AI Diffusion in Low Resource Language Countries

Amit Misra, Syed Waqas Zamir, Wassim Hamidouche · 2025

Read AI Diffusion in Low Resource Language Countries by Amit Misra, Syed Waqas Zamir, Wassim Hamidouche (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-02752.txt.

Introduction

Amit Misra and four colleagues ask whether weak language support slows the spread of AI in some countries. Frontier language models perform poorly on low-resource languages because training data is scarce. In a 2025 paper, the team tests whether that gap cuts the value people get from AI and so holds back adoption.

Using a weighted regression, they separate the language effect from income, education, and other social factors. Countries whose main languages are low-resource had a share of AI users about 20 percent lower than their baseline would suggest. The authors read this as proof that language access is a real, independent barrier. Closing it, they argue, is part of making AI growth fair across the world.

Catalog record

CREATORAmit Misra, Syed Waqas Zamir, Wassim Hamidouche
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE17 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-02752

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

AI Diffusion in Low Resource Language Countries Amit Misra, Syed Waqas Zamir, Wassim Hamidouche Abstract Artificial intelligence (AI) is diffusing globally at unprecedented speed, but adoption remains uneven. Frontier Large Language Models (LLMs) are known to perform poorly on low-resource languages due to data scarcity. We hypothesize that this performance deficit reduces the utility of AI, thereby slowing adoption in Low-Resource Language Countries (LRLCs). To test this, we use a weighted regression model to isolate the language effect from socioeconomic and demographic factors, finding that LRLCs have a share of AI users that is approximately 20% lower relative to their baseline. These results indicate that linguistic accessibility is a significant, independent barrier to equitable AI diffusion. 1 Introduction Artificial intelligence is the latest general‑purpose technology and is diffusing across the globe at historic speed [ 1 ] . However, as with past waves of technology like electricity and the internet, adoption is not uniform across countries. Familiar patterns of digital inequality are emerging: countries with stronger economies, higher education levels, and better digital infrastructure showing higher and earlier AI adoption [ 2 , 3 ] . A key difference with AI, however, is that language is not merely an interface, it is the substrate of the technology’s functionality. Modern AI, especially large language models (LLMs), relies on learning from vast amounts of text data. The open web, a primary source for training LLMs, is heavily skewed toward a few high-resource languages. Nearly half of all web content is in English [ 4 ] , only 5.1% of the global population are native English speakers [ 5 ] . Other major languages (such as Spanish, French, German, Russian, Mandarin Chinese) have roughly 10× less web data than English, yet still enjoy a strong digital presence [ 4 , 6 ] . Among these, languages in the Indo-European family, in particular, may also benefit from cultural and linguistic similarities with English. In contrast, the overwhelming majority of the world’s \sim 7,000 languages are low-resource, with minimal to no digital footprint [ 6 ] . The imbalance in data directly translates into LLM performance disparities. Multilingual benchmark tests consistently find that LLMs perform poorly on low-resource languages (LRLs) relative to high-resource ones [ 7 , 8 , 9 ] . In this paper, we present what is, to our knowledge, the first country‑level assessment of how language resourcing relates to AI diffusion. We classify countries by the prevalence of low‑resource languages—defining (LRLCs) via a tiered scheme that captures language resourcing and dominant language(s) in a country—and merge this classification with usage telemetry from 147 countries/economics [ 10 ] , while controlling for socioeconomic and demographic covariates (GDP per capita, electricity access, internet penetration, and age structure). Using cross‑country regressions and per‑capita usage metrics, we compare AI adoption in LRLCs versus higher‑resource‑language countries. Our results reveal a sizable language-linked gap: in raw terms, LRLCs exhibit less than half the per-capita use of AI tools. After controlling for covariates, we estimate an \sim 20% lower share of AI users is attributable to language factors. We also analyze recent trends to assess whether this gap is narrowing as systems improve. Taken together, the findings indicate that while income and digital infrastructure are necessary, they are not sufficient for inclusive AI diffusion; linguistic accessibility must be addressed. We conclude with implications for research and policy to promote AI adoption with—rather than at the expense of—marginalized languages. 2 Methods 2.1 Defining Low Resource Language Countries 2.1.1 Language Classification To quantify AI diffusion readiness by country, we first construct a taxonomy of languages based on the availability of digital content and performance in frontier LLMs. We leverage the FineWeb2 multilingual dataset [ 6 ] , which contains text corpora covering over 1,000+ languages, to categorize each language into one of three groups: • High-resource languages: This category comprises languages with large-scale digital presence and strong representation in existing LLMs. Examples include English, Spanish, French, German, Italian, Russian, Portuguese, Mandarin Chinese and Japanese. • Mid-resource languages: These are languages with moderate amounts of text data available online. LLMs can perform reasonably in these languages, however not as accurately as in high-resource languages. Examples include Arabic, Hindi, Bengali, Polish, Dutch, Indonesian, Vietnamese, Persian, Turkish, Thai, Korean, Ukrainian, Greek, Czech, Swedish, Hungarian, Danish, Finnish, Hebrew, Malay, and many others. • Low-resource languages: This category encompasses the majority of the world’s languages, those with very limited textual data, or none at all. Examples include Chichewa, Inuktitut and Guarani. 2.1.2 Country-Level Classification With the above language taxonomy defined, we assign each country an “AI diffusion readiness” label (High, Intermediate, Low) based on its dominant national language. Our primary reference for country languages is the CIA World Factbook [ 11 ] , which provides (regularly updated) information on official and widely spoken languages in each country. For each country, we apply the following steps. (1) Language extraction: We retrieve the languages field from the World Factbook. Since this language field contains unstructured data, we use GPT-5 reasoning model as parser to extract the dominant language name for that country. (2) Map the dominant language to a resource category: We check which of our three categories the identified language fell into. If the language is in our predefined high resource list, the country is labeled as a High-Resource Language Country (HRLC). If the language is in the Intermediate list, the country receives Mid-Resource Language Country (MRLC) label. All other countries default to Low-Resource Language Country (LRLC) category. 2.2 Estimating the Impact of Low-Resource Languages on AI Diffusion To estimate the impact of LRLC status on AI adoption, we model the relationship between the two using AI User Share, a Microsoft metric estimating the proportion of the working-age population using AI per country [ 10 ] . Our primary estimate comes from a fractional logit Generalized Linear Model (GLM), weighted to produce the Average Treatment effect on the Treated (ATT) [ 12 ] . This model is well-suited for a bounded outcome (a percentage) and estimates the change in AI usage if an LRLC were to become a non-LRLC. In all models, we control for GDP per capita (log scale) [ 13 ] , electricity access [ 14 ] , internet access [ 15 ] , and age structure [ 11 ] . Continuous covariates are standardized before propensity score estimation to stabilize optimization. To ensure our findings are robust, we compare the GLM results against several alternative estimators, including Inverse Probability Weighting (IPW), Augmented IPW (AIPW), and a standard Ordinary Least Squares (OLS) regression [ 16 , 17 , 18 ] . As shown in Table 1 , these methods consistently point to a meaningful negative impact for LRLCs and are consistent with the ATT-weighted GLM results. Table 1: Comparison of Treatment Effect Estimates for Low-Resource Language Status Method Estimate Std. Error 95% CI ESS OLS -1.88 0.96 [-3.77,\ \phantom{-}0.01] – ATT (IPW) -2.32 0.80 [-3.89,\ -0.75] 21.1 ATT (AIPW) -1.69 0.91 [-3.48,\ \phantom{-}0.09] 21.1 ATT-weighted GLM -2.07 0.86 [-3.76,\ -0.38] 21.1 • Effects are in percentage points on AI User Share. ESS = Kish effective sample size for controls. CIs use robust (HC3) SEs for OLS and ATT‑weighted GLM and percentile bootstrap for ATT estimators.

Read the complete text