Three months. That is all it took for the share of working-age adults using AI tools worldwide to jump from roughly 16.3 percent to 17.8 percent, according to a population-normalized metric developed by the Microsoft AI for Good Lab. The finding, published in a May 2026 preprint on arXiv, represents one of the first serious attempts to measure AI adoption not by headline-grabbing user counts but by scaling anonymized product telemetry against demographic baselines from the World Bank.
To put the pace in perspective: smartphone adoption, often cited as the fastest consumer technology rollout in history, took roughly a decade to move from niche to mainstream. AI tool usage appears to be compressing a comparable curve into quarters, not years. If the current trajectory holds, more than one in five working-age adults globally would be using AI tools before the end of 2026.
What the metric actually measures
The paper, titled “Measuring AI Diffusion: A Population-Normalized Metric for Tracking Global AI Usage” (arXiv ID 2511.02781), converts anonymized signals generated when people interact with AI-powered services into a per-capita adoption rate. The denominator draws on the World Bank’s open data indicator SP.POP.TOTL, which tracks total population across countries and territories, not working-age population specifically. The paper applies a subset of that indicator, filtering to working-age cohorts, to produce its adoption rate; however, it does not fully detail which age bands define “working-age” or how the filtering is performed. By anchoring raw usage numbers to a recognized demographic baseline, the metric sidesteps a common distortion: comparing absolute user counts across nations with vastly different population sizes. Ten million AI users in a country of 50 million tells a very different story than 10 million in a country of 500 million.
The preprint is hosted on arXiv, the open-access repository operated by Cornell University and supported by a consortium of research institutions. That gives the paper a stable, independently citable record outside Microsoft’s corporate channels, which matters when outside researchers want to scrutinize or replicate the work.
Where the numbers get fuzzy
A single preprint from a single corporate research lab is not the same as a peer-reviewed consensus. Several gaps limit how far anyone should push the 17.8 percent figure.
First, the paper does not disclose which specific products or telemetry sources feed the numerator. It is unclear whether someone who asks a chatbot one question a month is counted the same as a developer running AI code-generation tools for eight hours a day. That distinction matters enormously for understanding whether the metric captures deep integration into work or casual experimentation.
Second, regional breakdowns have not been published. The quarterly spike could be driven by a handful of high-adoption economies, or it could reflect broad global growth. Without country-level data, a hiring manager in Nairobi and a hiring manager in New York cannot use this metric to gauge local conditions.
Third, arXiv preprints receive basic screening for scientific relevance, but not the adversarial evaluation that journal peer review demands. The population-normalization step involves design choices, such as which age bands qualify as “working-age” and how telemetry proxies map to individual users, that independent statisticians have not yet audited. Until they do, the headline number carries the weight of one team’s methodological decisions.
Finally, the relationship between AI adoption and internet access in lower-income regions remains unresolved. The International Telecommunication Union’s Facts and Figures reports track global connectivity, but no primary research has directly tied AI diffusion rates to internet penetration in developing countries. Claims about whether this growth is narrowing or widening digital divides are, for now, inference rather than measurement.
How this compares to other adoption estimates
The Microsoft metric is not the only attempt to size up AI adoption, but it takes a different approach than most. A McKinsey Global Survey published in mid-2024 found that 72 percent of organizations reported using AI in at least one business function, up from 55 percent the year before. Pew Research Center surveys have tracked individual awareness and usage of tools like ChatGPT among U.S. adults. OpenAI itself has disclosed surpassing 100 million weekly active users.
What distinguishes the Microsoft metric is its denominator. Rather than surveying a sample or reporting raw platform numbers, it attempts to express adoption as a share of the entire global working-age population. That framing is more conservative: 17.8 percent sounds modest next to McKinsey’s 72 percent organizational adoption rate, but it describes a fundamentally different thing. It asks not “how many companies have tried AI” but “what fraction of all adults of working age are actually using it.”
What the speed means for workers and employers
Strip away the methodological caveats and one signal comes through clearly: AI tool usage is growing fast enough to shift measurably in a three-month window. That pace has practical consequences.
For businesses evaluating workforce planning or technology budgets, the trajectory suggests adoption pressure is accelerating, not plateauing. Companies that have delayed AI integration strategies face a narrowing window before competitors, and their own employees, force the issue.
For workers in roles most exposed to AI capabilities, particularly in text generation, code production, data analysis, and customer service, the tighter the adoption curve, the less time there is to build complementary skills. The World Economic Forum’s Future of Jobs Report 2025 estimated that 40 percent of existing worker skills would need updating by 2030. A quarterly adoption jump of 1.5 percentage points suggests that timeline may already feel optimistic.
For policymakers, the gap between adoption speed and regulatory readiness is widening. The EU’s AI Act entered its phased implementation in 2025, but most countries still lack comprehensive frameworks. A technology that reaches nearly a fifth of the global working-age population before governance structures are in place creates risks that are easier to name than to manage: labor displacement without retraining infrastructure, data privacy erosion, and concentration of economic gains among firms and nations that move first.
Why the denominator matters more than the headline number
The most honest thing about the Microsoft preprint is what it does not claim. It tells us speed. It does not tell us depth, quality, or direction. A 1.5-point quarterly jump is steep by the standards of technology adoption curves, which historically follow S-shaped patterns: slow early growth, rapid middle growth, then a plateau. Whether AI adoption is still in the steep middle section or approaching an early ceiling depends on variables no single telemetry-based metric can capture, including pricing changes, infrastructure investment in developing economies, and regulatory friction.
What the number does establish is a baseline. For the first time, there is a documented, reproducible attempt to answer a deceptively simple question: what share of humanity is actually using AI? The answer, as of early 2026, is roughly one in five and climbing. The next question, what that usage is doing to jobs, economies, and societies, will require data this metric was never designed to provide. But it gives researchers, employers, and workers a starting line from which to measure everything that comes next.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
