrate a macroeconomic model in which AI acts as permanent increasein the level of productivity with a differential impact across sectors.In particular,weassume that AI raises annual productivity growth by 1.5 percentage points

25、 for the nextdecade,in line with plausible estimates in the literature(Baily et al.,2023;GoldmanSachs,2023).We allocate the aggregate productivity increase across sectors using theAIIE measure.In addition to the usual set of nominal and real rigidities commonly usedto account for aggregate economic

26、fluctuations,the model features a detailed industrystructure on both the demand and production sides,following the work of Rees(2020).This allows it to account for industry-specific shifts in consumer preferences and workpractices,as well as industry-level and aggregate transmission mechanisms of AI

27、.Wefocus on 20 sectors in our baseline model,which broadly encompass consumption goodssectors closer to final demand,intermediate goods and investment goods.43It is therefore not surprising that the financial industry at large,and the central banking communityin particular,are actively engaging with

28、 AI.Banks and non-bank financial institutions have spent billionsupgrading their IT infrastructure in recent years and use it to analyse the large amount of data theypossess.The rise of fintech and big tech firms,which often rely on a combination of big data and machinelearning to provide their serv

29、ices,has further contributed to the rising footprint of artificial intelligence.At the same time,most central banks are already experimenting with machine learning and AI tools tosupport their economic analyses and policy decisions(Doerr et al.,2021;Araujo et al.,2024).4Whenever convenient to simpli

30、fy the exposition,we aggregate the 20 sectors into five:primaryindustries,secondary industries,distribution,professional services,and other services.2Our analysis considers two scenarios.In the first,households and firms observe theimpact of AI on productivity at each point in time and form expectat

31、ions about the futurepath of consumption,inflation and output based on those developments.However,theydo not foresee the boost to productivity from future AI developments.We refer to this asthe“unanticipated”case.In the second scenario,households and firms not only observethe effects of AI on produc

32、tivity that have already occurred,but also anticipate futureones.Accordingly,they adapt more quickly to adapt to the changes implied by the AI-induced productivity increase than in the”unanticipated”case.We refer to this as the“anticipated”case.We see these two cases as the extreme scenarios.Actual

33、expectationsare likely to fall somewhere in between.5We start with the unanticipated case and investigate the effects of AI on macroeco-nomic aggregates.Our results show that AI significantly raises output,consumption andinvestment in the short and long run,reflecting the positive effect of higher p

34、roductiv-ity growth on economic capacity.For inflation,initially the supply expansion acts as adisinflationary force,as higher TFP increases the economys output capacity.Yet afteraround four years,general equilibrium effects due to rising consumption and investmentraise aggregate demand and hence wa

35、ges sufficiently to make AIs impact inflationary.Responding to inflation,the policy rate first declines but then increases above its initiallevel to counteract the demand-driven rise in inflation.We then simulate the model under the anticipated scenario.The long-run responseof output,consumption and

36、 investment is identical to the unanticipated scenario.How-ever,as households fully anticipate the effects of AI on productivity today and in thefuture,they increase consumption more forcefully right away and postpone investment.Accordingly,output increases more slowly in this scenario.The paths of

37、inflation,interest rates and,to a lesser extent,the output gap,are starklydifferent in this scenario.Because demand increases in anticipation of future productivityincreases,AI adoption is initially inflationary.Inflation only begins to converge towardsthe unanticipated scenario as investment become

38、s positive and the productive capacityof the economy gets replenished.Policy rates,through the policy rules embedded in the5It is not obvious which of the two extremes is more likely.On the one hand,results from vector-autoregression models and the history of past general purpose technologies sugges

39、t that technologyshocks initially have disinflationary effects(Evans and Marshall,2009),which would be consistent withimperfect anticipation.At the same time,record stock market valuations of companies producing AIor the necessary hardware suggest that at least financial markets anticipate AI to sub

40、stantially raisegrowth.3model,rise immediately as inflation increases and decline only after around 10 years.Next,we inspect individual sectors to understand their relative importance in shapingaggregate dynamics.6The first insight we gain is that a sectors initial exposure to AIhas little relations

41、hip to its ultimate increase in value added output.The reason is that,ultimately,general equilibrium effects arising from higher demand for a sectors outputmatter much more than the initial increase in productivity as calibrated from our AIIEmeasure.Moreover,our results show that the ultimate increa

42、se in value added output issmaller in more labour-intensive sectors.An increase in productivity implies that moreoutput needs to be produced with a fixed amount of labour,which raises real wages.Sectors in which labour is a more important factor of production hence face higher costsand increase thei

43、r production by less.However,which sector is initially most affected by AI matters greatly for the responsein aggregate output and inflation.In our baseline calibration based on the AIIE measure,there is limited cross-sectoral variation in exposure to AI.While AIIE provides a solidgrounding for stud

44、ying the impact of AI across sectors,there remains a high degree ofuncertainty regarding the effects of AI.Our model allows us to contrast what wouldhappen if AI would,for example,mostly affect the services sector or the manufacturingsector.The counterfactual exercises show that,while output always

45、increases,it can growby up to twice as much for the same AI-induced increase in productivity when AI affectssectors producing consumption goods rather than investment goods.7When AI increasesproductivity in consumption goods sectors,freed-up labor moves to the investment goodssector,raising producti

46、on there,too.Since investment goods are used in the productionof intermediate and consumption goods,second round effects through sectoral linkageslead to an additional boost in output as the capital stock expands.Because of thesesecond-round effects,the initial decline in inflation is much more pron

47、ounced when AIaffects consumption-good sectors in the unanticipated case.In contrast,when AI raisesproductivity only in the investment goods sector,the aggregate responses in output andinflation are weaker.A second counterfactual exercise illustrates that results are qualitatively unchanged6We focus

48、 on the unanticipated case for ease of exposition,but the main insights go through moregenerally as we zoom in on long run effects across industries,which remain largely unchanged acrossscenarios.7For ease of exposition,our counterfactual exercises feature the productivity shock from AI as aone-off

49、event.Accordingly,the distinction between unanticipated and anticipated is immaterial.4if AI is a factor-specific technology rather than a general purpose technology.Focusingon the long run impact on output and inflation,we explore the consequences of makingAI either a labour-augmenting or capital-a

50、ugmenting technology.The long run effects onboth output and inflation are slightly more subdued when AI is factor-specific,especiallyif it is capital-augmenting.But our baseline result that output and inflation increase inthe long run remains qualitatively unchanged.Our findings inform the debate on

51、 the impact of AI on labor markets and output andhave implications for public policy.In particular,they suggest that public policy thatfosters the adoption of AI could lead to a Goldilocks scenario:in the case of imperfectanticipation,greater use of AI could ease inflationary pressures in the near-t

52、erm,therebysupporting central banks in their task to bring inflation back to target.In the longerterm,as inflation rises because of greater AI-induced demand,central banks task ofcontrolling inflation would become simpler,as they can dampen demand via monetarytightening.More generally,AIs positive c

53、ontribution to growth could also offset someof the detrimental secular developments that threaten to depress demand going forward,such as population aging,re-shoring and changes in global supply chains,as well asgeopolitical tensions and political fragmentation.These aspects underscore the need forp

54、olicies that foster the adoption of AI by firms and households.Policy efforts to spurAI adoption should focuses on sectors that produce consumption goods,as these promiseespecially high returns.Our main contribution is to provide a framework to model the effect of AI on aggregateoutput and inflation

55、 in both the short and long run.We thereby relate to recent studiesthat investigate the effect of AI on productivity and employment.Brynjolfsson et al.(2023)study the effects of AI on 5,000 customer-support agents working for a largeenterprise-software company.The agents were provided with an AI too

56、l,built on thelarge language models developed by OpenAI,with a staggered timing.Support agentswho used the AI tool could handle 13.8%more customer inquiries per hour and workquality,measured by the share of successfully resolved customer problems,improved by1.3%.Noy and Zhang(2023)asked experienced

57、business professionals from a variety offields,including marketers,grant writers,data analysts,and human-resource professionalsto write two business documents within their field.For the second document,half of theparticipants were randomly assigned to use ChatGPT.Professionals using ChatGPT werealmo

58、st 60%faster at writing and the rated quality was also higher.Finally,Peng et al.(2023)compare programmers using the GitHub Copilot AI tool to those who did not5use AI.They find that those using AI completed programming tasks more than twice asfast.Importantly,less-experienced programmers benefited

59、the most.Early evidence alsosuggest a positive correlation between AI adoption and firm productivity(Yang,2022;Czarnitzki et al.,2023)and innovation(Babina et al.,2024).Other papers discuss potential effects of AI on employment.Acemoglu et al.(2022)useestablishment-level data on online vacancies in

60、the United States:they find rapid growthin AI-related vacancies over 201018 that is driven by establishments whose workers en-gage in tasks compatible with AIs capabilities.AI adoption by establishments leads toreduced hiring in non-AI positions and a change in the skill requirements of remainingpos

61、tings.Yet,there is(so far)no significant effect of AI on aggregate employment andwage growth.Felten et al.(2019)provide evidence that,on average,occupations impactedby AI experience a small but positive change in wages,but no change in employment.Autor(2022)provides an overview of the labour market

62、implications of technologicalchange,with a focus on artificial intelligence.Lu(2021)develops a three-sector endoge-nous growth model to quantify the implications of AI for growth and welfare.To the best of our knowledge,our paper is the first to assesses the effect of AI onoutput and inflation in th

63、e short and long run.We do so by constructing industry-levelmeasures of exposure to AI and embedding these into a rich macroeconomic model.Ourpaper also provides novel insights on how to assess the direct and second-round effect ofAI on output,employment and hours worked in different sectors over ti

64、me.2Measuring the impact of AI on productivity acrossoccupations and industriesWhile AI,and in particular generative AI,is a general purpose technology,its impactdiffers across occupations and industries.Unlike automation through robots,which haspredominantly affected jobs and output in occupations

65、that require manual labour,AIis expected to have the largest impact in occupations with more cognitively demandingtasks.The reason is that much like other computer-based technologies,it can substitutefor routine cognitive tasks while complementing nonroutine cognitive tasks(Autor et al.,2003).This i

66、s best illustrated by contrasting surgeons and meat slaughterers.Thesetwo occupations require similar physical abilities(eg dexterity and steadiness),but their6cognitive content differs,with various forms of problem solving and logical reasoningbeing more relevant for surgeons(Felten et al.,2021).AI

67、 could affect productivity through several channels,but two stand out.The first channel is to directly raise the productivity of(cognitive)workers.Forexample,Brynjolfsson and McAfee(2017)show that access to a generative AI-basedconversational assistant improves customer support agents productivity b

68、y 14%.Forcollege-educated professionals,Noy and Zhang(2023)show that the chatbot ChatGPTsubstantially raised productivity in solving writing tasks,reducing the time required by40%and raising output quality by 18%.Meanwhile,Peng et al.(2023)find that softwaredevelopers that use AI could code more tha

69、n twice as many projects per week.The second channel is to spur innovation and thereby future productivity growth(Brynjolfsson et al.,2018;Baily et al.,2023).Most innovation,for example throughresearch and development but also through managerial activities,is generated in occu-pations that require h

70、igh cognitive abilities.Improving the efficiency of cognitive workhence holds large potential to generate further innovation that in turn improves efficiencyeven further.The macro-economic impact of AI on productivity growth could be sizeable.Assessingthis impact requires aggregating industry-specif

71、ic productivity gains,which can be doneby multiplying the size of the productivity increase with the relative size of the sector.8Improvements in productivity in an industry can hence have large aggregate effects.Different studies provide estimates for AIs impact on annual labour productivity growth

72、(ie output per employee)over the next decade,with estimates ranging from 1pp to 1.5pp(Baily et al.,2023;Goldman Sachs,2023).To assess the impact of AI on productivity in different industries,we first constructan industry-level measure of exposure to AI(AIIE)at the 2-digit NAICS level.9We buildon Fel

73、ten et al.(2021),who provide an index called AI Occupational Exposure(AIOE)that is widely used in the literature(see eg Acemoglu et al.(2022)and Autor(2022)and constructed as described below.We then attribute the estimated impact of AI on8This is well-established in the economics literature in the s

74、eminal work of Hulten(1978).See Baqaeeand Fahri(2019)for a recent extension to Hultens theorem.9NAICS refers to the North American Industry Classification System,the standard used by federalstatistical agencies in classifying business establishments for the purpose of collecting,analysing,andpublish

75、ing statistical data.See the dedicated webpage by the US Census Bureau for more details.7aggregate productivity to each industry,depending on its exposure to AI and its relativesize in the economy.We construct the AIIE measure in four steps.In a first step,ten AI applicationscovering AIs most likely

76、 use cases are linked to a list of 52 workplace abilities.10Foreach ability,survey respondents need to indicate whether they think the respective AIapplication can be used.The result is a relatedness measure for each occupation-abilitycombination that ranges between zero(no relation)and one(high rel

77、ation).In a secondstep,each abilitys exposure is constructed as the sum of the relatedness value acrossall AI applications.It ranges from zero(no exposure)to 10(high exposure).The thirdstep involves computing each occupations exposure to AI(AIOE)by taking the weightedaverage across the 52 abilities

78、exposures to AI,with weights given by abilities prevalencein each occupation(provided by O*Net).We standardise the resulting AIOE variable torange from zero to ten,with higher values indicating a greater importance of AI withinan occupation.To construct exposure to AI at the industry level(AIIE),we

79、use data onoccupations employment shares within each two-digit industry(provided by the Bureauof Labor Statistics)as weights to average across AIOEs.We then standardise AIIE sothat the industry with the highest exposure value has a score of one;all other industriesAIIE are then expressed as a fracti

80、on of the AIIE of the highest exposure industry.Figure 1 plots the five occupations with the highest and lowest AIOE scores.Asdiscussed in Felten et al.(2021),the highest AIOE scores consist mostly of white-collaroccupations that require advanced degrees,such as genetic counsellors,financial exam-in

81、ers and actuaries.The lowest-scoring occupations predominately require a high degreeof physical effort and include,for example,dancers,fitness trainers or iron and rebarworkers.Figure 2 plots the(standardised)AIIE at the industry level(grey bars)as well asthe respective employment shares(red diamond

82、s).The industry with the highest AIIEscore is finance and insurance,followed by management of companies and enterprises.Those with the lowest score are agriculture,forestry,fishing and hunting as well as trans-portation and warehousing.Overall,there is not very large variation in exposure across10Th

83、e ten applications are:image recognition,visual question answering,image generation,readingcomprehension,language modelling,translation,speech recognition,abstract strategy games,real-timevideo games and instrumental track recognition.A list of 52 cognitive,physical,sensory and psychomotorworkplace

84、abilities is provided by the O*Net database.Cognitive abilities include,for example,deductivereasoning or oral comprehension.Physical abilities include stamina or trunk strength,among others,whereas psychomotor abilities include arm-hand steadiness or finger dexterity.802468occupational exposure to

85、AI(AIOE)Genetic CounselorsFinancial ExaminersActuariesPurchasing Agents,Except Wholesale,Retail,and Farm ProductsBudget AnalystsPressers,Textile,Garment,and Related MaterialsReinforcing Iron and Rebar WorkersHelpersPainters,Paperhangers,Plasterers,and Stucco MasonsFitness Trainers and Aerobics Instr

86、uctorsDancersFigure 1:AI exposure of top/bottom occupations.industries,the reason being that within industries,there are a lot of occupations withvarious degrees of exposure to artificial intelligence.Our measure reflects exposure to AI and does not capture whether AI will substituteor complement an

87、y particular occupation.While there is general agreement that AI ispositive for productivity,there is no consensus on the scope for AI complementing orsubstituting tasks.One option is to use judgement to categorise each specific occupationas more/less at risk of displacement(Pizzinelli et al.,2023;C

88、azzaniga et al.,2024).Thedrawback is that this approach requires strong assumptions that are difficult to verify,which risks assuming the final impact across occupations from inception.Instead,wecalibrate exposure to AI and let the model decide which industries witness increased ordecreased employme

89、nt.In this way,the final impact across industries is disciplined bygeneral equilibrium effects and is not a function of our priors.90.05.1.15.2employment share0.25.5.751AIIE score34244454849562digit NAICS codeindustry exposure to AI(left)industry employment share(rig

90、ht)Figure 2:The direct impact of AI across industries3A multi-sector modelThis section describes the building blocks of our multi-sector model and explains how weuse it to chart the economic effects of AI.We then provide intuition for our core resultsusing simplified versions of the full model.3.1Th

91、e modelWe work with the multi-industry New-Keynesian Dynamic Stochastic General Equilib-rium model presented in Rees(2020).The model features a detailed industry structurein both its demand and supply sides.This allows it to capture the key industry-level andaggregate transmission mechanisms of AI.W

92、e use the model to account for the directeffects of AI on industry-level productivity and then to trace through the effects of thesechanges across the economy to assess the implications for aggregate outcomes.The model consists of a closed economy featuring households,firms,the governmentand the cen

93、tral bank.1111For the sake of expositional clarity,we focus on the aspects of the model that are most relevant forour application.We provide the full model as a separate file.10Households make consumption,work,investment and saving decisions to maximisetheir lifetime utility,subject to an intertempo

94、ral budget constraint.Their utility functionis given by:Xt=0t?log(Ct hCt1)AN1+N1+t?(1)where Ctand Ntare household consumption and labour supply.The parameters,hand are the households intertemporal discount rate,its habits parameter and its Frischlabour supply elasticity.The intertemporal budget cons

95、traint in turn is given by:PC,tCt+PI,tIt+Bt+1Rt Bt+FXj=1?PC,trKj,tkj,t+wj,tnj,t?(2)where Itis the households total investment in physical capital,PC,tand PI,tare theprices of the consumption and investment goods,Bt+1is a risk free nominal bond thatpays one unit of the consumption good in period t+1

96、and Rtis the interest rate of thatbond.The economy features F industries.The variables kj,tand nj,trepresent the totalsupply of capital and labour from the household to industry j.In turn,rKj,tand wj,tarethe return on capital and nominal wage paid by that industry.The aggregate consumption and inves

97、tment goods in Equation(2)consist of bundlesof products from individual industries.For example,the aggregate consumption bundleis:Ct=FXj=11cjc1j,t1(3)where cj,tis the amount of output of industry j used to produce consumption goodsat time t.Conditional on the relative prices of industries output,the

98、 parameter cjdetermines the weight of industry j in the aggregate consumption bundle.The parameter determines the degree of substitutability between the output of different industries inconsumption:a higher(lower)value means that the output of different industries is more11(less)substitutable in con

99、sumption.The aggregate labour supply that appears in the household utility function is also aweighted sum of labour supply to individual industries:Nt=FXj=11njn+1j,t+1(4)where njcaptures the relative disutility the household receives from supplying labourto industry j and controls the substitutabili

100、ty of work across industries.In the limit,if =,workers are indifferent between working in different industries.On the production side of the model,each industry consists of many firms producingdifferentiated product varieties under monopolistic competition.Individual firms produceoutput using a mult

101、i-stage production process.The first stage combines labour andcapital according to the following production function:fj,t(i)=?1f,jnj,t(i)1+(1 f,j)1ksj,t(i)1?1(5)where fj,tis an aggregate of labour and capital used by firm i in industry j and nj,t(i)andksj,t(i)are the amount of labour and capital ser

102、vices employed by the firm.The parameter is the elasticity of substitution between capital and labour.In a second stage,firms combine the labour and capital bundle,fj,t,with intermediateinputs sourced from other industries:yj,t(i)=aj,t?1y,jfj,t(i)1+(1 y,j)xj,t(i)1?1(6)where yj,t(i)is the gross outpu

103、t of firm i in industry j and xj,t(i)is the amount ofintermediate inputs used by the firm.The term aj,trepresents total factor productivity,which is industry-specific and common to all firms in that industry.The parameter is the elasticity of substitution between intermediate inputs and the aggregat

104、e of labourand capital.The intermediate input is itself a bundle of intermediate goods from the other indus-tries,with the elasticity of substitution between different varieties of intermediate goods12given by.Value-added output is equal to a firms gross output minus its intermediate inputs.In the c

105、ase where all final goods prices are:yvaj,t(i)=yj,t(i)xj,t(i)(7)Market clearing requires that the gross output of industry j(yj,t)equals the sumof demand for the good as a consumption good(cj,t),investment good(ij,t)and publicdemand good(gj,t),or as an intermediate input(where xj,k,tis the output of

106、 industry jused as an intermediate input in industry k):yj,t=cj,t+ij,t+gj,t+FXk=1xj,k,t(8)The presence of intermediate inputs creates a rich network of linkages across its in-dustries.For instance,higher productivity in the manufacturing industry lowers costs forfirms operating in industries that us

107、e manufacturing goods as an input.Two factors dic-tate the importance of inter-industry linkages in the model:the weight of intermediateinputs in industry production functions and the substitutability between intermediateinputs and other factors.The model includes a number of nominal and real rigidi

108、ties.Firms face Calvo-styleprice rigidities,with the degree of price stickiness varying across industries.Householdsare assumed to unionise,giving them a degree of monopoly power in the labour market.There too,Calvo wage rigidities exist.The models real rigidities include habits,invest-ment adjustme

109、nt costs and capital utilisation costs.While these rigidities do not alterthe long run consequences of AI adoption,they materially influence the models short-rundynamics.The two remaining agents are the government and the central bank.We assume thatgovernment expenditure as a share of nominal GDP fo

110、llows an autoregressive process,funded by lump-sum taxation.The central bank adjusts its policy interest rate in responseto deviations of inflation from target and the output gap,defined as the deviation of realGDP from the models flexible price benchmark.The model includes 20 industries,correspondi

111、ng to the two-digit NAICS classification13described in Section 2.We calibrate the model to match key features of the US economy,using input-output tables to pin down the weights of capital,labour and industry-specificintermediate inputs in the industry production functions,the weights of consumption

112、,in-vestment and government spending in domestic demand and the weights of each industryin the consumption,investment and government spending bundles.Most of the otherparameters controlling the dynamics of the model,such as the habits and investmentadjustment cost parameters,or the aggregate labour

113、supply elasticity,are taken from theliterature.Table 1 provides a summary of these for details we refer the reader to Rees(2020).Table 1:Calibration of key parametersParameterDescriptionValueDiscount rate0.99hHabits0.70SppInvestment adjustment cost3.00Elasticity of substitution in demand CES0.90Elas

114、ticity of substitution between capital and labour0.95Elasticity of substitution between intermediates and capital/labour0.60Elasticity of substitution between intermediates0.40Frisch labour supply elasticity2.00wLabour supply elasticity across industries5.00Depreciation rate0.02rTaylor rule-autoregr

115、essive parameter0.80Taylor rule-response to inflation1.50gapTaylor rule-response to output gap0.25stickyCalvo-sticky price sectors0.80semiflexCalvo-semi-flexible price sectors0.50flexCalvo-flexible price sectors0.25pPrice indexation0.20wCalvo-wages0.75Note:Sticky price sectors are Agriculture and Mi

116、ning;Semi-flexible price sectors are Utilities,Manufac-turing,Retail trade,Wholesale trade and Transport.Finally,we use the AIIE measure constructed above to simulate the impact of AIadoption by industries.Guided by the estimates in the literature,we do this in a waysuch that the annual impact on to

117、tal TFP growth aggregates to a value of 1.5%for adecade.3.2Modelling the effects of AIWe simulate the effects of AI as follows.14We first linearise the model along its balanced growth path.This provides us witha baseline case in which there is no AI-induced boost to productivity.We present all ofour

118、 results as deviations from this baseline.The models linearised structural equationstake the form:Ayt=C+Byt1+DEyt+1+Ft(9)where ytis the vector of the models endogenous variables and tis the vector of structuralshocks,which we take to be i.i.d.without loss of generality.Note that in solving the model

119、we keep track of the steady state of the linearised variables,resulting in the inclusion ofthe vector C in Equation(9).If it exists and is unique,the standard rational expectations solution to Equation(9)is:yt=J+Qyt1+Gt(10)We model the effects of AI as a sequence of permanent increases in the level

120、of industry-specific productivity(i.e.the steady state of the terms aj,tin Equation 6).As well astheir direct effects on industry output,these changes have general equilibrium effects thatalter the economys entire balanced growth path.We account for these changes usingthe approach outlined in Kulish

121、 and Pagan(2017).Specifically,for each period in oursimulations,we calculate a new system of linearised equations:Atyt=Ct+Btyt1+DtEyt+1+Ftt(11)where At,Bt,Ct,Dtand Ftrepresent the linearised structural equations for the industry-specific productivity levels in period t.The presence of forward-lookin

122、g expectations in Equation(11)requires us to take astance on economic agents beliefs about the future path of productivity to solve for thereduced form of this system.Various plausible alternatives suggest themselves.Advances in AI technology havereceived considerable public attention,particularly a

123、fter the release of ChatGPT 3.0 inNovember 2022.Rapid increases in the share prices of AI-related firms suggest that finan-cial market participants,at least,have started to factor the transformative possibilitiesof these technologies into their investment decisions.That said,considerable uncertainty

124、exists about the real-world application of existing AI models,not to mention their future15evolution.As such,it would not be surprising if households and firms responded cau-tiously to AI developments,adjusting their economic decisions as the technology evolvesand its real-world applications become

125、more apparent.Acknowledging this uncertainty,we consider two cases.In the first,which we referto as the“unanticipated”case,models agents observe AI-induced productivity increaseswhen they occur,but do not anticipate further gains in the future.In this case,thesolution to the reduced form of the mode

126、l is a time varying VAR of the form:yt=Jt+Qtyt1+Gtt(12)where Jt,Qtand Gtare the standard rational expectations solutions to Equation(11)in each period.In the second case,agents observe AI-induced productivity increases as they occur andcorrectly anticipate their further evolution.In this“anticipated

127、”case we solve the modelrecursively.We start in period T,at which point we assume that no further AI-inducedproductivity increases will occur.After solving for the reduced form solution matrices,JT,QTandGTin that period,we can then use the formulas provided in Kulish andPagan(2017)to determine all p

128、revious reduced form matrices.3.3Building intuition through parsimonious modelsThe effects of AI on macroeconomic aggregates in our full model reflect a rich set ofdynamics and mechanisms.To build intuition for our results,we first show how AI-induced increases in industry productivity propagate in

129、two simpler versions of the model.Each highlights a specific mechanism at work in the full model.The first scaled-down model illustrates the role of cross-industry differences in produc-tion technologies.The model has three industries:the first is labour intensive,the secondis capital intensive and

130、a third is balanced between those two extremes.We calibratethe model so that each industry accounts for the same share of aggregate consumption,investment and intermediate inputs before the introduction of AI.We vary these weighsin the second scaled-down model below.We consider four scenarios.In the

131、 first three scenarios,AI adoption leads to animmediate,permanent and one-time 10%increase in the TFP of a single industry,and16has no effect on productivity in the other industries.In the fourth,AI adoption raisesproductivity in each industry by 313%.12The path of GDP is similar across the scenario

132、s(Figure 3,left-hand panel).Thatis,the response of GDP to AI does not depend on whether its effects are concentratedin labour-or capital-intensive industries.It rises immediately by around 3%.13It thenincreases further over time,even though the level of productivity in each industry remainsconstant

133、after the first period.The additional output increase,beyond that implieddirectly by the increase in productivity,occurs because higher productivity enables anexpansion in the economys capital stock and the production of more intermediate inputs,which both deliver an additional boost to the economys

134、 productive capacity.The finalincrease in GDP of around 12%is hence much larger than the initial rise.051015Years-5051015%GDPAll industriesLabour-intensiveBalancedCapital-intensiveAI boosts TFP in:051015Years-5051015%GDP and industry value addedAI boosts TFP in all industriesGDPLabour-intensiveBalan

135、cedCapital-intensiveLong run change in input and output pricesAI boosts TFP in all industries0 0 0Return oncapitalReal wage Relative price-5051015%Labour-intensiveBalancedCapital-intensiveFigure 3:Simple model 1 Cross-industry differences in production technologiesAt the industry level,output trajec

136、tories differ substantially.In Scenarios 1-3,theindustry that experiences the productivity boost naturally experiences the largest increasein output.But even in Scenario 4,where AI delivers the same productivity boost to eachindustry,cross-industry output differences arise(Figure 3,centre panel).In

137、particular,value added output rises more in capital-intensive industries than labour-intensive ones.The behaviour of input and output prices provides intuition for this result.In the long-run,the size of each industrys capital stock adjusts to equalise the rate12Because the industries are initially

138、equally-sized,the weighted average increase in industry produc-tivity in Scenario 4 is the same as that in Scenarios 13.13This is roughly the increase in GDP that would result if the introduction of AI had no effect on thesupply of other factors of production and influenced GDP only through its effe

139、ct on TFP.17of return on capital across industries.AI adoption that raises the return on capital ina given industry encourages investment in that industry.In time,higher investmentincreases the size of that industrys capital stock and drives down its return.14Hence,although capital returns vary alon

140、g the transition to a new balanced growth path,as longas the effects of AI on productivity growth are temporary,and do not affect householdsdiscount rates or capital depreciation rates,the return on capital will ultimately convergeto its original level(Figure 3,right-hand panel,first three bars).The

141、 scope to increase the labour force is much more limited.Hence,for workers,AI-induced productivity gains lead to permanently higher real wages.And,when AIraises productivity by the same amount in each industry,the increases in real wages deflated by aggregate consumer prices are also similar(Figure

142、3,right-hand panel,centre bars).15The behaviour of rates of return to capital(flat)and real wages(higher)implies thata proportional increase in TFP will raise input costs more in labour intensive industries.As firms price output as a markup over their marginal costs,the relative output pricesof labo

143、ur-intensive industries rise and those of capital-intensive industries fall(Figure3,right-hand panel,right-hand bars).This,in turn,induces households and firms tosubstitute towards more capital-intensive goods and services,explaining the differencesin value-added across industries in Figure 3,centre

144、 panel.The second scaled-down model illustrates the implications of differences in the use ofindustry output.As before,we build a three-sector version of the model,calibrated sothat each sector accounts for the same share of value added output before the adoptionof AI.In this case,industries share s

145、ame production technology,i.e.their use of capital,labour and intermediate inputs are initially identical.However,they differ in how theirproducts are used.We calibrate the weights of the three industries in the consumption,investment and intermediate-input CES bundles so that each industry speciali

146、ses in theproduction of goods for one particular use.We simulate the same four scenarios as forthe previous model.In contrast to the first scaled-down model,the trajectory of aggregate GDP differsmarkedly across scenarios(Figure 4,left-hand panel).When the effects of AI are con-14Investment adjustme

147、nt costs prevent the adjustment in the size of the capital stock from takingplace immediately.15Because we assume that labour is imperfectly substitutable across industries,wage rates do notequalise.18centrated in the industry specialising in the production of consumer goods,GDP risesby more than 20

148、%in the long run(blue line).When instead they are concentrated inthe industry producing investment goods,the long-run rise in GDP is around 5%(orangeline).The other two scenarios lie between these two extremes.051015Years-50510152025%GDPAll industriesConsumer goodsInvestment goodsIntermediate goodsA

149、I boosts TFP in:Long run change in relative pricesAI in consumergoodsAI in investmentgoods-10-50510%Consumer goodsInvestment goodsIntermediate goodsRelative price of:Long run change in real wagesAI in consumergoodsAI in investmentgoods0510152025%Consumer goods(C)Investment goods(C)Intermediate goods

150、(C)Consumer goods(P)Investment goods(P)Intermediate goods(P)Consumer(C)/producer(P)real wages in:Figure 4:Simple model 2 Differences in the use of industry outputOnce again,the intuition for this result starts with an examination of the behaviourof prices and input costs.A disproportionately large i

151、ncrease in productivity due to AI adoption within a par-ticular industry lowers the relative price of that industrys output.For example,if higherproductivity occurs in the consumer goods sector,the relative price of consumption goodsfalls(Figure 4,centre panel,left bars).If instead it is concentrate

152、d in the investmentgoods sector,the relative price of investment goods falls(right bars).As factor mobility limits divergences in wages and capital returns,AI adoption leadsto a similar rise in real wages deflated by consumer prices in all industries.Andit does so even when the effects of AI are con

153、centrated in only one industry(Figure 4,right-hand panel).But what matters for firms are factor costs deflated by their own output prices.Andmovements in relative prices mean that real producer wages(i.e.,nominal wages deflatedby a firms own prices)can vary substantially,even when real consumer wage

154、s(i.e.,nominal wages deflated by the overall consumer price index)move by a similar amount.19For example,when the effects of AI are concentrated in the consumer goods sector,realproducer wages for firms producing consumer goods rise much more than those in othersectors because of the relative declin

155、e in consumer goods prices.When the effects of AIconcentrate in the investment goods sector,the difference in real producer wages is evenlarger again a function of the large relative price swings for investment goods in thatscenario.Profit maximisation by firms creates a link between a firms costs a

156、nd its production.At the margin,firms that face higher costs evaluated in producer prices will cut backon production.When the effects of AI are concentrated in consumer goods industries,which lie at the end of the economys production chain,this constraint on productionhas positive spillovers.It free

157、s up resources to increase the production of investment andintermediate goods,both of which further expand the economys productive capacity.Incontrast,if lower relative prices of investment goods decrease that industrys production,the spillovers are negative as the economys ability to produce other

158、goods and servicesis also diminished.The scaled-down models also provide some intuition for the possible implications ofAI adoption for inflation and interest rates(Figures 13 and 14 in Appendix A).In the first model,inflation trajectories are similar across the four scenarios.For thespecific calibr

159、ation we consider,the introduction of AI initially lowers inflation(Figure13,left-hand panel).Inflation is only temporarily lower,however.It soon turns positive,alongside a positive output gap and a rise in the policy rate(centre and right-handpanels).The responses of inflation,the output gap and po

160、licy rates for the scenarios con-structed using the second scaled-down model are qualitatively similar,albeit with greatervariation in outcomes across the four scenarios(Figure 14).In one extreme,when theeffect of AI concentrates on the consumer goods industry,the effects on inflation,outputgap and

161、policy rates substantially more pronounced.In the other extreme,effects areweakest when AI boosts TFP in the investment goods sector.Why is the introduction of AI which expands aggregate supply inflationary inthese scenarios?The answer is two-fold.First,the introduction of AI raises incomes,which in

162、creases aggregate demand.Second,it creates the need for additional investmentto expand the capital stock and the production of intermediate inputs.This additional20investment increases aggregate demand immediately,and aggregate supply only with alag,and hence is inflationary.Taken together,the resul

163、ts from our scaled-down model variations deliver useful in-sights to assess the results from our main model,which we present in the next section.Differences in production technologies as reflected in the relative use of factors of produc-tion are not the main driver of the evolution of key macroecon

164、omic outcome variables.16Differences in the use of industry output,in contrast,can lead to quite different dy-namics in response to an AI-induced increase in productivity.When the effects of AIconcentrate in industries that are closer to final demand(i.e.,“downstream”),generalequilibrium price effec

165、ts shift resources to upstream industries,which in turn reinforcesthe positive direct effects on output through their high production linkages.174The macroeconomic impact of artificial intelligenceHaving built intuition for they key mechanisms driving our model dynamics,we nowpresent results from th

166、e full model.Recall that AI is represented as an increase inproductivity growth of 1.5 ppt annually over the next decade,allocated to sectors basedon the AIIE measure constructed in Section 2.In what follows we first discuss howAI adoption affects key macroeconomic aggregates,before analysing its im

167、pact acrossindustries.4.1The macroeconomic impact of AIThe unanticipated case.We first discuss results when households and firms do notanticipate future AI-induced productivity increases.Productivity improvements from AI adoption lead to a significant increase in GDP.Growth is fastest in the first 1

168、0 years i.e.the period in which AI directly raises industry-level TFP at which point GDP is almost 30%higher than it would have been without16To be sure,there are cross-industry differences even when AI delivers the same productivity boostto all industries.As discussed above,this can be rationalised

169、 by assessing general equilibrium effectsthrough prices.But they will not be a driver of large differences in the aggregate.17That said,we emphasise that the patterns of inflation,the output gap and policy rates obtain fromthese models are specific to the calibration used in this exercise.21AI adopt

170、ion(Figure 5,green line in panel(a).GDP continues to increase even after thedirect productivity gains from AI adoption are exhausted,albeit at a slower pace,as ittakes time for firms to adjust their capital stock and use of intermediate inputs to takefull advantage of AI.The level of GDP ultimately

171、stabilises around 35%above the no-AIbaseline.The paths of aggregate consumption and investment broadly resemble that of GDP(Figure 5,green lines in panels(b)and(c).The level of investment overshoots,therebydelivering the required increase in the economys capital stock,before converging to itslong-ru

172、n level.051015Years-20-5060%UnanticipatedAnticipated(a)GDP051015Years-20-5060%(b)Consumption051015Years-20-5060%(c)InvestmentFigure 5:Output,consumption,and investmentIn this scenario,AI adoption initially lowers inflation(Figure 6,green line in panel(a).Faster produ

173、ctivity growth increases the economys productive capacity,loweringfirm costs and relaxing supply constraints,thus acting as a disinflationary force.Thedemand response,on the other hand,takes time to materialise,in part because of frictionssuch as habits in consumption and investment adjustment costs

174、.18Policy rates respondaccordingly in this initial phase and decline with inflation.19Over time,however,theeffects of AI adoption on aggregate demand start to dominate its effects on aggregate18The timing of inflation is also a function of,among other things,the price stickiness of the industriesaff

175、ected and where they sit in the production chain(i.e.whether they supply consumption or investmentgoods),since the inflation measure we consider in the Figure is for consumption goods.This is discussedin some more detail in the context of our scaled-down models in Section 3.3.19The lagged term in th

176、e Taylor rule that determines the policy rate response prevents the central22supply.Consequently,AI adoption ultimately raises the inflation rate after around fouryears.Inflation peaks about 0.75 percentage points above its level without AI adoption.In response to higher demand-driven inflation,the

177、central bank steadily raises its policyrate(Figure 6,green line in panel(c).051015Years-101234pptUnanticipatedAnticipated(a)Inflation051015Years-101234%(b)Output gap051015Years-101234ppt(c)Policy rateFigure 6:Inflation,output gap and policy rateThe anticipated case.We now discuss results when agents

178、 correctly anticipate theentire future path of AI-induced productivity increases.GDP increases more slowly in the anticipated case than the unanticipated one(Figure5,orange line in panel(a).This is because households,who correctly foresee that AIwill raise productivity in the future,bring forward co

179、nsumption in order to smooth itstrajectory over time(panel(b).To accommodate higher consumption,which runs aheadof actual AI-induced productivity improvements,investment falls(panel(c).As a result,the economys capital stock grows more slowly than in the unanticipated case,resultingin a lower level o

180、f GDP compared to the unanticipated case.The paths of inflation,interest rates and the output gap differ markedly in the antic-ipated scenario from their trajectories in the unanticipated one(Figure 6,orange lines).Although the level of output is lower in the anticipated case than the unanticipated

181、one,bank form lowering rates quickly enough to prevent disinflation(or raising them quickly enough toprevent inflation later).23inflation is significantly higher,peaking at more than 2 ppt higher than it would havebeen without AI adoption.It then remains above its baseline level throughout the sce-n

182、ario.This,in turn,induces a large rise in policy rates.The output gap is modestlypositive throughout the scenario.Taken together,these results illustrate that while different expectation formationmechanisms change the transitory dynamics of macro variables,they do not affect thelong-term impact of A

183、I on the economy.Our results can be seen as supporting theproductivity effect in the task-based framework of Acemoglu and Restrepo(2018).4.2The impact of artificial intelligence across sectorsWe now turn to the effects of AI across sectors.We focus on long-run outcomes,whichare the same in both the

184、anticipated and unanticipated cases.Figure 7 shows the long run increase in value-added output by industry.For ex-position,we color-code the industries into five groups primary industries,secondaryindustries,distribution,professional services and other services.As a reference point,thehorizontal lin

185、e shows the economy-wide long-run increase in aggregate GDP due to theintroduction of AI.Three observations stand out.First,value-added output rises in all industries,re-flecting the nature of AI as a general-purpose technology.Second,the impact variessignificantly across industries,ranging from a n

186、early 50%increase in value-added outputin manufacturing and real estate services to about 20%in education and management ser-vices.Third,there is no direct mapping between an industrys initial exposure to AI andthe long run increase in value added output(see Figure 8a).Some sectors that are ex-pecte

187、d to receive the smallest initial productivity boost from AI,such as arts&recreationservices,record a relatively large output increase.Others that are expected to receive aparticularly large direct boost from AI,such as management services,see smaller outputincreases.In general,primary and secondary

188、 industries display the largest increases invalue added output,while professional services are in the bottom half of the distribution,with a couple of notable exceptions such as Information&Communications and RealEstate Services.What explains the differences in value added growth across industries?I

189、t turns out24EducationManagementProfessionalOtherGovernmentAdministrationHealthArtsConstructionTransportRetailRecreationWholesaleFinanceUtilitiesMiningInformationAgricultureManufacturingRealEstate01020304050%Primary industriesSecondary industriesDistributionProfessional servicesOther servicesFigure

190、7:Long-run increase in industry value added99.51010.511AI-driven productivity increase(ppt)20253035404550Increase in value added(%)Primary industriesSecondary industriesDistributionProfessional servicesOther services(a)AI and value-added increase0204060Labour share of gross output(%)20253035404550In

191、crease in value added(%)(b)Value added increase and labour shareFigure 8:Value-added,AI and labour sharesthat the mechanism in the first simple model presented in Section 3 differences in thecapital and labour intensity of production is crucially important.The relationship25between an industrys labo

192、ur-intensity,as proxied by the income share of labour in thatindustrys gross output,and its increase in value-added output as a result of AI adoptionis close.More labour-intensive industries record the smallest increases in value addedoutput(Figure 8,panel(b).As in the simple model,differences in la

193、bour intensitytranslate into movements in relative prices,with the relative prices of labour-intensiveindustries rising relative to more capital intensive ones(Figure 9a).AI adoption also leads to a reallocation of labour across industries(Figure 9b).Ingeneral,employment increases in services indust

194、ries that experience higher relative pricesand the smallest increases in value added output.Higher selling prices allow firms inthese industries to raise wages more than firms in industries whose relative prices decline.This induces workers to adjust their labour supply towards these industries.In c

195、ontrast,in the capital-intensive industries that record the largest increases in value-added,andwhere relative prices decline,hours worked falls.Why do the industries that record the largest increases in hours worked not recordlarger increases in output than those where hours worked declines?The ans

196、wer lies inthe behaviour of capital.The productivity boost from AI adoption allows for a materialincrease in the economys capital stock.Because labour and capital are imperfect substi-tutes,a given proportional increase in the capital stock delivers a larger output increasein capital intensive indus

197、tries(i.e.those where y,jis lower)than labour intensive ones.The above discussion describes changes in value added output at the industry level.Inpractice,much of the increase in the output of primary and secondary industries will beused in downstream industries as intermediate inputs in the product

198、ion of final goods andservices for households and businesses.To illustrate these effects,Figure 10 decomposesthe increase in value added by industry into the increase in its value added that shows upas final demand(i.e.consumption,investment or government expenditure)and the partthat is used as inte

199、rmediate inputs by other industries.It further breaks down these twocategories into the change due to“income”effects(i.e.higher aggregate demand)and“substition”effects(i.e.changes in relative prices).The overall picture that emerges isthat income effects matter much more than substitution effects.26

200、20304050Increase in value added(%)-10-505101520Increase in relative price(%)Primary industriesSecondary industriesDistributionProfessional servicesOther services(a)Value added and relative prices20304050Increase in value added(%)-6-4-2024Increase in hours worked(%)(b)Value added and hours workedFigu

201、re 9:Value-added,hours worked and relative pricesAgricultureMiningUtilitiesConstructionManufacturingWholesaleRetailTransportInformationFinanceRealEstateProfessionalManagementAdministrationEducationHealthArtsRecreationOtherGovernment-30-20-5060%Income:Final demandIncome:Intermediate demand

202、Substitution:Final demandSubstitution:Intermediate demandFigure 10:Decomposition of industry value-added275Counterfactual scenariosOur setting has two distinctive features that matter for the interpretation of our mainresults.First,the effect of AI as calibrated from the AIIE measure shows relativel

203、y littlevariation across sectors.The measure is both theoretically and empirically grounded.However,substantial uncertainty remains as to how AI will ultimately affect individualindustries.Second,AI is neither capital-nor labour-specific,but is rather a generalpurpose technology affecting overall TF

204、P.This is in line with the common understandingin the literature.But again,uncertainty remains as to whether AI affects more capitalor labour.In this section we leverage the flexibility of our model to perform counterfactualanalyses that explore how the results change when we relax each of these fea

205、tures.Tosimplify the exposition,we show results from only the case where future TFP adoptionis unanticipated.However,the qualitative message from the anticipated case is similar.Heterogeneity in initial impact across industries.We first consider how our re-sults would change if the productivity boos

206、t from AI adoption was concentrated in specificindustries.We focus on the impact on output and inflation and consider four scenariosmarked by which sectors are affected.Concretely,we assess the effects of AI raisingproductivity only in professional services,construction,healthcare or manufacturing.W

207、echoose the first industries because their output is used primarily as intermediate inputs,investment goods and consumption services,respectively.We choose the manufacturingindustry because of the diversity of its final output destinations.To make the exercisescomparable,for each industry we assume

208、that the increase in productivity growth fromAI adoption is equal to 1.5%divided by that industrys share of total value added.Forcomparison,we also show the results from the baseline model.20All other aspects of themodel remain as in the baseline case.Figure 11 presents the results of these exercise

209、s.Output grows irrespective of whichindustries the effects of AI concentrate on,although there is meaningful cross-industryvariation(panel(a).For example,when the effects of AI affect only professional servicesor construction,the long run effect on output is smaller than when all industries are20The

210、 results when we assume that the effects of AI are felt equally in all industries are similar to thebaseline.28equally affected.Conversely,if the effect of AI concentrates on healthcare or manufac-turing,the long run effect on output can over one third larger than in the equally-affectedcase.The res

211、ult that the increase in GDP is larger when the productivity boost fromAI is concentrated in consumer-facing industries mirrors that in the second simple modeldescribed in Section 3.3.051015Years00%EqualHealthcareConstructionProfessional servicesManufacturing(a)Gross Domestic Product05101

212、5Years-2-1012ppt(b)Inflation(year-on-year)Figure 11:GDP and InflationInflation displays considerable heterogeneity across the scenarios.When the effectsof AI are concentrated in the professional services industry,there is no initial disinfla-tion at all,even though that scenario features the smalles

213、t increase in GDP.When onlyconstruction is affected,the long run effect on inflation is small despite an initial mild dis-inflation.In contrast,when the effect concentrates on manufacturing,there is substantialdisinflation early on,and it takes at least a decade for inflation to become positive.The

214、different inflation responses reflect,in part,different required paths of relativeprices.Higher productivity in a single industry lowers its relative price.When the pro-ductivity boost occurs in an industry that accounts for a large share of consumptionoutput like healthcare the relative price adjus

215、tment shows up largely in lower con-sumer price inflation and higher inflation in the investment goods deflator.In contrast,when the productivity boost occurs in an upstream industry,like professional services orconstruction,inflation in the investment goods deflator tends to fall significantly,whil

216、ethe consumer price inflation decreases only slightly,or even rises.29General purpose versus factor-specific technology.We now use our model tostudy the evolution of output and inflation under two counter-factual scenarios:1)AIas a labour-specific productivity increase and 2)AI as a capital-specific

217、 productivityincrease.Figure 12 presents the results of this exercise,where for reference we also present thebaseline scenario of a general purpose technology.The differences between the alternativesconsidered are smaller here than in the previous counterfactual exercise.To be sure,thereare some qua

218、litative differences in terms of the effect on output:the long run impactis smaller when AI is a factor-specific technology(especially so for capital-augmentingTFP).But regardless of how AI enters the production function,the long run impact onoutput is positive and considerable.For inflation there a

219、re even less differences across thealternatives,which all show a very similar dynamics.That said,AI as capital-augmentingTFP leads to a smaller long run impact in terms of inflation.051015Years010203040%BaselineLabour-agumenting TFPCapital-augmenting TFP(a)Gross Domestic Product051015Years-1-0.500.5

220、11.52pptBaselineLabour-agumenting TFPCapital-augmenting TFP(b)Inflation(year-on-year)Figure 12:GDP and Inflation6ConclusionConsidering AI as a general purpose technology that improves productivity growth,ouranalysis has shown that AI adoption raises aggregate output,consumption and invest-30ment.The

221、 impact on inflation depends on households and firms anticipation of futureincome growth from AI.In the case of imperfect anticipation,inflation declines in theshort run but eventually increases relative to its initial level.In contrast,when incomeincreases due to the AI-induced rise in productivity

222、 are fully anticipated today,inflationincreases already in the short run and remains elevated in the longer run.Our analysis also delivers important insights on the impact of AI on different indus-tries.For one,a sectors initial exposure to AI is uncorrelated with the ultimate impactof AI on that se

223、ctors output,as indirect effects through spillovers and linkages arewhat matters.However,for aggregate output,which sectors are initially most affected byAI is important.When the effect of AI concentrates on sectors producing consumptionrather than investment goods,output can grow by twice as much.P

224、olicies that fosterthe adoption of AI by firms should thus focus on sectors producing consumption goods.In addition,the adoption of AI could lead to a Goldilocks scenario for monetary policy:greater use of AI could ease inflationary pressures in the near-term,thereby supportingcentral banks in their

225、 task to bring inflation back to target.In the longer term,as inflationrises because of greater AI-induced demand,central banks task of controlling inflationwould become easier,as they can dampen demand via monetary tightening.However,ifhouseholds and firms anticipate the boost to growth from AI,pol

226、icy rates need to riserapidly already today.More research is hence needed to understand households andfirms expectations about gen AI,and how they differ across subgroups of the population(Aldasoro et al.,2024).31ReferencesAcemoglu,Daron and Pascual Restrepo,“Artificial Intelligence,Automation andWo

227、rk,”NBER Working Papers,2018,(24196).,David Autor,Jon Hazell,and Pascual Restrepo,“Artificial Intelligence andJobs:Evidence from Online Vacancies,”Journal of Labor Economics,2022,40(S1),S293S340.Aghion,Philippe,Benjamin F Jones,and Charles I Jones,“Artificial intelli-gence and economic growth,”in“Th

228、e economics of artificial intelligence:An agenda,”University of Chicago Press,2018,pp.237282.Agrawal,Ajay,Joshua Gans,and Avi Goldfarb,The economics of artificial intel-ligence:an agenda,University of Chicago Press,2019.Aldasoro,I naki,Olivier Armantier,Sebastian Doerr,Leonardo Gambacorta,and Tommas

229、o Oliviero,“Survey evidence on gen AI and households:job prospectsamid trust concerns,”BIS Bulletin,2024,86.Araujo,Douglas,Sebastian Doerr,Leonardo Gambacorta,and Bruno Tissot,“Artificial intelligence in central banking,”BIS Bulletin,2024,84.Autor,David,“The Labor Market Impacts of Technological Cha

230、nge:From UnbridledEnthusiasm to Qualified Optimism to Vast Uncertainty,”NBER Working Paper,2022,(w30074).,Frank Levy,and Richard J.Murnane,“The Skill Content of Recent Technolog-ical Change:An Empirical Exploration,”The Quarterly Journal of Economics,2003,118(4),12791333.Babina,Tania,Anastassia Fedy

231、k,Alex He,and James Hodson,“Artificial intel-ligence,firm growth,and product innovation,”Journal of Financial Economics,2024,151,103745.Baily,Martin N.,Erik Brynjolfsson,and Anton Korinek,“Machines of Mind:The Case for an AI-Powered Productivity Boom,”Brookings,May 2023.Baqaee,David and Emmanuel Fah

232、ri,“The Macroeconomic Impact of MicroeconomicShocks:Beyond Hultens Theorem,”Econometrica,2019,87(4),11551203.32Brynjolfsson,Erik and Andrew McAfee,“The Business of Artificial Intelligence,”Harvard Business Review,2017.,Daniel Li,and Luis R.Raymond,“Generative AI at Work,”NBER WorkingPaper,2023,(w311

233、61).,Daniel Rock,and Chad Syverson,“Artificial intelligence and the modern produc-tivity paradox:A clash of expectations and statistics,”in“The economics of artificialintelligence:An agenda,”University of Chicago Press,2018,pp.2357.Cazzaniga,Mauro,Florence Jaumotte,Longji Li,Giovanni Melina,Augustus

234、Panton,Carlo Pizzinelli,Emma Rockall,and Marina M Tavares,“Gen-AI:Artificial Intelligence and the Future of Work,”IMF Staff Discussion Notes,2024,(2024/001).Czarnitzki,Dirk,Gonzalo P.Fern andez,and Christian Rammer,“ArtificialIntelligence and Firm-Level Productivity,”Journal of Economic Behavior&Org

235、ani-zation,2023,(211),188205.Doerr,Sebastian,Leonardo Gambacorta,and Jose M.Serena,“Big Data andMachine Learning in Central Banking,”BIS Working Papers,2021,(930).Evans,Charles L and David A Marshall,“Fundamental economic shocks and themacroeconomy,”Journal of Money,Credit and Banking,2009,41(8),151

236、51555.Felten,Edward,Meera Raj,and Robert Seamans,“The Effect of Artificial Intel-ligence on Human Labor:An Ability-Based Approach,”2019,2019(1).,and,“Occupational,Industry,and Geographic Exposure to Artificial Intelli-gence:A Novel Dataset and Its Potential Use,”Strategic Management Journal,2021,42(

237、12),21952217.Fernald,John and Bing Wang,“The Recent Rise and Fall of Rapid ProductivityGrowth,”FRBSF Economic Letter,2015,(4).Furman,Jason and Robert Seamans,“AI and the Economy,”Innovation policy andthe economy,2019,19(1),161191.Goldman Sachs,“Generative AI:Hype,or Truly Transformative?,”Global Mac

238、ro Re-search Report,2023,(120).33Hulten,Charles R.,“Growth Accounting With Intermediate Inputs,”The Review ofEconomic Studies,1978,pp.511518.Kulish,Mariano and Adrian Pagan,“Estimation and Solution of Models with Ex-pectations and Structural Changes,”Journal of Applied Econometrics,2017,32(2),255274

239、.Lu,Chia-Hui,“The impact of artificial intelligence on economic growth and welfare,”Journal of Macroeconomics,2021,69,103342.Noy,Shiri and Wei Zhang,“Experimental Evidence on the Productivity Effects ofGenerative Artificial Intelligence,”Science,2023,381(6654),187192.Peng,Siyuan,Eirini Kalliamvakou,

240、Peter Cihon,and Mehmet Demirer,“TheImpact of AI on Developer Productivity:Evidence from Github Copilot,”arXivpreprint,2023.Pizzinelli,Carlo,Augustus Panton,Marina Mendes Tavares,Mauro Caz-zaniga,and Longji Li,“Labor Market Exposure to AI:Cross-country Differencesand Distributional Implications,”IMF

241、Working Papers,2023,(2023/2016).Rees,David,“What Comes Next?,”BIS Working Paper,2020,(898).Yang,Chun-Hung,“How Artificial Intelligence Technology Affects Productivity andEmployment:Firm-Level Evidence from Taiwan,”Research Policy,2022,(51.6).34Appendix35AAdditional figures051015Years-3-2-101234pptIn

242、flationYear on year051015Years-3-2-101234%Output gap All industriesConsumer goodsInvestment goodsIntermediate goodsAI boosts TFP in:051015Years-3-2-101234pptPolicy rate Figure 13:Simple model 1051015Years-3-2-101234pptInflationYear on year051015Years-3-2-101234%Output gap All industriesConsumer good

243、sInvestment goodsIntermediate goodsAI boosts TFP in:051015Years-3-2-101234pptPolicy rate Figure 14:Simple model 236BDecomposition of industry value addedThis appendix describes the decomposition of industry value added into the contributionsof final and intermediate demand and income and substitutio

244、n effects described in Section4.2.An industrys gross output can be expressed as:yj,t|zGross output=cj,t+ij,t+gj,t|zFinal demand+FXk=1xk,j,t|zIntermediate demand(B.1)Demand for an industrys output as a final or intermediate good are given by:cj,t=c,j(j,t)Ct(B.2)ij,t=i,j j,tI,t!It(B.3)gj,t=g,j j,tG,t!

245、Gt(B.4)xk,j,t=x,k,j j,tX,k,t!Xk,t(B.5)In the models initial steady state all relative prices are equal to 1.We define final demandincome effects,FDIEtas any change in the final demand for an industrys output fromits initial steady state that is unrelated to changes in relative prices.This is given b

246、y:FDIEt=c,j(Ct C0)+i,j(It I0)+g,j(Gt G0)(B.6)where C0,Iiand G0are aggregate consumption,investment and government expenditurein the models initial steady state.We define intermediate demand income effects,IDIEt,equivalently as:IDIEt=FXk=1k,j(Xk,t Xk,0)(B.7)where Xk,0is the total demand for intermedi

247、ates in industry k in the models initialsteady state.We define final demand substitution effects,FDSEt,as any change in final37demand due to changes in relative prices.Because any change in industry-level demandthat is not due to changes in aggregate demand must be due to price changes,this isgiven

248、by:FDSEt=cj,t+ij,t+gj,t FDIEt(B.8)and equivalently,intermediate demand substitution effects,IDSEtare given by:IDSEt=FXk=1xj,t IDIEt(B.9)The above calculations decompose gross output.To decompose value-added output,wefirst calculate the total value of intermediate inputs used in an industry:xj,t=FXk=

249、1xj,k,t(B.10)Denoting the value added decompositions by the superscript V A,the equivalent calcula-tions are:FDIEV At=FDIEt(1 xj,t/yj,t)(B.11)IDIEV At=IDIEt(1 xj,t/yj,t)(B.12)FDSEV At=FDSEt(1 xj,t/yj,t)(B.13)IDSEV At=IDSEt(1 xj,t/yj,t)(B.14)38 Previous volumes in this series 1178 April 2024 Finterne

250、t:the financial system for the future Agustn Carstens and Nandan Nilekani 1177 March 2024 Pre-publication revisions of bank financial statements:a novel way to monitor banks?Andre Guettler,Mahvish Naeem,Lars Norden and Bernardus F Nazar Van Doornik 1176 March 2024 The effect of Covid pension withdra

251、wals and the Universal Guaranteed Pension on the income of future retirees in Chile Carlos Madeira 1175 March 2024 Unmitigated disasters?Risk-sharing and macroeconomic recovery in a large international panel Goetz von Peter,Sebastian von Dahlen,and Sweta Saxena 1174 March 2024 The impact of informat

252、ion and communication technologies on banks,credit and savings:an examination of Brazil Flavia Alves 1173 March 2024 The macroprudential role of central bank balance sheets Egemen Eren,Timothy Jackson and Giovanni Lombardo 1172 March 2024 Navigating by falling stars:monetary policy with fiscally dri

253、ven natural rates Rodolfo G Campos,Jess Fernndez-Villaverde,Galo Nuo and Peter Paz 1171 March 2024 DeFi Leverage Lioba Heimbach and Wenqian Huang 1170 March 2024 Monetary Policy Transmission in Emerging Makerts:Proverbial Concerns,Novel Evidence Ariadne Checo,Francesco Grigoli,and Damiano Sandri 116

254、9 February 2024 Risk-based pricing in competitive lending markets Carola Mller,Ragnar E.Juelsrud,Henrik Andersen 1168 February 2024 Corporate payout policy:are financial firms different?Emmanuel Caiazzo,Leonardo Gambacorta,Tommaso Oliviero and Hyun Song Shin 1167 February 2024 Monetary Policy with P

255、rofit-Driven Inflation Enisse Kharroubi and Frank Smets 1166 February 2024 Tracing the adoption of digital technologies Vatsala Shreeti 1165 February 2024 The Term Structure of Interest Rates in a Heterogeneous Monetary Union James Costain,Galo Nuo,and Carlos Thomas All volumes are available on our website www.bis.org.