Key Findings: Decomposing the Capability Overhang

At the World Economic Forum in January 2026, OpenAI reframed the central challenge of AI development. The issue, it argued, was no longer capability. Current AI systems could already outperform most professionals across a wide range of knowledge work tasks. The problem was use. Between what AI could do and what people actually did with it lay a gap so large that closing it had become as urgent as any further technical advance. OpenAI called this the capability overhang.

What the capability overhang concept does not provide is a theoretical account of what this gap actually is. It treats the distance between AI’s potential and its realization as a single undifferentiated phenomenon, and proposes correspondingly uniform remedies: more access, more education, more infrastructure. This document presents the key findings of the working paper Decomposing the Capability Overhang, which argues that those prescriptions are not wrong but are undifferentiated in a way that makes them frequently counterproductive, because capability overhang is structurally two-dimensional, not one.

Operationalized against the Anthropic Economic Index dataset of 818,492 conversations across 173 countries, the AI Matrix produces a three-stage typology: education-first adoption (Indonesia, Peru, Ecuador), professional adoption (India, Brazil, Pakistan), and consumer diffusion (the Scandinavian countries, Germany, the UK, the United States). The first step in getting the intervention right is knowing which stage you are in. The capability overhang will not be closed by access provision alone.