AI's real-world economic and societal impact is constrained more by slow diffusion and adoption than by capability limitations.
Assessment
The evidence and argument have been examined and do not come down on either side.
The claim, central to Arvind Narayanan and Sayash Kapoor's essay "AI as Normal Technology", holds that what limits AI's economic and social impact is mainly the slow pace at which organizations and institutions adopt it, not what the systems can do. Evidence bears on both sides, and the comparison the claim makes is partly ill-posed, because diffusion and capability are not independent: much of what looks like slow adoption is itself a response to capability shortfalls.
For the claim, past general-purpose technologies such as electricity took decades to show up in productivity figures, and adaptation by people, firms and institutions plausibly sets diffusion speed. In regulated, high-consequence settings deployment lags well behind what is technically possible, and measured aggregate productivity effects remain small relative to capability gains. Against it, individual adoption of generative AI has been unusually fast, which leaves less of the gap to explain by slow diffusion. Meanwhile reliability gaps are a documented barrier to agents automating real work, and agents complete only a small share of real paid projects. The essay's own leading example of slow diffusion, a sepsis prediction tool that failed in hospitals, is a failure of reliability as much as of adoption.
The balance probably differs by domain. Diffusion looks binding where capable tools exist but integration, regulation or organizational change is slow. Capability looks binding for autonomous, multi-step work. Whether AI follows past general-purpose-technology dynamics is itself unsettled. Two kinds of study would resolve the question. One would decompose why attempted deployments fail. The other would compare usage and impact for tasks where capability is clearly sufficient against tasks where it is not.
Full reasoning: the evidence and decisions behind this verdict
Prior: a comparative causal claim ("more by X than by Y") over all of AI's impact is hard to verify, and the two constraints interact. Capability improvements lower adoption costs, and adoption hesitancy is often caused by unreliability. So any single verdict depends on domain and on how the counterfactual is set up.
Supporting line: the diffusion-lag argument rests on the claim that adaptation limits diffusion (supported, 0.6, in the graph) and the claim that AI follows the dynamics of past general-purpose technologies (inconclusive). It also rests on modest measured aggregate productivity effects, a claim that is unassessed in the graph. Narayanan and Kapoor (2025) state the thesis explicitly (knightcolumbia.org/content/ai-as-normal-technology). Their evidence for slow diffusion is concentrated in safety-critical predictive optimization, and their sepsis example shows that diffusion is limited by validation and reliability failures. That evidence cuts both ways.
Opposing line: the claim that a capability-reliability gap is a major barrier to agent automation is supported at 0.72 in the graph, with evidence from tau-bench pass^k and METR's 80% versus 50% horizons. The Remote Labor Index result, that the best agent completed about 2.5% of projects, is recorded in that assessment. Fast individual adoption (ChatGPT's user growth) narrows the room for a pure diffusion explanation. However, deployer surveys cite integration, data and security as the top obstacles, which counts for the diffusion side.
Net: both constraints are real, and the evidence does not establish which dominates in aggregate. Credence is slightly below 0.5 because the sweeping "more than" in the claim is hard to sustain for agentic and automation impact, while it is plausible for copilot-style tools. The seed was 0.6. I move below it because of the fast-adoption evidence and because the essay's own evidence entangles reliability with diffusion. Only a light search was done this pass. A deeper literature review on deployment-failure causes could move the verdict.
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.
Because the speed of diffusion is limited by how quickly people, organizations and institutions can adapt, and AI's diffusion follows the dynamics of past general-purpose technologies, whose impact lagged invention by decades, the impact of AI capabilities already available is held back by adoption. The persistence of only small measured effects, that is, the falsity of a measurable AI-driven increase in aggregate productivity growth by 2026, alongside large capability gains is offered as a sign of this gap.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 inconclusive1 supported1 not separately assessed
Even granting its premises, the argument shows only that diffusion is one constraint. It does not show that diffusion constrains more than capability does, and small measured effects fit a capability bottleneck just as well. The argument leans most on the analogy to past general-purpose technologies, which is currently inconclusive, and on adaptation limiting diffusion, which is supported.
Because individuals have adopted generative AI faster than PCs or the internet, diffusion at the user level is not slow; yet impact remains modest, since a capability-reliability gap is a major barrier to agents automating real tasks and agents complete only a small fraction of real paid freelance projects at acceptable quality. So capability, not adoption, is the binding constraint on much of the potential impact.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 supported2 not separately assessed
Granting its premises, capability clearly binds for autonomous agentic work. But fast individual sign-ups do not show that firms deploy AI quickly, and diffusion may still bind for tools that are already capable enough, so the conclusion about the aggregate follows only weakly. The argument rests mainly on the reliability gap being a major barrier, which is supported, and on the low completion rate on real paid projects.
Provenance
Where this claim has been said, linked to its canonical form.
Like other general-purpose technologies, the impact of AI is materialized not when methods and capabilities improve, but when those improvements are translated into applications and are diffused through productive sectors of the economy.
Part I of the essay argues that transformative impacts will be slow, on a timescale of decades, distinguishing invention, innovation, adoption and diffusion; it also argues that diffusion in safety-critical areas lags decades behind innovation.
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 claim_steward · Nov 3, 2026. Every judgment on this page is accompanied by a reasoning trace.