What Data Do We Have on AI Adoption?

A first map of the evidence: what current AI adoption indices measure, where they are useful, where they are thin, and what they tell us about the direction of the AI economy.

25 May 2026

AI adoption is often described as if it were a single curve. It is not. One chart says a large share of workers has tried ChatGPT. Another says only a small share of firms uses AI in production. A platform report shows millions of real conversations, mostly writing, coding, and practical guidance. A macro index ranks countries by their preparedness to benefit from AI. An exposure paper says whole occupations could be affected, while labor-market data still shows only weak aggregate effects.

All of these can be true at the same time.

The reason is simple: AI adoption is not one thing. It can mean technical exposure, individual use, organizational deployment, national readiness, workflow integration, or measurable economic impact. The current data landscape is valuable precisely because it is fragmented. Each dataset shows one layer of the transition. None of them, alone, tells us whether AI is already transforming the economy.

This is the right starting point for gitthub labs: before making claims about AI and development, inequality, productivity, or the future of work, we need to ask what can actually be observed.

1. Exposure: what could AI touch?

The first family of measures does not observe adoption directly. It estimates where AI could matter.

The canonical approach is task-based. Researchers start with an occupational taxonomy such as O*NET, list the tasks attached to each job, and ask whether a model could perform, assist, or speed up those tasks. The influential early example is Eloundou, Manning, Mishkin, and Rock’s “GPTs are GPTs”, which estimated how much of the U.S. labor market is exposed to large language models. Related exposure measures appear in IMF work on generative AI and the future of work, where exposure is combined with a concept of complementarity: some exposed jobs may be displaced by AI, while others may become more productive because AI works with human judgement rather than replacing it.

Exposure indices are useful because they are broad. They let us compare occupations, sectors, countries, and income groups before the technology is fully diffused. They are especially useful for policy: if advanced economies have more high-exposure cognitive work, and low-income economies have less infrastructure to benefit from AI, then the distributional question appears early.

But exposure is not adoption. It tells us where AI might go, not where it has gone. It can overstate impact if tasks are technically feasible but legally constrained, socially resisted, poorly integrated, or economically unprofitable. It can understate impact if new workflows emerge that are not described in existing task databases. Exposure is the map of possible contact, not the record of use.

2. Platform data: what are people actually doing?

The second family of measures looks at real usage inside AI products. This is where the field has changed most quickly.

Anthropic’s Economic Index uses privacy-preserving analysis of Claude conversations and API traffic, maps observed tasks to O*NET, and separates some uses into augmentation and automation patterns. Its March 2026 report sampled one million Claude.ai conversations and one million first-party API transcripts. A few findings stand out: coding remains central, but Claude.ai use has become less concentrated; the top ten O*NET tasks fell from 24 percent of traffic in November 2025 to 19 percent in February 2026. Anthropic also reports that about 49 percent of jobs have seen at least a quarter of their tasks performed using Claude, while global per-capita usage remains highly concentrated in a small set of countries.

OpenAI has a similar but broader consumer-facing measurement effort. The NBER working paper How People Use ChatGPT, written by OpenAI researchers and David Deming, studies ChatGPT usage through July 2025. It estimates that ChatGPT had reached around 10 percent of the world’s adult population, finds that non-work usage grew to more than 70 percent of all use, and shows that practical guidance, information seeking, and writing account for nearly 80 percent of conversations. Work use is more concentrated among educated users in higher-paid professional occupations.

These datasets are important because they observe behavior, not just capability. They tell us what people actually ask AI systems to do. They can reveal shifts that surveys miss: coding moving from chat interfaces into APIs, personal use growing faster than workplace use, or task categories becoming more diverse as the user base broadens.

The weakness is selection. Claude data is not the AI economy; it is Claude. ChatGPT data is not the AI economy; it is ChatGPT. Platform usage has no clean denominator for “all possible users” or “all work done.” A conversation is also not the same as a completed economic task. A user may ask for code, reject the answer, and do the work manually. A firm may use an API call in a production workflow, or it may be a prototype nobody depends on. Platform data is the closest thing we have to revealed behavior, but it still needs context.

3. Worker surveys: who says they use AI?

The third family asks people directly.

Bick, Blandin, and Deming’s “The Rapid Adoption of Generative AI” is the key worker-side reference. It finds that 23 percent of employed U.S. respondents had used generative AI for work at least once in the previous week, and 9 percent used it every workday. Relative to the timing of earlier mass-market technologies, work adoption was about as fast as the personal computer, while overall adoption was faster than either PCs or the internet. At the same time, the authors estimate that only 1 to 5 percent of all work hours are currently assisted by generative AI, with reported time savings equivalent to about 1.4 percent of total work hours.

This combination is crucial. Many people have touched the technology; much less work has been reorganized around it.

Surveys are valuable because they have denominators. They can tell us who uses AI by age, education, occupation, sector, and work arrangement. They capture tools beyond a single platform, including informal use that firms may not see. But they also have familiar weaknesses: self-reporting, vague definitions, recall bias, and the difference between “I used AI” and “AI changed my job.”

Worker surveys are therefore best read as adoption breadth. They tell us how widely AI has entered people’s routines. They say less about depth, quality, and organizational consequence.

4. Firm data: has AI entered production?

The fourth family asks firms. This is where adoption looks much slower, at least when the bar is actual use in producing goods and services.

The U.S. Census Bureau’s Business Trends and Outlook Survey is one of the strongest real-time sources. The NBER paper Tracking Firm Use of AI in Real Time uses BTOS data from September 2023 to February 2024 and finds that bi-weekly estimates of business AI use rose from 3.7 percent to 5.4 percent, with expected use around 6.6 percent by early fall 2024. Common uses included marketing automation, virtual agents, and data/text analytics. Many AI-using firms reported training staff, developing new workflows, or buying cloud services, while few reported employment reductions.

More recent firm-level survey work raises the bar in a different way. The BFI working paper Firm Data on AI surveys nearly 6,000 senior executives in the U.S., U.K., Germany, and Australia. It reports that 69 percent of firms actively use AI, but also that executives who use AI average only 1.5 hours a week and that nine in ten report no impact on employment or productivity over the last three years. The forward-looking expectations are more optimistic than the realized effects.

Firm data is the bridge between individual experimentation and economic transformation. It can capture whether companies have changed processes, purchased infrastructure, trained workers, or embedded AI in operations. It also exposes a core tension in the adoption debate: high reported use can coexist with low measured impact.

The main problem is definitional. If one employee uses ChatGPT for a memo, is the firm an AI adopter? If a company has a pilot, is that production? If a vendor product quietly embeds an AI feature, has the firm adopted AI or just bought software? Firm surveys are indispensable, but the question wording matters enormously.

5. Readiness and diffusion: who is positioned to benefit?

The fifth family moves from workers and firms to countries.

The IMF’s AI Preparedness Index compares economies across digital infrastructure, human capital and labor-market policies, innovation and economic integration, and regulation. Its purpose is not to count chatbot users. It asks which countries are structurally positioned to absorb AI, manage transition costs, and convert exposure into broad gains.

Microsoft’s Global AI Adoption in 2025 report takes a more direct diffusion approach, using aggregated and anonymized telemetry adjusted for device access, internet penetration, OS and device market share, and population. It estimates that global generative AI adoption reached 16.3 percent of the world’s population in the second half of 2025, up from 15.1 percent in the first half. It also finds a widening divide: 24.7 percent adoption in the Global North versus 14.1 percent in the Global South.

These measures matter because AI adoption is not just a workplace story. It is also an infrastructure story, a skills story, a regulatory story, and a development story. The same model can produce very different economic effects depending on broadband access, education systems, firm capabilities, data availability, language coverage, managerial quality, and trust.

The weakness is aggregation. Country indices can hide large within-country gaps. They also depend on weighting choices: how much should regulation matter relative to infrastructure, or human capital relative to cloud availability? Still, for macro analysis, these indices provide the frame that platform and worker data cannot.

One of the most interesting recent directions comes from behavioral and social research. Adoption is not only a response to measured productivity. It is also shaped by fear of falling behind, peer behavior, managerial pressure, and status competition.

In Social Dynamics of AI Adoption, Leonardo Bursztyn, Alex Imas, Rafael Jimenez-Duran, Aaron Leonard, and Christopher Roth study parental demand for unrestricted AI tools in education. The exact setting is not the workplace, but the mechanism is important: demand rises strongly when parents believe more peers are using the technology. Information about possible harms changes beliefs and increases support for bans, but does not necessarily reduce individual demand. The fear of being left behind can sustain adoption even under uncertainty.

That point generalizes. AI diffusion may not wait for clean evidence of productivity gains. People and firms may adopt because others adopt, because clients expect it, because workers experiment informally, or because managers want to signal modernity. This helps explain why adoption measures can look hot while productivity statistics remain cold.

7. What the evidence says so far

The strongest current conclusion is not “AI changes everything” or “AI changes nothing.” It is more specific:

AI has diffused unusually fast at the individual level.
Real usage is concentrated in knowledge-work tasks, especially writing, coding, information seeking, practical guidance, analysis, and support.
Use is broadening, but it remains uneven across countries, occupations, income groups, firms, and workflows.
Individual use is ahead of organizational transformation.
Technical exposure is much larger than observed adoption.
Measured labor-market and productivity effects are still modest, noisy, or not yet visible at aggregate scale.

Anthropic’s Labor Market Impacts of AI paper captures the state of the field well. It combines O*NET tasks, theoretical exposure estimates, and real Claude usage into an “observed exposure” measure. The result is a better bridge between what AI could do and what people are actually using it for. Even there, the authors find no systematic increase in unemployment for highly exposed workers since late 2022, though they do find suggestive evidence that hiring into exposed occupations has slowed for younger workers.

That is the present contradiction. The usage data is real. The exposure is real. The organizational and macro effects are not yet settled.

8. Where measurement needs to go next

The next generation of AI adoption data should measure depth, not just contact.

We need to know whether AI is used once or every day; whether it supports a peripheral task or a core workflow; whether output quality improves; whether human verification remains central; whether firms redesign data models and operating procedures around agents; whether teams maintain the systems they build; and whether the gains accrue to experts, novices, firms, workers, or customers.

In other words, the key variable is not adoption alone. It is conversion: the conversion of access into capability.

This is where the current data landscape points. Exposure indices tell us where the opportunity space is. Platform data tells us what people are trying. Worker surveys tell us how broadly the tools have entered everyday practice. Firm surveys tell us whether organizations have moved beyond experimentation. Readiness indices tell us which countries have the complementary systems to benefit. Social research tells us why adoption can accelerate before evidence is settled.

The coming question is whether people, firms, and countries can build the layer above the model: the data structures, workflows, governance, maintenance routines, and human capabilities that let AI work with us reliably.

That is also where the most interesting research begins. AI may enable human potential, but it does not do so automatically. The opportunity space is large. The trajectories will be uneven. The job of measurement is to see the difference.