The AI Literacy Divide: Access, Capability, and Inequality in the Age of Generative AI

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The rapid diffusion of artificial intelligence (AI), particularly generative AI (GenAI), has shifted the digital-inequality question from whether people can access digital technologies toward whether they can understand, evaluate, and use AI effectively. Yet these dimensions remain fragmented across datasets and measurement approaches. This study develops an integrated measurement framework for examining the emerging AI literacy divide across six conceptually related stages: digital access, AI access, AI literacy, demonstrated capability, effective use, and realised benefit. The framework is used as an analytical architecture rather than as an empirically demonstrated causal pathway. The empirical analysis combines secondary-data quantitative analysis with structured cross-dataset evidence synthesis. Individual-level analysis is conducted using 1,205 observations from the FITPED AI-literacy dataset, representing 1,146 unique participants, while Eurostat, ITU, World Bank, UNESCO, OECD PIAAC, IEA ICILS, Pew Research Center, Stanford AI Index, ILO, and selected scholarly studies provide independent contextual and population-level evidence. FITPED measures perceived or self-reported AI literacy rather than demonstrated GenAI capability. The five-item AI-literacy scale has a mean of 3.991 on a five-point scale, a standard deviation of 0.676, and Cronbach’s alpha of 0.688. Perceived AI literacy is positively associated with AI readiness (ρ = .580), relevance (ρ = .484), career motivation (ρ = .450), and behavioural intention (ρ = .407), and negatively associated with AI anxiety (ρ = −.165). Male participants and participants in IT programmes report higher AI-literacy scores than their respective comparison groups. The findings indicate that current evidence can illuminate distinct components of AI inequality but does not yet support a complete individual-level model linking access, capability, effective use, and realised benefit.