Deployment Overhang

In the AI Story

The evidence for the overhang is empirical. Surveys of knowledge-work adoption (McKinsey 2024, Anthropic's own economic-impact research, OpenAI's deployment analytics) consistently find that frequent, task-integrated use of AI tools is a minority behavior even in professions where the tools are clearly productive (programming, writing, legal research, customer support). The gap between capability and use is larger in domains with higher regulatory overhead (medicine, finance, law), larger in organizations with older workforces, and larger where tool use is not already part of the professional identity. This is ordinary technology diffusion behavior; it is also unusual because the rate at which capability is advancing exceeds the rate at which diffusion historically closes.

Clarke's satellite forecast is instructive here as well. In 1945 the extra-terrestrial relay concept was technically plausible; the geostationary orbit was understood; the engineering challenges were estimable. It took nineteen years to the first operational geostationary satellite and another decade for the technology to reshape telecommunications. The gap was filled by launch infrastructure, international coordination, ground-station networks, billing relationships between carriers, and regulatory regimes. Each was a deployment step, not a capability step, and each was harder than it looked. The present AI situation has a compressed timeline but the same basic structure: capability arrives before the enabling substrate of skills, workflows, integrations, and norms.

The policy-relevant consequence is that most of the social effects of a given AI capability are not felt at the moment the capability is demonstrated; they are felt during the diffusion period after deployment at scale. This has two implications. First, the temptation to evaluate a capability by its immediate, visible effects is misleading; the evaluator is looking at the leading edge and missing the bulk. Second, the window during which societal adjustment is still possible is longer than it looks at the moment of capability release; it is the length of the diffusion period, not the length of the demo period. Whether the societal adjustment that happens during diffusion is sufficient or insufficient is the open question.

The overhang is also a risk category. Capabilities that sit unused are also unevaluated in their downstream effects; when the overhang closes, the effects arrive quickly. The transition from "AI can do X in principle" to "millions of users do X daily" can be short (weeks for consumer products) or long (years for enterprise deployments), but the risk profile of the deployed-at-scale state is only discoverable after deployment at scale. This is not a unique feature of AI; it applies to every transformative technology. It does mean that the period of greatest uncertainty about a new AI capability is not the moment of release but the moment of scaled integration, which lags by months to years.

Origin

"Compute overhang" entered the AI-safety literature through Bostrom's Superintelligence (2014) and has been elaborated by Epoch AI's compute-trend research. The "deployment overhang" terminology is more recent, appearing in policy discussions around 2022–2023 as the gap between demo and integrated use became the salient strategic variable.

Key Ideas

Capability is not consequence. The demo does not have the effects; the integrated deployment does, and these differ in kind and timing.

Diffusion has a substrate. Workflows, skills, billing, regulation, and norms each gate the speed of adoption.

The adjustment window is diffusion-length, not release-length. Societal adaptation has more time than it looks to have from the pace of the leading edge.

Risk lags capability release. The deployed-at-scale state is where new risk profiles emerge; the release itself is often the quietest period.