Artificial intelligence (AI) is remodeling what machines can do, from processing pure language to analysing complicated datasets and producing photos. Latest advances in generative AI (as an illustration, massive language fashions similar to ChatGPT) are additionally animating a vigorous debate about the potential for big productivity gains that might permit economies to flee the disappointing productivity development of the previous twenty years in lots of OECD nations (Goldin et al. 2024, Winker et al. 2021, Andre and Gal 2024).
Opinions on this debate fluctuate broadly (Determine 1). Some view AI as a transformative general-purpose know-how that would unleash productivity development throughout a variety of financial actions and ship massive macroeconomic productivity gains over the subsequent decade (Baily et al. 2023). Others argue that present AI know-how just isn’t notably helpful in most financial actions and predict that the mixture productivity gains from AI will likely be modest (Acemoglu 2024). Our new paper (Filippucci et al. 2024) contributes to this debate by assessing the mixture productivity gains from AI beneath totally different situations for sectoral productivity development and by discussing the position of sectoral reallocation.
Determine 1 Divergent views about the mixture productivity gains from AI
Predicted improve in annual labour productivity development over a 10-year horizon on account of AI (in proportion factors)
Notes: When the supply presents a variety of estimates as the predominant outcome, the decrease and higher bounds are indicated by striped areas. In instances the place predictions are made for complete issue productivity, predicted labour productivity gains are obtained by assuming a normal long-run multiplier of 1.5 concerning the adjustment of the capital inventory (Acemoglu 2024, Aghion and Bunel 2024, Bergeaud 2024 and OECD). The estimates check with the nations proven in brackets.
Sources: See references at the finish of the paper; for Goldman Sachs (2023), the underlying reference is Briggs and Kodnani (2023); for IMF (2024) the underlying reference is Cazzaniga et al. (2024); for OECD, the vary from Filippucci et al. (2024) predominant situations are proven (Desk 2 final row in Part 3.1).
Sources of disagreement concerning the mixture productivity gains from AI
A rising physique of analysis paperwork that AI can considerably improve the efficiency of employees in particular enterprise contexts, similar to customer support (by 14%), enterprise consulting (by 40%), or software program growth (by greater than 50%) (see Filippucci et al. 2024a and 2024b for a evaluation of current research on the worker-level productivity impacts of AI). Given the mounting proof of considerable productivity gains in particular domains, it could be stunning that opinions about the mixture productivity advantages of AI stay so diverse. Nonetheless, predicting mixture gains by extrapolating from proof on the affect of AI in particular components of the financial system is difficult. The economy-wide affect of AI will depend upon how broadly AI might be adopted to enhance the manufacturing processes throughout many components of the financial system – sometimes called ‘publicity’ to AI – and on how quickly companies will undertake AI.
As well as, mixture productivity development additionally is dependent upon the relative demand for the items and providers produced in several sectors of the financial system. Particularly, a Baumol impact (Baumol 1967, Nordhaus 2008) can come up on the whole equilibrium if productivity gains from AI are concentrated in a number of sectors and relative sectoral demand reacts little to relative worth modifications. On this case, sectors the place AI-driven productivity gains are low (e.g. development, agriculture, and private providers) might develop as a share of GDP. Combination development may become restricted “not by what we do nicely however somewhat by what is important and but exhausting to enhance” (Aghion et al. 2019).
We assess the macroeconomic productivity gains from AI beneath totally different situations for publicity to AI, the velocity of AI adoption, and drivers of Baumol’s development illness. In our predominant situations, we undertaking that AI may contribute between 0.25 and 0.6 proportion factors to annual complete issue productivity development in the US (or between 0.4 and 0.9 proportion factors to annual labour productivity development, assuming a normal long-run multiplier of 1.5 concerning the adjustment of the capital inventory) over the subsequent decade. Estimates for different economies are of comparable magnitude, although considerably decrease, provided that adoption of AI is predicted to be slower and extremely AI-exposed sectors are comparatively smaller in these economies.
These predictions, in the event that they certainly materialise, suggest a considerable contribution to labour productivity in the context of weak productivity development throughout the OECD over the previous many years, which has been in the vary of 1%–1.5% per yr. The higher finish of our estimates suggests a productivity acquire from AI that’s of equally massive magnitude as what has been attributed to ICT in the US throughout the high-growth decade beginning in the mid-90s (round 1% per yr; see Byrne et al. 2013 and Bunel et al. 2024).
From micro to macro
To derive projections for macroeconomic productivity development, we proceed in two steps. First, impressed by Acemoglu (2024), we receive sectoral productivity gains by combining estimates of worker-level efficiency gains with measures of sectoral publicity to AI (Determine 2) and projections of future adoption charges primarily based on the historic expertise with earlier general-purpose applied sciences (Determine 3). The ensuing ten-year sectoral gains in complete issue productivity vary from 1–2% in manual-intensive actions (agriculture, fishing, mining) to fifteen–20% in knowledge-intensive providers (ICT, finance, skilled providers), relying on the particular assumptions on AI adoption and publicity.
Determine 2 Publicity to AI varies throughout sectors
Determine 3 Totally different situations for the adoption path of AI
In the second step, we derive the implied macroeconomic productivity gains utilizing a calibrated multisector general-equilibrium mannequin that accounts for sectoral input-output linkages and the position of demand in driving worth changes and issue reallocation throughout sectors (Baqaee and Farhi 2019). Macroeconomic productivity gains are derived beneath totally different situations concerning the magnitude of micro-level productivity gains, sectoral publicity to AI, the velocity of adoption, and structural determinants of sectoral reallocation (Determine 4). The combination productivity gains from AI might be decomposed into three results: (1) a direct impact of elevated productivity at the sectoral degree; (2) an input-output multiplier impact as productivity gains in a single sector additionally profit different sectors by lowered prices of intermediate inputs; and (3) a Baumol impact.
Determine 4 Macro-level productivity gains from AI beneath totally different situations
Estimated affect on annual development charges of complete issue productivity over a 10-year horizon
Notes: The bars correspond to totally different situations concerning the adoption, capabilities, and micro-level gains of AI (as in Determine 1). In situations 1 and a pair of, the elasticity of substitution between the output of various sectors is shut to 1, and the components of manufacturing (labour and capital) can reallocate freely throughout sectors. In situations 3–5 with adjustment frictions, the elasticity in consumption is assumed to be very low, and components can’t reallocate throughout sectors. See extra particulars in part 3 of Filippucci et al. (2024).
AI adoption is a key driver of productivity development, however uneven sectoral gains may restrict mixture development by a Baumol impact
A key perception that emerges from this evaluation is that the macroeconomic affect of AI will rely totally on the adoption velocity and the diploma to which AI can profit financial actions throughout a variety of sectors in the financial system. At the moment, adoption varies strongly throughout companies and sectors, with country-level adoption charges being typically low, in the vary of 5%–15%, as reported by official statistics of companies and firm-level research (e.g. Calvino and Fontanelli 2023a, 2023b). A comparability of situations 1 (low adoption) and a pair of (excessive adoption and expanded capabilities) exhibits that quick and productive integration of AI in a wider vary of financial actions by expanded AI capabilities (e.g. additional integration with different digital instruments) is critical for the emergence of enormous macroeconomic gains.
A damaging Baumol impact on mixture productivity development arises if the productivity advantages of AI are concentrated in a number of sectors, as in situation 3 (excessive adoption and expanded capabilities, plus uneven sectoral gains and adjustment frictions), the place sectoral gains are extra uneven as a result of knowledge-intensive sectors similar to ICT and finance are assumed to undertake AI extra shortly.
Productivity gains in the earlier technology-driven growth (throughout the ICT growth decade beginning in the mid-90s) had been concentrated in a number of sectors. On this spirit, situation 4 (very massive gains, concentrated in most uncovered sectors, plus adjustment frictions) considers a focus of sectoral gains which are nearer to what was noticed throughout that interval.
Right here, the Baumol impact reduces mixture productivity gains by a 3rd.
In distinction, no Baumol impact arises if AI gains are extra widespread throughout sectors, as an illustration if AI is healthier built-in with robotics know-how, which might imply that not solely cognitive but additionally manual-intensive actions may benefit from AI (situation 5, AI mixed with robotics know-how, plus adjustment frictions).
We additionally discover how mixture productivity results may depend upon the presence of frictions by their affect on modifications in the sectoral composition of the financial system. Particularly, we think about the chance that components of manufacturing (capital and labour) can’t be freely reallocated throughout sectors over our projection horizon. We present that such frictions may amplify the damaging Baumol impact by requiring steeper declines in the relative output costs of AI-boosted sectors to create sufficient demand for his or her elevated output. This is able to result in a bigger decline of their GDP share, particularly if demand is inelastic.
Therefore, regardless that such frictions would forestall the reallocation of things from high- to low-growth sectors, a normal equilibrium perspective clarifies that mixture productivity development would nonetheless be harmed by stopping the environment friendly allocation of things in the direction of sectors the place they’re most valued.
Total, AI holds promise to revitalise productivity development in OECD nations and past. Governments also can play a job in shaping the macroeconomic gains from AI, for instance by resolving authorized uncertainties round accountability, which can maintain again productive AI adoption by companies (OECD 2024a). At the similar time, governments can foster a aggressive setting (each in the AI-using in addition to the AI-producing sectors; see Aghion and Bunel 2024, OECD 2024b) that’s conducive to innovation and experimentation, whereas monitoring potential labour market disruptions and supporting employees as they transition into new roles in the AI financial system (e.g. Acemoglu et al. 2023a,b, Baily et al. 2023, OECD 2023).
References
Acemoglu, D (2024), “The easy macroeconomics of artificial intelligence”, Financial Coverage, eiae042.
Acemoglu, D, D Autor, and S Johnson (2023a), “Can now we have pro-worker AI? Selecting a path of machines in service of minds”, MIT Shaping the Way forward for Work Initiative, coverage memo.
Acemoglu, D, D Autor, and S Johnson (2023b), “How AI can become pro-worker”, VoxEU.org, 4 October.
Aghion, P, and S Bunel (2024), “AI and development: The place can we stand?”.
Aghion, P, B Jones, and C Jones (2019), “Artificial intelligence and financial development”, in The Economics of Artificial Intelligence: An Agenda, College of Chicago Press.
Andre, C, and P Gal (2024), “Reviving productivity development: A evaluation of insurance policies”, OECD Economics Division Coverage Paper No. 1822.
Baily, M, E Brynjolfsson, and A Korinek (2023), “Machines of thoughts: The case for an AI-powered productivity growth”, Brookings Establishment, 10 Could.
Baqaee, D R, and E Farhi (2019), “The macroeconomic affect of microeconomic shocks: Past Hulten’s theorem”, Econometrica 87(4): 1155–203.
Baumol, W J (1967), “Macroeconomics of unbalanced development: The anatomy of city disaster?”, The American Financial Overview 57(3): 415–26.
Bergeaud, A (2024), “Financial coverage in an period of transformation”, ECB Discussion board on Central Banking.
Bunel, S, G Bijnens, V Botelho, E Falck, V Labhard, A Lamo, O Röhe, J Schroth, R Sellner, J Strobel, and B Anghel (2024), “Digitalisation and productivity”, Occasional Paper Sequence 339, European Central Financial institution.
Byrne, D M, S D Oliner, and D E Sichel (2013), “Is the info know-how revolution over?”, Worldwide Productivity Monitor 25: 20–36.
Hatzius, J, J Briggs, D Kodnani, and G Pierdomenico (2023) “The possibly massive results of artificial intelligence on financial development (Briggs/Kodnani)”, International Economics Analyst, Goldman Sachs.
Calvino, F, and L Fontanelli (2023a), “A portrait of AI adopters throughout nations: Agency traits, property’ complementarities and productivity”, OECD Science, Know-how and Business Working Papers No. 2023/02.
Calvino, F, and L Fontanelli (2023b), “Firms’ use of artificial intelligence: Cross-country evidence on business characteristics, asset complementarities, and productivity”, VoxEU.org, 14 June.
Cazzaniga, M, F Jaumotte, L Li, G Melina, A J Panton, C Pizzinelli, E J Rockall, and M Mendes Tavares (2024), “Gen-AI: Artificial intelligence and the future of labor”, IMF Workers Dialogue Observe SDN2024/001.
Filippucci, F, P Gal, C Jona-Lasinio, A Leandro, and G Nicoletti (2024a), “The affect of artificial intelligence on productivity, distribution and development: Key mechanisms, preliminary proof and coverage challenges”, OECD Artificial Intelligence Papers No. 15.
Filippucci, F, P Gal, C Jona-Lasinio, A Leandro, and G Nicoletti (2024b), “Should AI stay or should AI go: The promises and perils of AI for productivity and growth”, VoxEU.org, 2 Could.
Goldin, Ian, P Koutroumpis, F Lafond, and J Winkler (2024), “Why is productivity slowing down?”, Journal of Financial Literature 62(1): 196–268.
Nordhaus, W D (2008), “Baumol’s ailments: A macroeconomic perspective”, The BE Journal of Macroeconomics 8(1).
OECD (2023), OECD Employment outlook 2023: Artificial intelligence and the labour market, Paris: OECD Publishing.
OECD (2024a), “Suggestion of the Council on Artificial Intelligence”, OECD Authorized Devices, OECD/LEGAL/0449.
OECD (2024b), “Artificial intelligence, information and competitors”, OECD Artificial Intelligence Papers No. 18.
Winkler, J, P Koutroumpis, F Lafond, and I Goldin (2021), “Re-evaluating the sources of the recent productivity slowdown”, VoxEU.org, 31 Could.