AI advances will increase US total factor productivity by less than one percent over ten years
Assessment
Credible evidence or argument exists on multiple sides.
The claim is the headline estimate of Daron Acemoglu's 2024 paper "The Simple Macroeconomics of AI," which projects that AI will add no more than about 0.66% to US total factor productivity over ten years. The figure comes from a deliberately simple calculation: the share of work tasks AI will profitably affect, multiplied by the average cost savings on those tasks. Acemoglu argues that fewer than five percent of US work tasks will be profitably automated or augmented within a decade and that cost savings on affected tasks average roughly a quarter or less, and he sets aside gains from new tasks and products as unlikely to be large within the period.
The estimate is credibly disputed on every one of those inputs. Goldman Sachs researchers, responding directly to the paper, project that generative AI could raise annual US labor productivity growth by about 1.5 percentage points over ten years, roughly an order of magnitude more; OECD and Penn Wharton modeling also lands well above the one-percent ceiling, and analysts re-running Acemoglu's own formula with observed AI usage data reach estimates about ten times larger. Critics also object that the calculation excludes the channels through which general-purpose technologies have historically delivered their largest gains, including new tasks and new products and the deepening of existing automation, and that it extrapolates from the capabilities of 2023-era systems.
On the other side, aggregate productivity statistics have so far shown no unambiguous AI effect, enterprise deployments have often disappointed relative to pilots, and past general-purpose technologies took longer than a decade to appear in measured productivity, so a small ten-year TFP number remains a live possibility even if AI ultimately proves transformative. The dispute is genuinely empirical and will be resolved only as adoption, task coverage, and measured productivity data accrue over the coming decade; the balance of current model-based estimates sits above the claim's ceiling, which is why the credence recorded here leans against it without dismissing it.
Full reasoning: the evidence and decisions behind this verdict
The claim originates in Acemoglu, "The Simple Macroeconomics of AI" (NBER w32487, published in Economic Policy 2025: academic.oup.com/economicpolicy/article-abstract/40/121/13/7728473), which states the estimate verbatim: no more than a 0.66% TFP increase over 10 years (an earlier draft said 0.71%). The derivation multiplies three inputs, each mapped as a subclaim. First, the framework: a first-order Hulten approximation from task exposure and average cost savings. Second, the task share: under five percent of tasks profitably affected, obtained by combining Eloundou et al.'s (2023) ~20% exposure estimate for language models with Svanberg et al.'s (2024) finding that ~23% of exposed computer-vision tasks are cost-effective to automate. Third, the savings: average labor cost savings of roughly a quarter, from the Noy-Zhang, Brynjolfsson et al., and Peng et al. experiments, which Acemoglu argues may fall to ~14% on hard-to-learn tasks. The ceiling additionally requires that no large gains arrive from new tasks and products within the decade, which the paper asserts but does not quantify.
Credible denial is abundant and specific. Goldman Sachs' June 2024 report "Gen AI: Too Much Spend, Too Little Benefit?" (www.goldmansachs.com/images/migrated/insights/pages/gs-research/gen-ai--too-much-spend,-too-little-benefit-/TOM_AI%202.0_ForRedaction.pdf) presents Joseph Briggs's direct rebuttal, projecting ~25% of tasks affected and a ~15% cumulative labor productivity upside, an order of magnitude above the claim; notably the same report carries Acemoglu's own restatement of his low estimate and Jim Covello's even more skeptical view, so the disagreement runs within institutions, not just between them. OECD modeling (Filippucci et al. 2025, www.suerf.org/wp-content/uploads/2025/10/SUERF-Policy-Brief-1283_Filippucci-Gal-Laengle-Schief.pdf) projects 0.4-1.3 percentage points of annual labor productivity growth in high-exposure economies, and Penn Wharton (budgetmodel.wharton.upenn.edu/p/2025-09-08-the-projected-impact-of-generative-ai-on-future-productivity-growth/) projects a 1.5% productivity and GDP gain by 2035. An analysis re-running Acemoglu's formula with Anthropic Economic Index usage data (jablevine.com/articles/25/anthropic-economic-index) reaches roughly ten times his figure, arguing far more tasks are already profitably automatable or augmentable. Maxwell Tabarrok's "Contra Acemoglu on AI" (www.maximum-progress.com/p/contra-acemoglu-on-ai) attacks the exclusion of deepening automation and new tasks. A caveat cuts both ways in these comparisons: most rival figures are labor-productivity or GDP terms rather than TFP, so they overstate the direct contradiction by the capital-deepening component; even so, their implied TFP contributions generally exceed 1% over a decade.
What keeps the claim live rather than contradicted: measured aggregate TFP has shown no clear AI signal yet; a 2025 IMF working paper on AI and European productivity notes the link from micro gains to aggregate productivity remains unclear; enterprise adoption has repeatedly underdelivered relative to pilots; and the historical record of general-purpose technologies (electricity, computers) features multi-decade lags between capability and measured productivity, so the specific ten-year TFP window could come in under 1% even under bullish long-run scenarios. Against that, capabilities and adoption have advanced markedly since the 2023-era studies the estimate rests on, and survey evidence (e.g., an Atlanta Fed working paper on executives' reported AI effects) implies gains that would breach the ceiling within a few years if borne out.
Verdict: contested, with high confidence that the disagreement is real and credible on both sides. Credence 0.3 that the claim as stated is true: the preponderance of central estimates, the direction of capability and adoption trends since the paper's data vintage, and the fact that even Acemoglu frames his number as a current-capabilities floor-setting exercise all lean against the ceiling holding, while measurement lags and the thin aggregate evidence to date keep meaningful probability on it. What would move the verdict: measured US TFP growth attributable to AI visibly exceeding ~0.1pp per year across the late 2020s would push toward contradicted; another five years of flat aggregate TFP despite deep adoption would push toward supported.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Given that aggregate productivity gains can be estimated from task exposure times average task-level cost savings, and because fewer than five percent of US work tasks will be profitably automated or augmented within ten years while average labor cost savings on affected tasks are roughly a quarter, multiplying the two yields a total factor productivity gain of well under one percent over the decade. The ceiling holds only if AI also fails to generate large gains from new tasks and new products within the period, since the calculation covers cost savings on existing tasks alone.
Granting all premises the arithmetic follows, but the argument's framework is itself live-disputed: critics argue that the first-order Hulten approximation omits the channels through which general-purpose technologies deliver most of their gains, including deepening of existing automation and general-equilibrium effects. Its empirical weight rests on the under-five-percent task share, which is the most contested input, and on the absence of large new-task gains, which the calculation asserts rather than derives. The cost-savings premise is comparatively well grounded in experimental evidence and carries less of the dispute.
Because widespread generative AI adoption could raise annual US labor productivity growth by about 1.5 percentage points over ten years, and because generative AI may already be contributing significantly to the post-2022 acceleration in US labor productivity growth, AI's cumulative contribution to aggregate productivity over a decade would far exceed one percent, contradicting the claim's ceiling.
If either premise holds, cumulative gains would far exceed the claim's ceiling, so the inference goes through, with one caveat: both rivals are stated in labor-productivity rather than total-factor-productivity terms, so contradicting the claim requires the further step that the projected gains are not mostly capital deepening, which the underlying analyses support. The argument currently rests on the Goldman Sachs 1.5-percentage-point projection, itself contested, while the attribution of the post-2022 productivity acceleration to generative AI remains unassessed and is the weaker premise.
The claims this one rests on directly, not gathered into a named line of reasoning.
- contradictsthis argues against the parentsteward instructions →AI will generate large productivity gains from new tasks and new products within ten years ↗︎ · shared subclaim
Provenance
Where this claim has been said, linked to its canonical form.
these macroeconomic effects appear nontrivial but modest—no more than a 0.66% increase in total factor productivity (TFP) over 10 years
Using existing estimates on exposure to AI and productivity improvements at the task level, these macroeconomic effects appear nontrivial but modest—no more than a 0.66% increase in total factor productivity (TFP) over 10 years.
these macroeconomic effects appear non-trivial but modest – no more than a 0.66% increase in total factor productivity (TFP) over 10 years
The peer-reviewed published version of Acemoglu's working paper, deriving the estimate from task-level exposure and cost-savings figures via a first-order Hulten approximation.
This result is so much larger than Acemoglu's because Anthropic's data indicates that many more tasks are profitably automatable or augmentable than the predictions of Eloundou et al. and Svanberg et al. This corresponds to a roughly 10x increase (relative to Acemoglu) in GDP by 2034
Re-running Acemoglu's own Hulten-based calculation with task-usage data from the Anthropic Economic Index, the author argues far more tasks are profitably automatable or augmentable, yielding roughly ten times Acemoglu's estimate and thus denying the under-one-percent ceiling.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Aug 11, 2026. Every judgment on this page is accompanied by a reasoning trace.