AI User Share: Tracking AI Adoption

• AI User Share is a quantitative metric that represents the proportion of working-age individuals using AI tools, derived from anonymized telemetry and device scaling.
• It aggregates desktop and mobile usage data, adjusts for overlaps, and normalizes across populations to provide a robust indicator of AI diffusion.
• Empirical findings show a strong correlation with GDP per capita and highlight latent AI demand in low-income regions, aiding policy and digital inclusion efforts.

AI User Share is a quantitative metric denoting the fraction of a given country’s working-age population that actively uses AI tools. Developed to provide a population-normalized, cross-national indicator for tracking AI diffusion, AI User Share has gained prominence as both an empirical construct in recent measurement science and as an evaluative tool in AI policy, economics, and digital inclusion research. Unlike traffic-based or device-based proxies, AI User Share combines anonymized usage telemetry with device and connectivity adjustments to estimate actual societal uptake of AI-driven services and systems across economies.

1. Formal Definition and Mathematical Structure

AI User Share (often abbreviated as AUS) for country $c$ and period $t$ is mathematically defined as the proportion of individuals of working age (15–64) who, within a representative time window (typically one month or rolling quarter), meet a prescribed threshold of AI tool usage. In practice, this is operationalized by aggregating AI-active user counts from device telemetry (e.g., Microsoft Windows event logs) and scaling for device penetration and platform modality. Formally,

$$\mathrm{AUS}_{c,t} = S^{\mathrm{desk}}_{c,t} + S^{\mathrm{mob}}_{c,t} - S^{\mathrm{overlap}}_{c,t}$$

where:

• $S^{\mathrm{desk}}_{c,t}$ = estimated desktop AI user share,
• $S^{\mathrm{mob}}_{c,t}$ = estimated mobile AI user share,
• $$S^{\mathrm{overlap}}_{c,t} = S^{\mathrm{desk}}_{c,t} \times S^{\mathrm{mob}}_{c,t}$$ (assumes independence).

Desktop share is computed as:

$$S^{\mathrm{desk}}_{c,t} = \gamma^{*}_{c,t} \times D_c \times M_c$$

$\gamma^*_{c,t}$ is the opt-in–adjusted share of Microsoft desktop users in $c$ who use AI for ≥90 minutes/month, $D_c$ is the desktop penetration scaling factor, and $M_c$ corrects for mobile usage via observed traffic ratios. Opt-in adjustment for telemetry participation is handled by blending country and global means; device and traffic normalization ensure comparability across heterogeneous environments (Misra et al., 4 Nov 2025). The resulting AUS metric lies in the interval [0,1].

2. Data Sources, Processing, and Coverage

The current primary deployment of AI User Share is based on Microsoft anonymized telemetry, covering 147 countries/regions. This telemetry captures visits (or sustained usage) to a curated set of major AI platforms (19 in the cited paper), with users filtered to those meeting/exceeding a threshold (e.g., ≥90 minutes/month aggregated usage) to discard sporadic or accidental users.

Device penetration for scaling is derived using monthly active device (MAD) counts, calibrated against industry market-share data (e.g., StatCounter Windows OS share). Mobile-to-desktop traffic ratios, drawn from StatCounter, extend the metric to environments where PCs are rare and mobile is primary. Population and internet connectivity baselines are taken from authoritative sources (World Bank WDI, ITU, CEIC).

Coverage is restricted to geographies with robust telemetry and populations >2 million; small or data-sparse countries are allocated to regional aggregates, with opt-in blending and imputation strategies applied where necessary (Misra et al., 4 Nov 2025, Misra et al., 4 Nov 2025).

3. Empirical Findings: Disparities and Correlates

AI User Share measurements reveal pronounced disparities, both across and within regions:

Region/Economy	AUS (2025, %)
World Average	~15
United Arab Emirates	59.4
Singapore	58.6
North America	~27
Europe & Central Asia	~22
South Asia, Sub-Saharan Africa	<13
Low-Resource Lang. Countries	~9.9
Non-LRLCs	~21.3

A robust correlation is found between AUS and GDP per capita (Spearman ρ=0.83, p<10⁻⁶). However, analysis conditional on internet penetration exhibits high “latent demand” in lower-income contexts: in countries with the lowest internet access, AUS among connected users can reach 23%, even as the population-wide AUS remains below 10%. This suggests infrastructure, rather than interest, is often the primary constraint.

Product launches can dramatically increase national AUS; for example, following DeepSeek’s introduction in China (Jan 2025), China’s AUS rose from ~8% to 20% in months, an absolute gain exceeding 100 million users (Misra et al., 4 Nov 2025).

4. Methodological Controls and Causal Analysis

AI User Share is now central to causal studies of AI diffusion. For example, in “AI Diffusion in Low Resource Language Countries” (Misra et al., 4 Nov 2025), AUS is the dependent variable in fractional logit GLMs estimating the independent effect of low-resource language status. Key controls include log GDP per capita, electricity/internet access, and age structure. ATT-weighted estimation counters confounding via propensity-score–weighted log-likelihood:

$$\ell(\alpha, \tau, \gamma) = \sum_{i} w_i [Y_i \log \mu_i + (1-Y_i) \log(1-\mu_i)]$$

with

$$\mu_i = G(\alpha + \tau D_i + \gamma^T X_i), \quad G(z) = \frac{\exp(z)}{1+\exp(z)}$$

and ATT weights $w_i$ adjusted for treatment status and propensity $PS_i=P(D_i=1|X_i)$. This approach separates language effects from infrastructure and demographics.

Model-adjusted results show LRLCs have a ~2.07 percentage point lower AUS (95% CI [–3.76, –0.38]) out of a baseline of ~10%, corresponding to a 21% shortfall attributable to language resource status.

5. Limitations, Biases, and Interpretative Caveats

AI User Share, as constructed, is subject to several sources of bias:

• Telemetry capture is desktop-centric and excludes non-Windows and many mobile-only platforms.
• Opt-in rates vary, and blending may understate or overstate true rates in privacy-sensitive settings.
• Scaling assumptions (e.g., independence of desktop and mobile use, linear device-normalization) can compress or inflate coverage at the distributional tails.
• Absence of telemetry for some regions (e.g., Russia, Iran) necessitates imputation or incurs coverage gaps.
• Reliance on a single provider’s telemetry introduces sample selection concerns; generalization assumes Microsoft user behaviors reflect those of the broader population.

Such biases must be considered when interpreting cross-country or cross-group differences, and comparisons with survey-based or alternative-tracker-based metrics should account for these factors.

6. Policy and Research Applications

AI User Share now functions as a core benchmark for multiple stakeholders:

• Governments and multilaterals use AUS to track the effectiveness of interventions (e.g., broadband rollout, AI education, device subsidies).
• Researchers employ AUS in regression and difference-in-difference designs to estimate policy, infrastructure, or language impacts on diffusion.
• Observed rapid elasticity to product launches (e.g., DeepSeek) demonstrates utility for timely market, event, or regulatory sensitivity analyses.
• Analyses show that latent demand in digitally connected but low-income populations underscores the need for universal device and connectivity access alongside linguistic and digital-literacy adaptations.

AUS’s population-level focus distinguishes it from user-count, traffic-based, or device-based alternatives, enabling direct mapping onto epidemiological or economic models of innovation diffusion.

7. Future Directions and Metric Evolution

Ongoing work aims to integrate more comprehensive telemetry, including cross-platform (non-Windows, app-based) sources and granular demographic overlays. This addresses current platform biases and expands AUS’s representativeness. Additional research is incorporating AUS into benchmarking suites for “AI readiness” and as an evaluation axis in global AI policy frameworks. There is also a push for open-source tracking and standardized reporting to permit greater scrutiny and methodological consistency across researchers, regions, and policy-makers (Misra et al., 4 Nov 2025, Misra et al., 4 Nov 2025).

This suggests the role of AI User Share will continue to expand as global AI adoption intensifies and as demand for actionable, timely diffusion metrics grows across sectors and regions.