Widespread generative AI adoption could raise global GDP by about 7% over ten years
7 events · 3 assessments · 3 decisions
Reassessed (staleness check): re-affirmed contested, credence 0.3, confidence 0.88
Cadence-triggered re-check six days after the prior assessment. Ran two web searches for movement in the evidence landscape. Found new 2026 Federal Reserve system analyses (St. Louis/SF Fed tracking piece, Fed Board buildout note) showing AI's current GDP contribution running through the investment boom with productivity effects still unmeasured, plus MIT GenAI Divide data on enterprise pilots stalling and St. Louis Fed adoption survey data (55% of people, 37% of workers). This material is relevant context on the realized-data track the verdict awaits but resolves nothing: it addresses a different channel (capex, not adoption-driven productivity) and none of it asserts or denies the 7% projection in its own voice, so no new instances were recorded. Structure unchanged; no new dependencies discovered; the one unassessed subclaim (GPT diffusion lags) awaits its own steward. Re-recorded the assessment with the updated landscape woven in, confirmed both argument evaluations (verdicts unchanged, prose refreshed to reflect early realized data), held credence and confidence, and set marginal_yield low (0.15) since another pass now buys little; the meaningful updates will come from aggregate productivity data as it accrues. No status change, so no dependent notification sent. Importance left at 0.7, consistent with a consequential, actively argued macroeconomic projection.
Reassessed: still Contested
verdict confidence 0.88 · credence 0.30
Reassessed no status change
Triggered by first assessments on three subclaims: the required 1.5pp US productivity-lift premise (cf70edb4, now contested 0.85), the 300M-jobs exposure figure (8063ed7a, now supported 0.8/0.75), and the Acemoglu sub-1% TFP bound (8b075fb6, now contested 0.85, credence 0.30). All three are coherent with the existing contested verdict, so the status stands. I refreshed the assessment texts because the prior reasoning described the productivity premise as unassessed, which is no longer true; raised confidence slightly (0.85 to 0.88) since the key premise's contested standing is now an assessed finding; held credence at 0.30, judging that the credence lean against the low-end TFP bound mildly favors the higher-end side but does not close the order-of-magnitude gap to 7%. Updated the Goldman argument's written form to reflect the tightened US-scope canonical form of the productivity subclaim, and re-recorded both argument evaluations against the current premise assessments (both holds_with_caveats). No structural changes: the decomposition already captures the dispute. No new web search: the trigger was internal and the external evidence base is unchanged since the first pass. No dependent notification: status, credence, and the shape of the assessment are materially unchanged, so no dependent's verdict could turn on this refresh. Marginal yield 0.25: the dispute is well mapped; the main future mover is realized productivity data through the late 2020s, which a staleness pass should pick up.
Reassessed: still Contested
verdict confidence 0.85 → 0.88 · credence 0.30
Structured and assessed
First pass (structure_and_assess) plus three Curator suggestions concerning cf70edb4. Decomposition: adopted the Curator's 'requires' reading for cf70edb4 (the 1.5pp productivity lift) over the 'supports' alternative, since in the originating Goldman Sachs analysis the 7% GDP figure is computed from the productivity lift and, more broadly, a genAI-attributable GDP gain of that size has no channel other than productivity; grouped it with existing claim 8063ed7a (300M jobs exposed, supports) under a named for-argument (Goldman Sachs task-based projection). Attached existing claim 8b075fb6 (Acemoglu TFP <1% over ten years) as contradicts and minted one novel subclaim (Matcher-confirmed) on the historical GPT diffusion lag (contradicts, importance 0.35, seeded 0.75), both under a named against-argument. Declined to mint a separate "AI will raise GDP by about 1% over the next decade" claim despite the Matcher deeming it novel: the existing TFP claim already carries the Acemoglu rival in the graph and a parallel GDP-magnitude node would add little structure here (NOR/lean-graph judgment; the Curator can revisit). Also declined to mint the "quarter of US/Europe work tasks automatable" premise: it belongs in cf70edb4's own decomposition, which is its Steward's domain. Importance set to 0.7 (contestation 0.85), refining the Extractor's 0.75: a heavily cited, actively contested anchor of the AI-macro debate, below worldview-level central claims. Evidence pass (3 web searches): Goldman original and its public defense against Acemoglu; Acemoglu NBER w32487 (~1% GDP over a decade); Penn Wharton (1.5% by 2035); early cross-country CES estimate near Acemoglu's range; McKinsey higher than Goldman. Recorded two denying instances (Acemoglu, Penn Wharton) at reduced confidence as rival point estimates. Verdict: contested (0.85), credence 0.3, marginal_yield 0.4 since realized productivity data and the unassessed required subclaim could move this within a year or two. Canonical form kept: already neutral and correctly scoped. No dependents exist, so no notifications sent.
Assessed Contested
verdict confidence 0.85 · credence 0.30
The figure originates in a 2023 Goldman Sachs Research analysis by Joseph Briggs and Devesh Kodnani, which projected that generative AI, once widely adopted, could raise global GDP by about 7% (roughly $7 trillion) over a ten-year period. The projection is derived from a task-exposure accounting: because the equivalent of some 300 million full-time jobs worldwide is exposed to automation, the analysis expects annual labor productivity growth to rise by about 1.5 percentage points over ten years, which compounds into the GDP gain. The projection is explicitly conditional on widespread adoption, and Goldman itself expected the effects to appear in GDP data only gradually. The magnitude is credibly disputed. Working from similar task-level exposure data but different assumptions about cost savings and the share of tasks profitably automatable, Daron Acemoglu's task-based model finds a total factor productivity gain of less than one percent over ten years, implying a GDP effect near 1%, roughly an order of magnitude below the Goldman figure. The Penn Wharton Budget Model similarly projects a productivity and GDP gain of about 1.5% by 2035. Skeptics also note that past general-purpose technologies took decades to produce measurable aggregate productivity gains, which weighs against the ten-year horizon. On the other side, McKinsey Global Institute's estimates are larger than Goldman's, and Goldman has publicly defended its projection against Acemoglu's critique, attributing the gap to differing assumptions about automatable task shares, labor reallocation, and new task creation. The disagreement is empirical rather than definitional, and it is resolvable in time: realized productivity and GDP data through the late 2020s and early 2030s, together with firm-level evidence on whether generative AI adoption produces cost savings at the assumed scale, will show which family of estimates was closer. As of now, independent academic estimates cluster well below 7% for a ten-year window, while the higher projections remain defended by their authors and widely cited.
Claim entered the graph