AI will not generate large productivity gains from new tasks and new products within ten years
3 events · 1 assessment · 1 decision
Structured and assessed first pass
First pass (structure_and_assess). Decomposed into four subclaims under two named arguments. For: Acemoglu's directed-automation premise (new, seeded 0.55) and the historical GPT diffusion-lag claim (matched existing 03b71ec7-ba5e-440a-9b7f-2acd93aaf0be per Matcher, confidence 0.98). Against: the historical new-task/new-product channel claim (new, seeded 0.7) and generative AI's record adoption speed (new, seeded 0.85; Bick-Blandin-Deming w32966). All dependencies went through match_claim; three were novel. Both arguments given written forms and evaluated holds_with_caveats. Importance confirmed at 0.55 with contestation 0.75: the premise is where critics of Acemoglu's contested sub-1% TFP estimate concentrate. Assessed CONTESTED (confidence 0.8, credence 0.45, marginal_yield 0.35): Acemoglu credibly affirms, Goldman Sachs and other model-based estimates credibly deny; resolution is empirical and will accrue with adoption and productivity data. Used 4 web searches (Acemoglu w32487, critiques via AEI/Gans, Bick-Blandin-Deming adoption, GPT lag literature); the fifth search failed on the provider limit. Recorded NO instances this pass: none of the retrieved passages contained a verbatim assertion of the claim or its negation, and fabricating quote text would violate fidelity; the Acemoglu paper should be recorded as the originating affirming instance on a pass that reads the asserting passage directly. Escalated to Curator: canative form is a negation; the affirmative rewording (stance-flipping) is theirs to adjudicate, cheapest now while the claim has zero instances. Canonical form otherwise left unchanged (16 words, neutral within its polarity).
Assessed Contested
verdict confidence 0.80 · credence 0.45
The claim is a working premise of Daron Acemoglu's 2024 paper "The Simple Macroeconomics of AI," whose headline projection of modest ten-year productivity gains counts only cost savings on existing tasks. Acemoglu sets aside gains from new tasks and new products as unlikely to be large within the decade, arguing that the industry's effort is directed at automating and monetizing existing workflows, and that some new AI products may even carry negative social value. The case for the claim rests on economic history and on the direction of current investment. General-purpose technologies have historically taken decades before producing measurable aggregate productivity gains: the familiar shorthand runs about sixty years for steam, thirty for electricity, and fifteen for computing, largely because the deeper gains required reorganizing production and inventing new products, not just cheaper execution of old tasks. If AI development is directed mainly at automating existing tasks rather than creating new ones, the new-task channel would additionally be starved of investment during the window. The case against is that this history cuts the other way: new tasks and new products have historically been the main channel through which general-purpose technologies raised productivity, so excluding the channel is what critics of Acemoglu's estimate, including Goldman Sachs researchers and several academic commentators, identify as its central omission. And the lag analogy is weakened by evidence that generative AI is being adopted faster than the personal computer or the internet were, with new product categories such as conversational assistants and coding agents already generating revenue within two years of launch. The disagreement is empirical and time-indexed: it will narrow as adoption depth, new-product revenue, and measured productivity accrue over the coming decade. A residual definitional looseness in "large" means the two sides can partly talk past each other, but the core dispute, whether the decade window is long enough for this channel to show up in aggregate statistics, is genuine and unresolved.
Claim entered the graph