The Impact of Artificial Intelligence on Productivity and Economic Growth: A Meta-Overview of the Post-2017 Evidence

I. Introduction

Few questions in contemporary economics have moved as quickly from speculation to measurement as the effect of artificial intelligence on productivity. The public release of large language model (LLM) assistants in late 2022 gave firms and researchers a discrete, datable technology shock, and within two years a substantial body of experimental and quasi-experimental evidence had accumulated. At the same time, a parallel literature of macroeconomic assessments—some produced by academic economists, some by financial institutions—has attempted to translate task-level automation potential into projections of aggregate growth. The two literatures reach conclusions that appear, at first glance, sharply inconsistent: individual studies routinely report productivity gains of tens of percent, while sober macroeconomic calibrations imply aggregate total factor productivity (TFP) contributions of well under one percentage point per year.

This paper does not resolve that tension by re-estimating the underlying micro-data. Instead, it offers what the literature currently lacks: a structured, transparent meta-overview that (i) assembles a corpus of studies through systematic database retrieval, (ii) codes each study on a common set of design and effect-size dimensions, and (iii) synthesizes the resulting evidence separately for the micro and macro tiers, making explicit why the two are not directly comparable. Our aim is closer to a methods-and-findings map than to a formal meta-analysis: because reported effect sizes are heterogeneous in outcome definition and only inconsistently accompanied by standard errors, we favor descriptive distributions and vote-counting over a pooled random-effects estimate, and we are explicit about that choice.

Three findings emerge. First, the micro evidence is remarkably consistent in sign: across 14 studies reporting a quantifiable productivity effect, every estimate is positive, with a median of about 30 percent on the primary metric. Second, the micro evidence is equally consistent in a distributional pattern first documented by Brynjolfsson, Li, and Raymond (2025): gains are largest for less-experienced and lower-skilled workers, compressing the productivity distribution within occupations. Third, the macro evidence is characterized not by consensus but by dispersion: the ratio between the most conservative and most optimistic published annual-productivity projection exceeds twenty to one, and the disagreement is traceable to a small number of modeling choices—the share of tasks exposed, the cost of automating them, and whether displaced labor is redeployed productively.

The remainder of the paper proceeds as follows. Section II situates the current debate in the longer arc from the Solow productivity paradox to the productivity J-curve. Section III describes the corpus, retrieval strategy, and coding scheme. Section IV synthesizes the micro-experimental evidence; Section V the macro assessments. Section VI discusses reconciliation, Section VII treats heterogeneity, diffusion, and measurement, and Section VIII lays out open questions. Section IX concludes.

II. From the Solow Paradox to the Productivity J-Curve

The intellectual backdrop to the AI-productivity debate is Robert Solow’s 1987 observation that the computer age was visible everywhere except in the productivity statistics. The “Solow paradox” framed a generation of research on information technology and growth, and its central lesson—that general-purpose technologies (GPTs) diffuse slowly and require large complementary investments before their effects register in measured output—remains the most useful lens for the present episode.

Brynjolfsson, Rock, and Syverson (2021) formalize this lesson as a productivity J-curve. When a GPT arrives, firms must accumulate intangible complementary capital—new processes, worker skills, organizational redesign—that is largely uncounted in national accounts. During the build-out phase, resources are diverted from measured output toward this unmeasured investment, so measured productivity growth understates true productivity growth and may even dip. Only once the intangible stock matures do the gains surface, producing a delayed surge. The J-curve reconciles two facts that otherwise sit uneasily together: transformative micro-level potential and disappointing aggregate statistics.

This framing carries two implications for reading the current evidence. First, early aggregate productivity data will tend to under-record AI’s ultimate contribution, because the economy is still in the intangible-investment phase. Second, the micro field experiments—which measure output on well-defined tasks where the complementary capital (a working tool, a trained user) is already in place—capture something closer to the technology’s frontier potential than to its economy-wide average effect. The gap between the two tiers is therefore partly a gap between frontier and average, and partly a gap between a mature-complementarity setting and an economy mid-transition.

III. Data and Coding of Studies

C. Coding scheme

Each study was coded on: (i) strand—micro-experimental, micro-observational, macro-projection, macro-observational, theory/foundational, or review; (ii) design—e.g., randomized controlled trial (RCT), field experiment, difference-in-differences, task-based growth-accounting model; (iii) setting—occupation, sector, country, tool; (iv) outcome—throughput, time-to-complete, quality, TFP, GDP; and (v) effect size, recorded as reported, with sign. Figure 1 shows the composition of the corpus. The evidence concentrates in three strands of comparable size—micro-experimental (17), micro-observational (17), and theory/foundational (19)—with 14 macro-projection studies.

Key numbers, generated from evidence_table.csv:

Corpus	69 studies (61 retrieved + 8 landmark supplements)
Strands	17 micro-experimental · 17 micro-observational · 14 macro-projection · 1 macro-observational · 19 theory-foundational · 1 review
Micro quantified effects	n = 14, median +29.96%, mean +30.67%, IQR 12.98–38.5%, range +2.4% to +87%, 14/14 positive
Macro projections (annual pp)	0.05 to 1.5 across 3 quantified sources — a ratio above 30:1
Provenance	61 retrieved · 8 verified · 0 author-supplied

D. The broader literature landscape

The coded corpus of this review is deliberately narrow: economics studies that estimate or project an effect of AI or automation on productivity, output, TFP, GDP, or growth. That corpus sits inside a much larger research landscape. As supplementary context, we track a broader collection of 884 papers on AI, work, and the economy, classified into seven themes spanning from the pre-generative-AI automation literature to post-ChatGPT occupational evidence. Two facts from that broader landscape are worth keeping in view when reading the coded evidence.

First, the composition of the field has shifted sharply since the release of ChatGPT. Themes centered on firm adoption of earlier AI waves and on wage inequality have shrunk as shares of new output, while task-level exposure measurement and the study of generative AI’s effect on work and skills have grown to dominate the inflow (Figure 3). The coded corpus of this review—concentrated in generative-AI experiments and task-based projections—reflects where the field’s marginal effort now goes.

Second, the most-cited works in the broader landscape remain the pre-generative automation classics (Autor’s history of workplace automation, Acemoglu and Restrepo’s race between man and machine, the early robots literature), a reminder that today’s generative-AI evidence extends a two-decade research program rather than starting one.

Work	Year	Citations	Venue
Why Are There Still So Many Jobs? The History and Future of Workplace Automation	2015	3,535	The Journal of Economic Perspectives
The Race between Man and Machine: Implications of Technology for Growth, Factor Shares,…	2018	2,820	American Economic Review
The forthcoming Artificial Intelligence (AI) revolution: Its impact on society and firms	2017	1,808	Futures
A future that works: automation, employment, and productivity	2017	907
Robots and Jobs: Evidence from US Labor Markets	2017	903	National Bureau of Economic Research
Artificial Intelligence and Jobs: Evidence from Online Vacancies	2022	894	Journal of Labor Economics
The Risk of Automation for Jobs in OECD Countries	2016	838	OECD social employment and migration working papers
Hotel employee’s artificial intelligence and robotics awareness and its impact on turno…	2019	770	Tourism Management
Toward understanding the impact of artificial intelligence on labor	2019	665	Proceedings of the National Academy of Sciences
The wrong kind of AI? Artificial intelligence and the future of labour demand	2019	569	Cambridge Journal of Regions Economy and Society

Notable contributing authors in the landscape collection, with works in the collection: Daron Acemoglu (12), Erik Brynjolfsson (4), Anton Korinek (4), Ronnie Chatterji (3), Alex Imas (3), Luis Garicano (2), Chad Jones (1), Tom Cunningham (1), Susan Athey (1), Simon Johnson (1).

IV. Micro Evidence from Field Experiments

The micro tier is where the evidence is strongest and most internally consistent. ?@tbl-micro lists the field experiments and quasi-experiments in the corpus; Figure 4 plots their headline productivity gains.

The canonical studies share a common architecture: workers are randomly assigned, or exposed through a staggered rollout, to a generative-AI tool, and output is measured on a well-defined task. Peng et al. (2023) find that developers given GitHub Copilot complete a programming task 55.8 percent faster. Noy and Zhang (2023) report that access to ChatGPT cuts the time to complete professional writing tasks by about 40 percent while raising assessed quality. Brynjolfsson, Li, and Raymond (2025) study the staggered introduction of an LLM-based assistant among more than five thousand customer-support agents and find average productivity—issues resolved per hour—rises about 15 percent. Dell’Acqua et al. (2023) document large gains for management consultants on tasks within the AI’s capability frontier alongside degraded performance on tasks outside it, a caution against extrapolating uniformly.

Across the 14 studies reporting a quantifiable percentage productivity effect, the distribution is tight in sign and central tendency: median 30 percent, mean 31 percent, interquartile range 13–39 percent, and every estimate positive. The spread—from 2.4 percent to 87 percent—reflects genuine heterogeneity in task, tool, and outcome definition rather than disagreement about direction.

V. Macro Assessments and Growth Accounting

The macro tier attempts what the micro tier cannot: an estimate of AI’s effect on aggregate output. Because no economy has yet experienced full AI diffusion, these are necessarily model-based projections, and they diverge dramatically. ?@tbl-macro lists the macro-projection studies; Figure 5 places three influential annual-productivity projections in the wider forecast landscape.

At the conservative end, Acemoglu (2024) applies a task-based framework: the aggregate TFP gain equals the average cost saving on exposed tasks, weighted by the share of tasks affected. Plugging in empirical task-exposure shares and the observed cost savings from micro studies yields a total TFP increase of roughly 0.5–0.7 percent over ten years—about 0.05–0.07 percentage points per year. At the optimistic end, task-exposure forecasts such as Briggs and Kodnani (2023, Goldman Sachs) project US labor-productivity growth boosts on the order of 1.5 percentage points per year and a global GDP increase near 7 percent over a decade. Bottom-up use-case accounting (Chui et al. 2023, McKinsey) lands in between, at 0.1–0.6 percentage points per year of productivity-growth potential.

The disagreement is not primarily empirical; it is structural. Three modeling choices drive nearly all of it: (i) the share of tasks deemed exposed to AI, which Eloundou et al. (2023) estimate is large but whose translation into automatable tasks is contested; (ii) the assumed cost of automating exposed tasks, where Acemoglu’s insistence on realistic, often hard-to-automate tasks pulls estimates down; and (iii) whether displaced labor is reallocated to comparably productive work or to low-value activities, which determines whether task automation adds to or merely redistributes output.

The full forecast landscape is wider and more revealing than any single figure. Cunningham (2025) assembles roughly three dozen forecasts of AI’s effect on growth over 2025–2035 and documents a striking bimodality (Figure 5): professional economists’ forecasts cluster in a narrow 0.1–1.5 percent-per-year band, while forecasts by AI-industry insiders run from 3 to 30 percent per year—more than an order of magnitude higher. The macro estimates emphasized above fall in or just below this lower, economist band—McKinsey and Goldman Sachs inside it, Acemoglu’s more conservative 0.05–0.07 percent-per-year figure just beneath its floor—and the projections in ?@tbl-macro are drawn almost entirely from that same community. A meta-overview built from the economics literature therefore samples one mode of a bimodal distribution, and the omitted mode is not a fringe: it reflects the views of those closest to the technology’s development. We flag this so the reader does not mistake the internal coherence of the economist band for a settled consensus about AI’s growth effect.

VI. Reconciling Micro and Macro

Why do 30-percent micro gains coexist with sub-one-percent macro projections? The apparent contradiction dissolves once five wedges are made explicit.

First, task versus job. Micro experiments measure output on a single, AI-suited task; jobs bundle many tasks, most of which the tool does not touch. A 30-percent gain on the coding portion of a developer’s work is a far smaller gain on the developer’s total output, and smaller still on the firm’s value added.

Second, selection and external validity. Field experiments are run where AI is expected to help, on tasks chosen for measurability. Dell’Acqua et al.’s (2023) “jagged frontier” result—gains inside the frontier, losses outside it—warns that the economy-wide task mix includes many tasks on which current AI does not help or actively hurts.

Third, the J-curve and complementary capital. As Brynjolfsson, Rock, and Syverson (2021) argue, the aggregate payoff waits on intangible investment that is still being built. Today’s macro statistics sample the trough of the J-curve, not its eventual peak.

Fourth, diffusion and adoption. Even a large frontier effect contributes little to aggregate output until a large share of relevant work actually uses the technology. Adoption is partial and uneven, and the macro projections that assume near-complete diffusion are describing a long-run steady state, not the current economy.

Fifth, capabilities progress versus economics. The four wedges above concern the economics of a fixed technology. Cunningham (2025) argues that the largest source of forecast disagreement lies elsewhere—in assumptions about how far AI capabilities will advance, not in the economic questions of current impact, diffusion speed, or substitutability. Many economists’ forecasts, he observes, effectively treat AI as a one-time shock and exclude any effect of AI on the rate of innovation, implicitly assuming little further capability growth; forecasters who expect rapid progress toward broadly capable systems arrive at correspondingly higher growth numbers. On this reading the micro–macro gap and the economist–insider gap are partly the same phenomenon: estimates anchored to today’s tools, whether measured in a field experiment or fed into a task-based model, will understate the aggregate effect if the tools keep improving. This wedge is also the most speculative, since it turns on a forecast of technological progress that lies outside economics proper.

Taken together, these wedges imply that the micro and macro estimates are not in conflict; they are answers to different questions. The micro literature bounds the technology’s frontier potential on suited tasks; the macro literature projects an economy-wide, diffusion-weighted average that current data cannot yet confirm.

VII. Heterogeneity, Diffusion, and Measurement

Three cross-cutting themes recur across strands. Heterogeneity: the productivity effect is not a scalar. It varies by worker skill (larger for the less skilled), by task position relative to the capability frontier, and by outcome dimension—several studies find speed gains alongside flat or declining quality. Diffusion: the observational studies in the corpus, which track adoption and firm-level outcomes rather than randomized exposure, generally report smaller and noisier effects than the experiments, consistent with real-world adoption frictions eroding frontier gains. Measurement: national accounts were not designed to capture free or bundled AI services, quality improvements, or the intangible investment that the J-curve framework places at the center of the story. The mismeasurement is not a nuisance to be signed away; it is a first-order feature of GPT episodes.

VIII. Open Questions

Four questions stand out for future work. (1) Durability: do the large short-run experimental gains persist as tasks and tools co-evolve, or do they attenuate as the easy wins are exhausted? (2) Aggregation: what is the empirical mapping from task-level gains to firm-level and sector-level productivity, and how much is lost at each step of aggregation? (3) Distribution: does the within-occupation compression documented in the micro studies survive general-equilibrium adjustment of wages and employment? (4) Measurement: can new data—firm-level adoption panels, task-level output logging—close the gap between what experiments measure and what national accounts record? Progress on the AI-productivity question will come less from more single-task experiments than from studies that bridge these levels.

IX. Conclusion

The post-2017 literature on AI and productivity is large, fast-moving, and—once its two tiers are separated—more coherent than it first appears. Micro field experiments deliver a consistent verdict: generative AI raises productivity on suited tasks, typically by tens of percent, and most for the least-skilled workers. Macro assessments deliver a consistent disagreement, spanning an order of magnitude, that is fully attributable to a handful of modeling assumptions about task exposure, automation cost, and labor reallocation. The reconciliation lies not in splitting the difference but in recognizing that the micro estimates describe frontier potential on selected tasks while the macro estimates describe a diffusion-weighted, still-unrealized aggregate. The productivity J-curve suggests that today’s modest aggregate statistics are consistent with a large eventual payoff that current national accounts are structurally ill-equipped to see. Whether that payoff materializes on the timescale of the optimistic forecasts, or the more measured one of the task-based models, is the central open empirical question of the coming decade.

References

Prepared and maintained as an automated meta-overview. Effect sizes are reproduced from the cited sources; the accompanying evidence table (evidence_table.csv) documents provenance for every study, distinguishing in-session retrieved records from author-supplied landmark citations. This document synthesizes publicly available research and is intended as a literature map, not as peer-reviewed primary research.