Economy

Central banks see little proof yet that AI has lifted productivity

An RBA official's Nov. 9 speech reopens a dispute over how much AI will add to annual productivity growth, and why the estimates differ so widely.

12 min read
Watercolour of a half-built wooden footbridge spanning a misty river at dawn, with a gap in the middle.
AI-generated illustration

Reserve Bank of Australia Assistant Governor Sarah Hunter said there is limited evidence so far that AI has lifted aggregate productivity. Published estimates of the long-run effect run from about 0.1% to more than 2% a year, in part because they measure different things.

Central bankers are increasingly saying the same thing about artificial intelligence: investment is real, adoption is fast, and a clear payoff in national output per hour has not yet shown up in the data. On Nov. 9, 2026, Reserve Bank of Australia Assistant Governor Sarah Hunter put it plainly in a speech at a UBS conference in Sydney. In her words, "So far, there is limited evidence that AI has materially lifted aggregate productivity."

The statement matters because productivity sets how fast an economy can grow without stoking inflation, and because interest-rate decisions, asset prices and hiring plans increasingly rest on bets about AI. Those bets vary enormously. Secondary reports of the same speech say current estimates of AI's effect on annual productivity growth over the next decade run from more than 2% to just 0.1%. That spread is not a simple disagreement over one number, and understanding what lies underneath it is the key to reading the debate.

What the Reserve Bank of Australia actually said

According to the speech text on the RBA website, Hunter argued that firms are still taking time to invest and reorganize so they can capture the full benefits of AI, while also managing new operational and resilience risks. The speech cites a chart of the share of firms reporting AI productivity gains, with only a very small share reporting losses. Those figures come from the RBA's liaison program with businesses, which means they are self-reported.

Hunter also said the evidence suggests AI is spreading quickly, by many measures faster than earlier general-purpose technologies. Fast adoption and a missing aggregate payoff can coexist. New tools often need changes in workflows, training and organization before they show up in output per hour, a delay economists sometimes describe as a J-curve.

The speech also touched on financial stability. Press coverage of it reports that the RBA estimates realized data-center investment has risen 10-fold over three years, from a very low base. Hunter warned that asset prices could face sharp or even disorderly repricing if returns on AI investment fall short of expectations.

What US productivity data show, and what they cannot

The US numbers look strong. The Bureau of Labor Statistics reported on Nov. 5, 2026, in preliminary figures, that nonfarm business productivity rose at a 3.6% seasonally adjusted annual rate in the third quarter, while unit labor costs fell 0.7%. Output grew 3.7% and hours worked rose just 0.1%. Productivity was 2.1% above its level a year earlier.

The longer record is also firm. Since the fourth quarter of 2019, the last business-cycle peak, productivity has grown at a 2.2% annualized rate through the third quarter of 2026, according to the BLS. Output rose 2.7% a year over that stretch and hours 0.4% a year. Goldman Sachs economists put average nonfarm labor productivity growth at about 2% over the past five years, against about 1.5% before the pandemic, though that is Goldman's estimate rather than a figure verified here against BLS tables.

Several cautions apply. One quarter is a volatile reading, and the Q3 figure is preliminary, with a revision due Dec. 8, 2026. Summaries of the BLS release report that second-quarter growth was revised up to 2.1% from 1.4%, a reminder that early numbers move. Manufacturing productivity rose only 0.3% in Q3, so the strength is not uniform.

The bigger problem is attribution. Aggregate statistics do not record why productivity rose. The International Monetary Fund's April 2026 World Economic Outlook says recent AI investment and adoption could substantially lift productivity and medium-term growth sooner than expected, "as possibly foretold" by above-trend US productivity since 2020. The text retrieved for this article was garbled, though, and also suggests AI may not yet be the main force behind the stronger growth. Fed officials make a similar point, as the next sections show.

Why estimates run from 0.1% to more than 2% a year

The reported range comes from secondary coverage of the RBA speech, including FourWeekMBA, which quotes it as running from over 2% per annum to just 0.1%. The footnotes of the speech were not available for this article, so it is not confirmed which studies the RBA had in mind. Two widely cited estimates, though, illustrate how a gap that large can open up.

The cautious end: a task-by-task tally

MIT economist Daron Acemoglu's paper, The Simple Macroeconomics of AI (NBER Working Paper 32487, May 2024; later published in Economic Policy in 2025), builds up from task-level evidence. It finds macroeconomic effects that are nontrivial but modest: no more than a 0.66% increase in total factor productivity over 10 years. Because early evidence comes from tasks that are easy to learn, he argues the true gain is likely smaller, under 0.53%. His first estimate of the 10-year GDP effect is roughly 0.93%, assuming capital grows in step with productivity, and the paper's own annualized figure is about 0.064% a year.

The method multiplies the share of tasks exposed to AI, the share of those that can be automated profitably, and the cost saving. That makes the result sensitive to each assumption. It is also tied to the AI of roughly 2023 and 2024. Capabilities have changed since then, so the estimate may be out of date.

The ambitious end: a scenario built on widespread adoption

Goldman Sachs economists Joseph Briggs and Devesh Kodnani took a different approach in 2023. In their baseline, AI could lift US productivity growth by 1.5 percentage points a year, assuming widespread adoption over a 10-year period. That equals a 15% rise in the level of aggregate productivity over roughly a decade. In 2026, Goldman still estimates a 15% gross uplift to economy-wide US labor productivity after full adoption, according to a Goldman research report.

This is a conditional forecast, not a measurement. It is investment-bank research rather than peer-reviewed work, and Goldman has commercial interests in AI-related capital markets. The research behind it draws partly on firm anecdotes of 25% to 30% productivity gains, according to the report. Goldman's separate trend forecast for US productivity growth is 1.7% through 2029 and 1.9% in the early 2030s, which is a total, not an AI-only increment.

Two headline AI productivity estimates, as published and as converted to a yearly rate
Estimate What it measures Stated figure Horizon Approximate yearly rate (Jaroha's arithmetic unless noted)
Acemoglu 2024, upper bound Total factor productivity, level No more than 0.66% 10 years About 0.06% to 0.07% (paper states about 0.064%)
Acemoglu 2024, adjusted for hard tasks Total factor productivity, level Under 0.53% 10 years About 0.05%
Acemoglu 2024, first estimate GDP, level About 0.93% 10 years About 0.09%
Goldman Sachs 2023 and 2026 Labor productivity, level 15% About 10 years after widespread adoption About 1.4%
Goldman Sachs 2023 Labor productivity growth About 1.5 percentage points a year 10 years after widespread adoption 1.5% (as stated)

Why the same technology yields figures 20 times apart

Converted to annual terms, the two estimates differ by roughly 20-fold. Part of the gap is definitional. Acemoglu measures total factor productivity, output per unit of all inputs combined, while Goldman measures labor productivity, output per hour worked. Capital deepening, meaning more equipment per worker, raises labor productivity but not total factor productivity, so the latter will tend to come out smaller.

A second source of divergence is whether a number describes a level or a growth rate. A one-time 0.66% higher level over 10 years is very different from 0.66% of extra growth every year. Compounding also matters: 2% a year for a decade produces roughly a 22% higher level, well above Goldman's 15%, so even the top of the quoted range is not identical to Goldman's scenario.

Most of the gap, however, comes from assumptions: how many tasks AI touches, how much cost it saves and whether adoption becomes widespread. An IMF staff note on the global implications of AI adds a further brake. Physical, organizational and institutional bottlenecks limit aggregate gains, so even very high productivity in automated tasks may translate into only modest growth in the whole economy.

What central banks say about jobs

On employment, Hunter's speech finds little evidence so far of large effects on aggregate jobs in advanced economies. There are, she said, emerging signs of weaker hiring, particularly among younger workers in highly AI-exposed occupations. That reading rests on research the speech cites, which was not retrieved for this article, so it should be treated as a tentative signal rather than a settled result.

In Australia, RBA liaison contacts say AI has not yet affected headcount at most firms and they do not expect it to over the next year. Goldman's forecast looks further out: its 2026 report expects AI to displace some share of workers over the adoption cycle, though the retrieved text of that figure was fragmentary and is not relied on here.

Task-level studies show why individual workers may feel the change before the aggregate data do. Federal Reserve Governor Michael Barr, in a July 2026 speech, cited a study in which AI cut average assignment time by 40% and raised quality by 18%, with the biggest gains for the weakest unaided performers. The underlying study was not identified in the research, and the result applies to specific tasks, not the whole economy.

What central banks say about prices

The inflation story has two timelines. As reported by Reuters and other outlets, Hunter said that in the near term the positive shock to aggregate demand, set against the economy's current supply capacity, will put some upward pressure on inflation. Spending on data centers, equipment and infrastructure adds demand before any productivity gain adds supply. Over the longer run, she said, AI's effect on inflation is likely to be broadly neutral: if gains broaden, economies can produce more with the same inputs and the initial pressure should fade.

Fed officials lean somewhat more optimistic on timing. Governor Lisa Cook said in a Sept. 28, 2026 speech that she expects productivity gains to bring modest disinflation within the next few years. The payoff, she said, depends on how quickly and broadly businesses adopt AI tools and how fast any gains pass through to the labor market. She cited the Fed's Small Business Credit Survey, in which nearly half of small employer firms use AI and 71% of those report higher productivity as a result.

Others stress the risk of misjudging the timing. St. Louis Fed President Alberto Musalem describes himself as an AI and productivity optimist, but writes that the data to date are inconclusive on whether aggregate productivity has entered a sustained higher-growth regime, and that the risks of getting the productivity and inflation call wrong are too great. According to an Axios report from June, some Fed officials see clearer evidence of AI investment raising demand for labor, equipment and infrastructure than of widespread productivity gains. Axios quotes San Francisco Fed President Mary Daly saying companies tell her they have not seen the productivity yet. Those quotes were not checked against primary transcripts.

How strong is the evidence on each side?

For the optimistic reading

Productivity is running above its pre-2020 pace, and both Fed and IMF officials describe that as consistent with early AI effects. Musalem says US productivity growth has largely recovered from the pandemic slump, driven in large part by firms' investment in automation, including AI. St. Louis Fed research reportedly finds productivity gains in occupations where AI use has become more common, though that research was not read for this article. Small firms also report gains, and task-level experiments show large time savings.

For the skeptical reading

The same officials caution that the acceleration may owe something to pandemic-era changes. Barr's February 2026 speech refers to a period of higher productivity that began with the pandemic. He also lays out scenarios, including an "AI bust" in which AI cannot master hard-to-learn tasks or investment stops abruptly, leaving only a modest boost that fades. He adds that a boost relative to a world without AI does not necessarily mean growth well above its long-run trend, though he also calls a profound long-term positive impact on productivity growth very likely.

Survey evidence is softer than it looks. The 71% figure in the Fed's small-business survey is a self-assessment, and it covers only the firms that use AI, so optimism and selection could inflate it. Investment-based claims are weak productivity evidence too. Goldman notes that much AI spending goes on imported hardware that barely registers in domestic GDP, so the argument that AI is driving growth because investment is rising does not show that output per hour has improved.

What to watch next

The 0.1%-to-2% spread is best read as a map of uncertainty rather than a contest between two forecasts. At one end are narrow, task-based calculations tied to today's tools. At the other are conditional scenarios that assume broad adoption and large savings. Neither can be tested against the data yet, and the RBA's point is that the aggregate numbers do not currently favor either one.

Near-term signposts are concrete: the BLS revision to third-quarter US productivity due Dec. 8, 2026, whether strong readings persist as the effect of the pandemic fades, and whether weaker hiring of young workers in exposed jobs spreads. The longer-term question is whether the heavy spending now under way is eventually matched by gains in output per hour. If it is not, central banks have already warned that asset prices could adjust sharply.

Frequently asked questions

Is AI boosting productivity right now?

At the task level, studies and firm surveys report gains, such as a 40% cut in assignment time cited by Fed Governor Barr. In aggregate data, the RBA's Hunter said on Nov. 9, 2026 that there is limited evidence of a material lift. US productivity has been strong, but Fed officials say attributing that to AI is inconclusive.

Why do estimates of AI's productivity effect range from 0.1% to over 2% a year?

The estimates measure different things: total factor productivity in Acemoglu's work, labor productivity in Goldman's. Some are one-time level changes and others are annual growth rates. They also differ in how many tasks AI is assumed to affect, how much cost it saves and how widely firms adopt it.

Will AI cause inflation or reduce it?

The RBA expects some upward pressure in the near term as AI-related spending raises demand ahead of any gain in supply, and a broadly neutral effect over the longer run. Fed Governor Cook expects modest disinflation within the next few years if adoption and productivity gains come through. Musalem warns that errors in judging the productivity effect carry large risks.

Is AI costing jobs yet?

The RBA says there is little evidence of large effects on aggregate employment in advanced economies. It notes emerging signs of weaker hiring among younger workers in highly AI-exposed occupations. Australian firms contacted by the RBA say AI has not yet affected their headcount and they do not expect it to over the next year.

What did Daron Acemoglu estimate?

His 2024 paper puts the 10-year gain in total factor productivity at no more than 0.66%, and under 0.53% if hard-to-learn tasks yield less than the easy ones studied so far. The paper's annualized figure is about 0.064% a year. His estimate reflects the AI of roughly 2023 and 2024 and depends on assumptions about the share of tasks affected.

Sources (16)

  1. What Happens to the Economy as Intelligence Gets Cheaper? , Reserve Bank of Australia (2026-11-09)
  2. RBA's Hunter says AI has yet to deliver productivity boost, trigger job losses , Finwire (Reuters) (2026-11-09)
  3. Five Official Bodies Flagged AI Risks and Gains in Five Days , FourWeekMBA
  4. RBA's Hunter warns AI could trigger sharp asset repricing as boom adds to inflation , InvestingLive
  5. The Simple Macroeconomics of AI (NBER Working Paper 32487) , National Bureau of Economic Research (2024-05-01)
  6. AI may start to boost US GDP in 2027 , Goldman Sachs
  7. Goldman Sachs Research report on AI and jobs (Issue 149) , Goldman Sachs
  8. Productivity Home Page , U.S. Bureau of Labor Statistics (2026-11-05)
  9. US Labor Productivity Up 3.6% in Q3 2026, Unit Labor Costs Fall 0.7% , IndexBox
  10. Speech by Governor Cook on an update on AI and the economy , Board of Governors of the Federal Reserve System (2026-09-28)
  11. Speech by Governor Barr on artificial intelligence and the labor market , Board of Governors of the Federal Reserve System (2026-02-17)
  12. Speech by Governor Barr on artificial intelligence , Board of Governors of the Federal Reserve System (2026-07-14)
  13. How Should the Promise of Higher Productivity Growth Change the Reality of Monetary Policy Today? , Federal Reserve Bank of St. Louis
  14. Fed officials warn AI's economic costs may arrive faster than benefits , Axios (2026-06-01)
  15. World Economic Outlook, April 2026; Chapter 1 , International Monetary Fund
  16. Global Economic and Financial Implications of Artificial Intelligence , International Monetary Fund

Share this article

Back to list