The Economics of Artificial Intelligence

by Ajay Agrawal

  ent two such simulations, to examine: (a) a rise in the productivity of R&D,

  and (b) a rise in automation for middle- skilled tasks ( jobs currently requir-

  ing BA- level workers).

  13.6 Rise in R&D Productivity

  What happens to the structure of an economy when the returns to R&D

  rise because of a new general purpose technology such as transistors and

  342 Jeff rey D. Sachs

  computers in the 1950s or machine learning and artifi cial intelligence in the

  2020s. The experiment, to be precise, is a permanent, one- time step increase

  in

  , the productivity of high- skilled R&D workers. In this fi rst variant,

  R&D

  I assume that only low- skilled workers face the competition from automa-

  tion. In a sense, this illustration tracks the experience of the 1950s– 2010s,

  when the breakthroughs of the digital revolution enabled the automation of

  low- skill tasks. The full model and specifi c parameters are available in the

  supplementary materials. For the purposes here, I emphasize the qualitative

  results.

  In the numerical illustration, the rise in

  occurs in period 5 yet is

  R&D

  anticipated from period 1. Even before the rise in R&D takes hold, workers

  begin to raise their educational attainment in anticipation of the widening

  gap between low- skill and higher- skill wages. After the rise in

  the shift

  R&D

  in educational attainment is even stronger. The end result is a sharp decline

  in the proportion of low- skilled workers and a commensurate rise in middle-

  skilled and high- skilled workers, as shown in fi gure 13.7, which qualitatively

  tracks the same empirical pattern we saw for the US economy in fi gure 13.5.

  Automation initially gives rise to a fall in wages for unskilled workers,

  and a rise in wages for the intermediate and high- skilled sectors. The wage

  gap between high- skilled and low- skilled workers therefore opens, but then

  leads to the shift in educational attainment in fi gure 13.7, thereby tending

  to restore the preshock relative wages across skill levels.

  In the second simulation, the rise in

  for low- skill tasks (again start-

  R&D

  ing in period 5) is now accompanied by a similar rise in R&D productiv-

  ity for automation in intermediate- skill tasks (starting in period 10). Thus,

  automation replaces both low- skilled and intermediate- skilled workers. The

  Fig. 13.7 Labor by educational attainment

  Source: See appendixes A, B, and C.

  R&D, Structural Transformation, and the Distribution of Income 343

  Fig. 13.8 Labor by educational attainment: automation for low- skill and

  intermediate- skill tasks

  Source: See appendixes A, B, and C.

  result, of course, is to give an added boost to the attainment of advanced

  degrees, so that both L and L decline, while L rises. The pattern is shown U

  I

  H

  in fi gure 13.8, which may usefully be compared with fi gure 13.7.

  In the case of automation of both unskilled and intermediate- skill tasks,

  the main result is that market forces induce those receiving a bachelor’s

  degree to continue on to an advanced degree. The labor market ends up

  with just two kinds of labor, unskilled and highly skilled, with intermediate-

  skilled workers disappearing from the scene. Note that the model so far

  assumes that all workers are equally endowed with the skills needed for

  all levels of education; there is no “scarcity” value of STEM skills, for ex-

  ample, that would limit the supply of high- skilled workers. In a more real-

  istic model, we would grapple with the obvious fact that not all students have

  the aptitude for an advanced degree for high- skill work. Instead of the wage

  diff erentials being off set by highly elastic shifts in educational attainment, a

  premium on higher education would be sustained in the long run as a kind

  of natural rent on high educational aptitude.

  In both scenarios, the labor share of GDP declines markedly, as jobs are

  lost to automation. Figure 13.9 shows the time path of the labor share of

  GDP in the second scenario, in which automation for low- skilled workers

  takes off after period 5, and for intermediate- skilled workers after period

  10. The labor share of income begins to dip around period 5, but then soars

  again around period 10 as the wages of skilled workers increases. Over time,

  as workers raise their educational attainment, wages decline and the overall

  labor share of income falls sharply under the pressures of automation.

  344 Jeff rey D. Sachs

  Fig. 13.9 Labor share of GDP

  Source: See appendixes A, B, and C.

  13.7

  Next Steps

  So far, the conclusions of the simulations are wholly qualitative. The next

  steps in modeling will be to parameterize the model according to the main

  structural features of the US economy. Of course, there are many diffi

  cult

  modeling and conceptual choices ahead, both in validating a parametrized

  model according to recent history and using the model to project the impli-

  cations of future technological changes. Some of the diffi

  culties are the fol-

  lowing:

  1. modeling the automation process with empirical detail, for example,

  by identifying the classes of machines that are complementarity to versus

  substitutional with various skills and occupations;

  2. estimating the returns to automation- inducing R&D, and the implica-

  tions for the earnings of advanced technical workers;

  3. characterizing the supply and demand for higher education as a func-

  tion of wage diff erential, borrowing costs, and educational aptitudes;

  4. characterizing the relative roles of private and public fi nancing in deter-

  mining the investments in R&D and in education;

  5. creating realistic scenarios for the future evolution of smart machines

  and their interaction with occupations at various skill levels;

  6. modeling the intergenerational dynamics of automation as in Sachs

  and Kotlikoff (2012) and Benzell, Kotlikoff , LaGarda, and Sachs (2015);

  7. accounting for monopoly rents on patents and other changes in market

  structure associated with smart machines and artifi cial intelligence;

  R&D, Structural Transformation, and the Distribution of Income 345

  8. accounting for the income distributional consequences of big data and

  network externalities, for example, for giants such as Google and Amazon;

  9. accounting for the distributional implication of dematerialized pro-

  duction (ecommerce, ebooks, epayments) and the sharing economy (e.g.,

  vehicles on demand); and

  10. modeling the changes in past and future labor force participation

  and leisure time as the result of smart machines, artifi cial intelligence, and

  automation.

  Appendix A

  GAMS Equations

  Kf ( tf ). . . K( tf ) = e = K0;

  Hf ( tf ). . . H( tf ) = e = H0;

  Uf ( tf ). . . U( tf ) = e = U0;

  Sf ( tf ). . . S( tf ) = e = S0;

  IPPA f ( tf ). . .IPPA( tf ) = e = IPPA0;

  IPPAIf( tf ). . .IPPAI( tf ) = e = IPPAI0;


  Output( t). . . Q( t) = e = TA( t)**Alpha* M( t)**(1-Alpha); BAprod( t). . .BA( t) = e = MBA( t)**.2*SBA( t)**.2*HBA( t)**.6; PROFprod( t). . .PROF( t) = e = MPROF( t)**.2*ProdPROF( t)*

  HPROF( t)**.8;

  *PROFprod( t). . .PROF( t) = e = ProdPROF( t)*HProf( t); Health( t). . .HL( t) = e = MHL( t)**.2*LUHL( t)**.1*SHL( t)**

  .2*HHL( t)**.5;

  *HealthD( t). . .HL( t) = e = HLmin*IPP( t)**.2;

  HealthD( t). . .HL( t) = e = .01;

  Capital( t). . . K( t) = e = M( t) + MBA( t) + MPROF( t) + MHL( t) + RA( t)

  + RAI( t);

  Task( t). . .TA( t) = e = ( LU( t) + A( t))**Beta*( LS( t) + AI( t))**(1-Beta); Robot( t). . . A( t) = e = ThetaA( t)*HA( t)**Gamma*IPPA(t)**

  Delta*RA( t)**(1-Gamma- Delta);

  ArtInt( t). . .AI( t) = e = ThetaAI( t)*HAI( t)**Gamma*IPPAI( t)**

  Delta*RAI( t)**(1-Gamma- Delta);

  RDA( t + 1). . .IPPA( t + 1) = e = IPPA( t)*(1-depRD) +

  PRODRDA( t)*HRD( t);

  RDAI( t + 1). . .IPPAI( t + 1) = e = IPPAI( t)*(1-depRD) +

  PRODRDAI( t)*HRD( t);

  HighS( t). . . H( t) = e = HAI( t) + HA( t) + HRD( t) + HBA( t) + HPROF( t)

  + HHL( t);

  346 Jeff rey D. Sachs

  KNext( t + 1). . . K( t + 1) = e = K( t)*(1-dep) + FINV( t); Saving( t). . . C( t) = e = Q( t)—FINV( t) ; UNext( t + 1). . . U( t + 1) = e = U( t)*(1- n)—BA( t) + n*( U( t) + S( t) + H( t)) ; SNext( t + 1). . . S( t + 1) = e = S( t)*(1- n) + BA( t)—PROF( t); HNext( t + 1). . . H( t + 1) = e = H( t)*(1- n) + PROF( t); LaborU( t). . . U( t) = e = LU( t) + BA( t) + LUHL( t); LaborS( t). . . S( t) = e = LS( t) + 0.2*PROF( t) + SBA( t) + SHL( t); Utils( t). . .Ut( t) = e = log( C( t));

  KLast( tl ). . .KL( tl ) = e = K( tl )*(1-dep)+ FINV( tl ); CLast( tl ). . .CL( tl ) = e = Q( tl )—dep*KL( tl ); Utility. . .Util = e = sum( t,disc( t)* Ut( t)) + sum( tl,disc( tl )*log(CL( tl ))/

  Discrate);

  * Output

  Parameter WageU( t), WageS( t), WageH( t), Rrate( t), IPPArate( t), IPPAIrate( t), Lshare( t), Kshare( t), HAshare( t), RArate( t), Income( t), Lshare( t), LUshare( t), LSshare( t), LHshare( t);

  Parameter Kshare( t), IPshare(t), LULF( t), LSLF( t), LHLF( t), LF( t); WageU( t) = Alpha*Q.L( t)/ TA.L( t) * Beta * TA.L( t)/ (LU.L( t) + A.L( t)); WageS( t) = Alpha*Q.L( t)/ TA.L( t) * (1-Beta) * TA.L( t)/ (LS.L( t) +

  AI.L( t));

  Rrate( t) = (1-Alpha)*Q.L( t)/ M.L( t) ;

  WageH( t) = ThetaA( t)*Gamma*(A.L( t)/ HA.L( t))*WageU( t); HAshare( t) = HA.L(t)/ H.L( t);

  RArate( t) = (1-Gamma- Delta)*(A.L( t)/ RA.L( t))*WageU( t); IPPArate( t) = Gamma*(A.L( t)/ IPPA.L( t))*WageU( t);

  IPPAIrate( t) = Gamma*(AI.L( t)/ IPPAI.L( t))*WageS( t);

  Income( t) = WageU( t)*LU.L( t) + WageS( t)*LS.L( t) + WageH( t)*H.L( t)

  + Rrate( t)*K.L( t) + IPPArate( t)*IPPA.L( t) + IPPAIrate( t)*IPPAI.L( t); Lshare( t) = (WageU( t)*LU.L( t) + WageS( t)*S.l(t) + WageH( t)*H.L( t)) /

  Income( t);

  LUshare( t) = WageU( t)*LU.L( t)/ Income( t);

  LSshare( t) = WageS( t)*LS.L( t)/ Income( t);

  LHshare( t) = WageH( t)*H.L( t)/ Income( t);

  Kshare( t) = Rrate( t)*K.L( t)/ Income( t);

  IPshare( t) = (IPPArate( t)*IPPA.L( t) + IPPAIrate( t)*IPPAI.L( t))/

  Income( t);

  LF( t) = LU.L( t) + LS.L( t) + H.L( t);

  LULF( t) = LU.L( t) / LF( t);

  LSLF( t) = LS.L( t) / LF( t);

  LHLF( t) = H.L( t) / LF( t);

  R&D, Structural Transformation, and the Distribution of Income 347

  Appendix B

  Parameter Values

  Parameters Gamma, Alpha, Beta, Delta, Disc( t), dep, depRD, HLmin,

  Discrate;

  Gamma = .5;

  Alpha = .7;

  Beta = .7;

  Gamma = .3;

  Delta = .3;

  Discrate = .06;

  Disc(t) = (1/ (1+Discrate))**(ord( t)- 1);

  dep = 0.05;

  depRD = .05;

  HLmin = .1;

  Parameters ThetaA, ThetaAI, tfpRA( t), tfpRAI( t);

  ThetaA( t) = 1;

  ThetaAI( t) = 1;

  *tfpRA( t) = .01;

  *tfpRA( t)$(ord( t) ge 10) = 1;

  *tfpRAI(t) = .01;

  *tfpRAI(t)$(ord( t) ge 15) = 1;

  tfpRA( t) = 1;

  tfpRAI( t) = 1;

  Appendix C

  Initial Values

  Parameter K0, U0, S0, H0, ProdRDA( t), ProdRDAI( t), ProdPROF( t), IPPA0, IPPAI0, n, Start( t);

  K0 = 21.9;

  U0 = 7.3;

  S0 = 2.25;

  H0 = 0.15;

  ProdRDA( t) = .01;

  ProdRDAI( t) = .01;

  *ProdRDAI( t)$(ord( t) ge 10) = 1;

  ProdPROF( t) = 2;

  IPPA0 = 0.001;

  348 Jeff rey D. Sachs

  IPPAI0 = 0.001;

  n = 0.05;

  References

  Acemoglu, Daron, and David Autor. 2011. “Skills, Tasks and Technologies: Implica-

  tions for Employment and Earnings.” In Handbook of Labor Economics, vol. 4b,

  edited by Orley Ashenfelter and David E. Card. Amsterdam: Elsevier.

  Benzell, Seth G., Laurence J. Kotlikoff , Guillermo LaGarda, and Jeff rey D. Sachs.

  2015. “Robots Are Us: Some Economics of Human Replacement.” NBER Work-

  ing Paper no. 20941, Cambridge, MA.

  Chui, Michael, James Manyika, and Mehdi Miremadi. 2016. “Where Machines

  Could Replace Humans and Where They Can’t (Yet).” McKinsey Quarterly

  July 2016. https:// www .mckinsey .com/ business- functions/ digital- mckinsey/ our

  - insights/ where- machines- could- replace- humans- and- where- they- cant- yet.

  Elsby, Michael W. L., Bart Hobijn, and Ays¸egül S¸ahin. 2013. “The Decline of the

  U.S. Labor Share.” Federal Reserve Bank of San Francisco Working Paper Series

  no. 2013– 27, Federal Reserve Bank of San Francisco, September.

  Frey, Carl Benedikt, and Michael A. Osborne. 2013. “The Future of Employment:

  How Susceptible are Jobs to Computerization.” Working Paper, Oxford Martin

  School, University of Oxford. September. https:// www .oxfordmartin.ox.ac .uk

  / downloads/ academic/ The_Future_of_Employment .pdf.

  International Labor Organization (ILO) and Organisation for Economic Co-

  operation and Development (OECD). 2015. “The Labor Share in the G20 Econo-

  mies.” G20 Employment Working Group, February.

  Kaldor, Nicholas. 1957. “A Model of Economic Growth.” Economic Journal 67

  (268): 591–624.

  Karabarbounis, Loukas, and Brent Neiman. 2013. “The Global Decline of the Labor

  Share.” NBER Working Paper no. 19136, Cambridge, MA.

  Koh, Dongya, Raul Santaeulalia- Llopis, and Yu Zheng. 2015. “Labor Share Decline

  and the Capitalization of Intellectual Property Products.” Working Paper, January.

  McKinsey Global Institute. 2017. “A Future that Works: Automation, Employ-

  ment, and Productivity.” Report, January. https:// www .mckinsey .com/ mgi/ over

  view/ 2017-in-review/ automation- and- the- future- of-work/ a- future- that- works

  - automation- employment- and- productivity.

  Sachs, Jeff rey D., Seth G. Benzell, and Guillermo LaGarda. 2015. “Robots: Curse

  or Blessing? A Basic Framework.” NBER Working Paper no. 21091, Cambridge,

  MA.

  Sachs, Jeff rey D., and Laurence J. Kotlikoff . 2012. “Smart Machines and Long- Term

  Misery.” NBER Working Paper no. 18629, Cambridge, MA.

  14

  Artifi cial Intelligence and Its

  Implications for Income

  Distribution and Unemployment

  Anton Korinek and Joseph E. Stiglitz

  14.1 Introduction

  The introduction of artifi cial intelligence (
AI) is the continuation of a

  long process of automation. Advances in mechanization in the late nine-

  teenth and early twentieth centuries automated much of the physical labor

  performed by humans. Advances in information technology in the mid- to

  late twentieth century automated much of the standardized data processing

  that used to be performed by humans. However, each of these past episodes

  of automation left large areas of work that could only be performed by

  humans.

  Some propose that advances in AI are merely the latest wave in this long

  process of automation, and may in fact generate less economic growth than

  past technological advances (see, e.g., Gordon 2016). Others, by contrast,

  emphasize that AI critically diff ers from past inventions: as artifi cial intelli-

  gence draws closer and closer to human general intelligence, much of human

  labor runs the risk of becoming obsolete and being replaced by AI in all

  domains. In this view, progress in artifi cial intelligence is not only a continua-

  Anton Korinek is associate professor of economics and business administration at the

  University of Virginia and Darden GSB and a research associate of the National Bureau of Economic Research. Joseph E. Stiglitz is University Professor at Columbia University and a research associate of the National Bureau of Economic Research.

  This chapter was prepared as a background paper for the NBER conference The Economics

  of Artifi cial Intelligence. We would like to thank our discussant Tyler Cowan as well as Jayant Ray and participants at the NBER conference for helpful comments. We also acknowledge

  research assistance from Haaris Mateen as well as fi nancial support from the Institute for New Economic Thinking (INET) and the Rewriting the Rules project at the Roosevelt Institute, supported by the Ford, Open Society, and the Bernard and Irene Schwartz Foundations. For acknowledgments, sources of research support, and disclosure of the authors’ material fi nancial relationships, if any, please see http:// www .nber .org/ chapters/ c14018.ack.

  349

  350 Anton Korinek and Joseph E. Stiglitz

  tion, but the culmination of technological progress; it could lead to a course

  of history that is markedly diff erent from the implications of previous waves

  of innovation, and may even represent what James Barrat (2013) has termed

  “Our Final Invention.”

  No matter what the long- run implications of AI are, it is clear that it has

  the potential to disrupt labor markets in a major way, even in the short and