Widespread generative AI adoption could raise annual US labor productivity growth by about 1.5 percentage points over ten years
5 events · 2 assessments · 2 decisions
Reassessed
Trigger: subclaim_change (three first assessments landed: exposure estimate 8063ed7a supported 0.8; Acemoglu bound 8b075fb6 contested with credence 0.30 on the low bound; post-2022 attribution a570c6e5 contested credence 0.35). Judged materiality: the changes pull in offsetting directions and none forces a status change. The supported exposure figure firms the projection's input but carries the explicit caveat that exposure is not adoption, which is where the dispute lives; the credence lean against Acemoglu's low bound mildly strengthens the high-end side without vindicating 1.5pp; the attribution subclaim confirms no visible aggregate effect yet with emerging differential evidence. One recency web search (Goldman March 2026: no economy-wide AI-productivity relationship yet, projection maintained) confirmed the landscape's shape; no new instance recorded since that piece reports current visibility, not an assertion or denial of the ten-year projection. Verdict stays CONTESTED, confidence nudged 0.85 to 0.87; credence still omitted (modal claim, conditional timeline, false precision). Refreshed the reader-facing assessment and reasoning trace, which had gone stale (described the exposure subclaim as unassessed), and re-recorded both argument evaluations against the new premise standings (rival-estimates argument now noted as leaning less heavily on the Acemoglu bound). No structural change needed: the attribution debate is already held as a subclaim. No notification to dependents: status and overall shape unchanged, so nothing material propagates; the parent GDP claim's contested standing is unaffected. Marginal yield 0.25: resolution will come from time-series data over the coming years, not from more analysis now.
Reassessed: still Contested
verdict confidence 0.85 → 0.87
Structured and assessed first pass
First pass on the Goldman Sachs 1.5pp productivity projection. Decomposition: adopted the Curator's suggested edges — the customer-support 14-15% experiment as supports and the Danish earnings/hours null as contradicts — and declined the METR-experiment link since the assessment does not cite that experiment. Added the 300M-jobs exposure claim as requires (the derivation's base, already in the graph under the sibling GDP claim), Acemoglu's sub-1% TFP bound and the GPT-diffusion-lag claim as contradicts. Organized these under two named arguments (for: task-exposure derivation; against: rival estimates and slow diffusion), both written and evaluated as holds_with_caveats. Minted one novel subclaim after match_claim confirmed novelty (confidence 0.95): whether generative AI significantly contributes to the post-2022 US productivity acceleration — the live 2026 attribution dispute, seeded at 0.4 with importance 0.45/contestation 0.85. Canonical form tightened to add the US scope and the widespread-adoption condition, per the source report's Exhibit 13 and how critics engage the figure; same proposition, no identity change. Importance set to 0.6 (major: mechanism behind a widely cited projection, order-of-magnitude dispute among credible parties), contestation 0.85, superseding the Extractor's 0.55. Verdict CONTESTED at 0.85 confidence, no credence (modal projection, unresolved attribution; a single number would be false precision). Coherent with the parent GDP claim's contested assessment, per the Curator's coordination note (PO, EU, ND). Recorded two in-the-wild instances read during evidence gathering: Goldman's 2025 reaffirmation (affirms) and Acemoglu's NBER paper (denies). Marginal yield 0.35: the dispute is well mapped, but new aggregate data lands continually; a staleness re-pass in a year or so would have material to digest.
Assessed Contested
verdict confidence 0.85
The projection originates in a 2023 Goldman Sachs Research analysis by Joseph Briggs and Devesh Kodnani, which estimated that generative AI, once widely adopted, could raise annual US labor productivity growth by about 1.5 percentage points over a ten-year period. The figure is derived from task-exposure accounting: because the equivalent of some 300 million full-time jobs worldwide is exposed to automation, applying assumed task-level cost savings across roughly a quarter of US work tasks yields the aggregate lift. It is the mechanism behind the same analysis's widely cited projection that generative AI could raise global GDP by about 7% over ten years. Goldman has stood by the estimate, restating it in 2025 as a roughly 15% lift in the level of developed-market labor productivity at full adoption. Real deployments give the mechanism some footing: the best-identified field experiment found that a generative AI assistant raised customer support agents' productivity by roughly 14-15%. The magnitude and timeline are credibly disputed. Working from similar exposure data but different assumptions about which tasks can be profitably automated, Daron Acemoglu's task-based model finds that AI will raise total factor productivity by less than one percent over ten years, roughly an order of magnitude below the Goldman figure, and the Penn Wharton Budget Model reaches similarly modest numbers. Skeptics add that general-purpose technologies historically took decades to produce measurable aggregate productivity gains, and early worker-level evidence points the same way: chatbot adoption showed no detectable effect on workers' earnings or hours within two years in Danish administrative data, with users reporting time savings near 3% of work hours. Three years into the projection window, the aggregate data have not settled the question. US labor productivity growth has run above its pre-pandemic trend since 2023, but whether generative AI is a significant contributor to that acceleration is itself disputed: adoption remains shallow, and comparisons of AI-exposed and less-exposed industries show no clear differential signal, while proponents read the pickup as the front edge of a productivity J-curve. The disagreement is empirical and resolvable in time: realized US productivity data through the early 2030s, and whether AI-exposed industries pull measurably ahead, will show whether the high-end projections or the order-of-magnitude-lower academic estimates were closer.
Claim entered the graph