The access–conversion gap across countries: how generative AI adoption exceeds learning use

1 Introduction

Generative artificial intelligence (AI) has moved into education with remarkable speed since late 2022 (Achuthan et al., 2026; Kasneci et al., 2023; Jyothy et al., 2024). Learners can now have access to systems that explain code, draft solutions, and answer questions on demand. Computer science education sits at the centre of this change, as programming tasks map closely onto what these tools do well (Prather et al., 2023; Raman et al., 2024). The pace of adoption has outrun the evidence about how people learn. This paper addresses one part of that gap: the distance between reaching such a tool and using it to learn.

Computer science is an unusually suitable discipline for this question. The field has a large, globally distributed population of learners, many studying outside formal institutions through online courses and peer communities (von Garrel and Mayer, 2023). Programming also generates frequent, well-defined tasks, such as fixing an error or drafting a function, on which an assistant can act directly (Yilmaz and Karaoglan Yilmaz, 2023). These features make the use of AI for learning observable at scale and across very different national contexts, and they speak to a population that policy on digital skills increasingly seeks to reach.

Commentary on the change has settled into two accounts. The optimistic one holds that generative AI widens access to expert-like support and thus narrows long-standing divides (Chan and Hu, 2023). The cautious one holds that the same tools track existing advantage, so early and skilled adopters in richer settings gain more (Wang et al., 2024). Population evidence supports the cautious account at the level of access: adoption rises with income, education, age (Bick et al., 2024), and, at the country level, with national income and infrastructure (Liu and Wang, 2024). The divide is real.

Both accounts, however, treat adoption as the outcome that matters. Adoption is a thin measure of what a learning technology does (Daepp and Counts, 2025). A learner can reach a tool and use it only at the surface, and the value obtained varies sharply with how skilfully it is used (Dell'Acqua et al., 2023). Whether a tool is reached and whether it is used for learning are separate questions, and the second depends more directly on competence. Reading adoption alone can mislead since it counts having a tool, rather than using it, as learning.

This paper addresses this distinction for the case of computing learners. Where access is already broadly shared, the question is not who reaches these tools but who turns reaching them into using them for learning and whether that conversion varies across income and region. Access becomes a precondition, and conversion becomes the object of study.

This focus sets the paper apart from two adjacent literatures and fixes its scope. A newer strand asks whether a divide is emerging in the use of AI rather than in access alone (Suarez and Garcia-Marinoso, 2025). By design, the paper studies a population for whom first-level access is broadly shared so that any divide in use and outcomes can be examined on its own (van Deursen and Helsper 2015). Reading returns places the study at the third level of the divide (Ragnedda, 2017). The scope is the connected population of computing learners.

The paper develops a construct that names the distance between reaching a tool and using it for learning, then asks what widens that distance and what follows from it. The construct is the access-conversion gap. The reasoning is abductive, moving from the pattern to a candidate explanation (Brandt and Timmermans, 2021). Access is operationalised as everyday use of AI, and conversion as the breadth of AI use for studying computer science. It indexes the capability to turn a tool into learning, which the capability approach formalises (Sen, 1999).

The framing matters most for learners in lower-resource settings, where the stakes are highest. If access were the binding constraint among the engaged, the task would be to place tools before those who lack them. Where access is instead broadly shared, the task shifts to ensuring access can be used, and what then drives use is the value and ease learners perceive (Ivanov et al., 2024). Experimental work suggests generative AI can help less-experienced users most (Noy and Zhang 2023), a pattern echoed among lower-skilled workers (Brynjolfsson et al., 2025). An assets-based reading locates resources in learners that dominant tools may overlook (Yosso, 2005).

The paper makes four contributions. It reframes the question of AI and learning from access to use, asking whether a divide can persist among learners who already reach these tools. It introduces the access-conversion gap as a measurable construct at the second and third levels of the digital divide, with an explicit operationalisation. It brings a multilevel and configurational analysis to the construct, separating whether adoption translates to learning use from how widely it is then used. It provides a cross-national account of where conversion fails, showing through a learner typology that the disengaged are a minority present across every income tier.

Three research questions organise the analysis:

RQ1. Does everyday adoption of AI vary across income and regional context once individual characteristics are held constant?

RQ2. Among those who adopt, does conversion to any learning use, and its breadth among converters, differ across income and regional context?

RQ3. Does the association between AI use and self-rated proficiency differ across context, and in which direction?

The paper proceeds as follows. The next section sets out the theoretical framework and defines the access-conversion gap. The methods section describes the data, measures, and multilevel and configurational strategy, with attention to weighting, missing data, and the selected sample. The results and discussion section reports the pattern by income tier and region, the models for each question, a learner typology, and the conditions behind a wide gap. The latter sections describe the implications, limitations, and conclusions.

2 Theoretical framework

2.1 Access, use, and functioning

Early work framed the digital divide as a binary of access, and later research showed that access alone explains little about usage once most people are connected (Lythreatis et al., 2022). A second-level divide concerns differences in skills and in the ways people use a technology once they hold it. A usage gap describes how groups with similar access still put a technology to systematically different uses (Scheerder et al., 2017). A third-level divide concerns the returns people obtain, so that the same activity yields unequal outcomes across groups (Ragnedda, 2017). Across these levels, inequality migrates from access to what people do with a tool and to what they gain from it (van Dijk, 2020).

The tradition also warns that new tools can widen gaps, and that disadvantage can be reproduced even after connection (Heeks, 2022). Those already advantaged in terms of skills and support tend to extract more from each new resource, so benefits accumulate where capacity already exists. Studies of internet use have repeatedly found such reinforcement, as the divide shifts from physical access to the material and skill conditions of use (van Deursen and van Dijk, 2019). Optimistic accounts, by contrast, treat availability as the key variable and infer benefit from reach (Bearman et al., 2023). The digital-inequality literature shows that this inference is unsafe because the decisive variation sits at the level of use and outcome.

The capability approach distinguishes a resource from the functioning that the resource enables and identifies the difference in conversion factors (Robeyns, 2005; Sen, 1999). Conversion factors are the personal, social, and environmental conditions that let a person turn a resource into something they can do or be, a lens that carries directly into technology and development (Kleine 2013). Applied here, an AI assistant is a resource, and the functioning of learning depends on a usable device, a connection, a shared language, and support for study. A learner may hold the resource and convert little of it, which is a conversion failure.

2.2 Why does adoption happen?

Theories of technology acceptance explain why adoption is often high, even where the divide literature would expect friction, and the field of AI in higher education has grown quickly around this question (Crompton and Burke, 2023). The unified theory (i.e., UTAUT) attributes use to performance expectancy, effort expectancy, social influence, and facilitating conditions (Venkatesh et al., 2003). A generative assistant scores well on the first two since it lowers the effort of many learning tasks and appears to deliver useful output at once (Shahzad et al., 2025; Strzelecki, 2024). This helps explain a pattern that puzzles the divide account, in which adoption is widespread across very different settings (Dwivedi et al., 2023). Acceptance theory is better at explaining that a tool is taken up than at explaining what use then follows, and considerations for practice make that gap explicit (Farrelly and Baker, 2023). Facilitating conditions enter as a predictor of use, yet the theory says little about the distance between adopting a tool and using it well. That distance is where the present construct sits, and it is why adoption and conversion are treated as separate outcomes throughout.

2.3 Why does conversion fail?

Two literatures explain why conversion can fall short even where adoption is high. The first concerns the regulation of learning. Research on hybrid human-AI regulation shows that a learner can transfer regulation to a tool and complete a task without building the self-regulation that durable learning requires (Molenaar, 2022). Related work describes how ready answers can reduce the monitoring and effort that learning depends on, an instance of cognitive offloading (Liu et al., 2026). Use that looks active can therefore still fail to convert into competence (Gerlich, 2025), and the risk rises where guidance is thin (Fan et al., 2025). The point is not that AI use is inert, but that use and learning can come apart, which is precisely what a measure of conversion should detect.

The second literature concerns epistemic fit. Decolonial scholarship argues that tools built on non-local data and norms carry defaults that suit some learners poorly (Zembylas, 2023). A learner whose language, examples, and context sit outside a tool's training may convert less of what they access, even at equal adoption (Mohamed et al., 2020). An assets-based reading locates resources in learners and communities that dominant tools overlook and treats conversion as contingent on whether a tool meets those resources (Yosso, 2005). Read together, the regulation and the epistemic accounts predict that conversion depends on facilitating conditions and on fit and that the gap should be widest where devices, support, and linguistic match are weakest.

2.4 The access-conversion gap

These strands converge on the need to understand the access-conversion gap. This gap is the distance between access and conversion for a learner or a group. For a population that is already connected, physical access to generative AI is largely settled, so it is not what separates learners. The relevant lower rung is whether a learner actually reaches the tool in regular use, which this study treats as effective access. Conversion is the further step from reaching the tool to putting it to use for learning. Conversion is a behavioural threshold, the entry into learning use, and not a measure of learning gained. The access-conversion gap is the distance between these two steps, between reaching a tool and using it to learn, in the sense the capability approach makes precise, where a resource yields a functioning only if a person can convert it (Robeyns, 2017; Sen, 1999). The digital-divide literature places the construct at the second and third levels since it concerns use and its returns (van Deursen and Helsper, 2015). Acceptance theory explains the high adoption that makes the gap visible, and the regulation and decolonial literatures supply the mechanisms that widen it. Figure 1 shows how the elements combine, from the deductive anchors that define the construct, through an inductive typology that populates it, to a configurational step that locates the conditions behind it.

The construct is intended as measurable and carries a validity claim that the analysis examines (Flake and Fried, 2020). The claim is that breadth reflects the range of learning purposes a tool is put to. Two features guard against reading engagement as conversion. Breadth is defined over distinct learning purposes, so it rewards range. A decoupling indicator marks adoption without any learning use, which isolates the clearest case of a wide gap. The construct is offered as middle-range theory, close enough to measurement to be tested and general enough to travel beyond the present setting (Merton, 1968). The access-conversion gap is related to but distinct from the usage gap. The usage gap describes different kinds of use among those with similar access. The access-conversion gap concerns a prior threshold, whether access becomes any learning use at all, and locates that threshold in the capability to convert (Scheerder et al., 2017).

The construct is distinct from three related concepts. It is not the first-level divide because access is held roughly constant among the adopters the gap compares. It is not general engagement, because breadth is defined over distinct learning purposes and is paired with the decoupling indicator. It is not proficiency, which is treated as a separate outcome that the gap may predict. Keeping these apart is what lets the paper question whether a wide gap carries consequences.

3 Data and methods

3.1 Data and scope

The analysis draws on a publicly archived international survey of computer science learners and practitioners (Dzialets et al., 2025). The survey covers 18,032 respondents. Data was collected in the first half of 2024. The instrument records a completion status and a survey weight for each response. The analytic sample comprises respondents who reached the section on AI use, defined by a non-missing answer to the everyday-use item. This rule yields 14,653 respondents across 166 countries. Respondents who did not reach that section are excluded from the AI outcomes by construction. Table 1 summarises the sample. Respondents are predominantly male, at 84 percent, and early in their careers. About two-thirds are under 30, and a third report no professional coding experience. Roughly half are formal students, and two-thirds use AI in everyday life. Educational attainment centres on a bachelor's degree. The skew toward male and early-career respondents follows from the recruitment channels and is noted among the limitations. One respondent is from an economy without a World Bank income classification and is therefore excluded from analyses by income tier, which use 14,652 respondents.

Table 1. Sample characteristics (weighted, n = 14,652).

The inclusion rule and the weighting decision are fixed in advance and applied identically across every stage. Missing covariate data are addressed by multiple imputation, so the regression models retain the full analytic sample of 14,653 respondents, of whom 9,537 are everyday users. Figure 2 describes the analysis stages. Countries were assigned to eight regions following the World Bank scheme, with some labels shortened for brevity (BIC for Bayesian Information Criterion); the full country-to-region and country-to-tier mapping is given in the Supplementary Table S2.

The survey is opt-in, and its documentation records two recruitment channels. External responses were gathered through social-media advertising and peer referrals, and internal responses through the social accounts and email list. The sample is a population of computing learners, and the findings describe this engaged population. The scope condition is deliberate. The paper studies the divide in the use of AI among learners for whom first-level access is largely settled, so a population in which access is close to saturated is the appropriate setting.

Every respondent in the analytic sample studied computer science in the past 12 months, whether in a formal institution or through self-education, so the sample contains no work-only adopters. The decoupling measure therefore cannot be inflated by respondents who use AI outside any study context since all are active learners.

Missingness on the model covariates is low. The largest fraction is education at 5.8%, with proficiency and coding experience at 2.6%, and the remaining covariates, age, gender, device, language, income, and region, complete. Table 2 records the pattern. Missing values are handled by multiple imputation, described in section 3.4. The low-income cell is small (106 respondents of whom 82 are everyday users)s so estimates specific to that cell are read as indicative.

Table 2. Missingness on model covariates.

3.2 Measures

The two focal measures operationalise the construct. Effective access is a binary indicator of everyday AI use. Conversion is a count of the distinct AI features a respondent reports using for studying computer science. This includes using AI for at least 1 of the 20 learning purposes in the survey. Breadth is the count of those purposes. The measured predictors map onto the capability approach's conversion factors. Personal factors are coding experience, education, and age. Social factors are student status and the country and regional context, entered as fixed and random effects. Environmental factors are the study device and the near-universal connectivity of the sample. Two further country indicators were considered and set aside. Gross national income per capita is the basis of the World Bank income tiers already central to the analysis, so it would duplicate the tier variable. Internet penetration is closely tied to income at the population level (almost 93% in high-income countries and 27% in low-income countries in 2024), and the sample already conditions on being connected, so penetration is near-constant within it and cannot account for variation among respondents (ITU, 2024).

In a sample of engaged learners, for whom first-level connectivity is near-universal, everyday use is the meaningful lower rung, so it stands as proxy for effective access rather than mere availability. Self-rated proficiency is recorded on a 10-point scale. At the group level, the access–conversion gap is the difference between the adoption rate and the learning-use rate, as in equation 1. At the individual level, a decoupling indicator marks a respondent who reports everyday use yet no learning use_.In Equation 1, g indexes the group, an income tier or region. The first term is the design-weighted share in group g reporting everyday AI use and the second the share reporting any AI use for learning. The gap is their difference.

Gapg=π^gaccess−π^glearn(1)

Two contextual factors structure the comparisons. National income group is coded from the country field into the four World Bank tiers, and world region into the World Bank regional grouping, both following the current classification (World Bank, 2024). Income is treated as the theoretical axis and region as the empirical granularity. The facilitating-condition proxies are a device indicator marking smartphone-reliant respondents who report no laptop or desktop, and a language-match indicator. Individual covariates are coding experience, formal education, age band, employment status, and gender, with non-response for gender retained as its own category. Table 3 maps each construct to its operationalisation, source, and role. The procedures describe how variables are constructed and do not enumerate item-level content in advance of the analysis.

Table 3. Constructs and their operationalisation.

The 20 learning-use features are listed in Table 4, grouped by learning purpose. Breadth is the count of features a respondent reports, so it rewards the range of distinct purposes a tool is put to. The measure treats the features as equal and additive, which the discussion revisits as a validity consideration.

Table 4. The 20 AI learning-use features by purpose.

3.3 Analytic strategy

Estimation proceeds in five stages. The models differ in structure rather than being three-level throughout. The adoption model is design-weighted with country-clustered errors, while the conversion, breadth, and proficiency models carry a region fixed effect and a country random intercept, and the variance decomposition uses a null three-level model. The first stage establishes the descriptive picture by income tier and by region, with design weights and country-clustered bootstrap intervals. The second stage fits the regression models for the three research questions, namely adoption, learning-use intensity among adopters, and self-rated proficiency. The third stage recovers a learner typology through latent class analysis. The fourth stage examines the conditions associated with a wide gap through a configurational analysis. The fifth stage decomposes proficiency variance across region, country, and respondent. The reasoning across stages is abductive, moving from the descriptive pattern to the construct and forward to the theory stated in the discussion (Brandt and Timmermans, 2021; Tavory and Timmermans, 2014).

Conversion is modelled as a two-part process, which separates whether an adopter converts to any learning use from how widely converters then use the tool, matching the threshold logic of the gap. The sensitivity of the conversion result to the definition of learning use is examined by repeating it at higher feature thresholds. Adoption and the use-proficiency association are additionally tested for heterogeneity by gender and age. Breadth counts how many of the 20 features a respondent reports. The usage rate of each feature among adopters is given in Supplementary Table S1 in the supplementary material. Explaining code and generating code are the most common, near 60%, and interactive simulations the least, near 10%. A value of zero denotes no reported learning use rather than missing data since the feature list was shown only to everyday AI users.

3.4 Estimation, weighting, and missing data

Two estimators are reported side by side, because no single approach in this setting captures both the survey design and the multilevel structure. The first is a design-weighted generalised linear model with country-clustered robust standard errors (Cameron and Miller, 2015). The second is a multilevel model with a country random intercept and region fixed effects (Snijders and Bosker 2012). The survey weights are calibration weights. Dzialets et al. (2025) state that they correct the country distribution of the externally recruited responses against an external benchmark to counter the under-representation of smaller and more remote countries and to improve external validity. The deposit provides a single weight column, ranging from 29 to 4,351 with a mean of 751, and no channel indicator, so the external and internal recruitment streams cannot be separated in these data. They therefore support the cross-country and cross-tier comparisons the paper makes, which is why the descriptive estimates and the design-weighted models apply them. They do not correct selection into the survey, so weighted estimates remain conditional on the engaged population that responded. Where the two estimators agree, the conclusion is reported as stable. Adoption is modelled by design-weighted logistic regression on income-tier indicators and individual covariates, education, age, coding experience, student status, and gender, with standard errors clustered by country.

Conversion is modelled as a two-part, or hurdle, process among adopters, which separates the threshold from the intensity of use. The first part is a logistic model for whether an adopter converts to any learning use, as in Equation 2. The second part is a zero-truncated negative-binomial model for breadth among converters, as in Equation 3, which addresses the overdispersion in the feature counts (Hilbe, 2011). This pairing matches the mechanism the construct proposes since the gap may sit at the threshold, at the intensity, or at both, and the two parts estimate each separately.The term t is the vector of income-tier indicators with coefficients β; x is the individual covariates with coefficients θ; β₀ is the intercept; γ is the region fixed effect; and u is the country random interceptwhere b is breadth among converters, modelled as a zero-truncated negative binomial. The linear predictor carries the tier term tβ, the covariates xθ, the intercept β₀, the region effect γ, and the country intercept u, as in Equation 2.

logitPr(convertij=1|adoptij=1)=β0+tjTβ+XijT+γr(j)+uj(2)

logE[bij|bij>0]=βo+tjTβ+XijTθ+γr(j)+uj(3)

Proficiency is modelled as a linear outcome with learning use interacted with context, as in Equation 4, which gives the cross-level interaction central to the third research question.where prof is self-rated proficiency from 0 to 10; learn is any learning use with main effect δ; β is the tier effect; κ the coefficients on the learning-by-tier interaction; θ the covariate coefficients; γ the region effect; and ε the residual.

profij=β0+δlearnij+tjTβ+KT(learnij.tj)+XijTθ+γr(j)+εij(4)

The contextual share of proficiency variance is the region-plus-country intraclass correlation from a null three-level model, that is, the share of total proficiency variance lying between regions and countries rather than within them (Snijders and Bosker, 2012). Missing covariate data are addressed by multiple imputation with 20 imputations under a missing-at-random assumption, and estimates are pooled by Rubin's rules (Rubin, 1987; van Buuren, 2018). The imputation model included income tier, region, age, gender, education, coding experience, student status, everyday AI use, and breadth of learning use. The complete-case models are retained as a reference, and the pooled estimates agree with them, as the low missing fraction would predict. Descriptive intervals come from a bootstrap that resamples whole countries before reweighting, so the intervals reflect the clustering of respondents within nations (Cameron et al., 2008).

3.5 Typology and configurational analysis

The typology is estimated by latent class analysis of the format, device, help-seeking, and feature items (Collins and Lanza, 2010). The number of classes is selected by the Bayesian information criterion, i.e., BIC (Nylund et al., 2007; Schwarz, 1978), and class separation is reported through relative entropy (Celeux and Soromenho, 1996). The classes are described by their use profiles and by their composition across income tiers, and full fit indices across the class solutions are reported in the results. The configurational analysis is a separate step. It is conducted on individual-level conditions, namely low income, coding inexperience, low education, smartphone-only study, and language mismatch, and not on the latent-class assignments. It reports necessity, sufficiency consistency, and coverage, defined as in (Ragin, 2006). Because the decoupling outcome is rare, near 3%, sufficiency consistency is bounded low by construction, so necessity and configurational structure are the interpretable quantities. A fuzzy-set analysis with formal calibration is identified as the planned extension, and the present pass is reported as descriptive set analysis.

where c is a case's membership in a condition C and d its membership in the decoupling outcome D. Consistency is the share of the condition covered by the outcome, and coverage is the share of the outcome covered by the condition.

4 Results and discussion

4.1 The access–conversion gap

Everyday adoption of AI is high across income levels, and within this engaged population it is not lower in poorer settings. That direction reflects who enters such a survey, so it is read as a property of the sample. Table 5 reports the weighted picture by income tier. Adoption sits above 59% in every tier, and learning use is a few points below. The quantity that separates the tiers is the gap, the distance between adoption and learning use, which is near three points in the upper tiers and widens to eight points in the low-income tier. Reading that distance changes what the data say. On adoption, the low-income tier looks well served since its adoption rate matches the middle-tier rates. On the gap it looks least able to turn adoption into study. The same learners are well supplied and poorly served at once, which only a measure that separates access from use can show.

Table 5. Adoption, the three use groups, the access–conversion gap, and breadth, by income tier (weighted %).

Separating the groups makes the point concrete. Every respondent falls into one of three, shown in Table 5. Some do not use AI in everyday life. Some use it every day but not for learning. Some use it every day and for learning. On the full-sample columns the low-income tier is not the least engaged, since more of its members reach and learn with AI than in the high-income tier. Its everyday-not-learning group, at 8%, is nonetheless larger than in any other tier. A learner who adopts AI but reports no learning use is decoupled. Expressed as a share of everyday users, that same group is the decoupling rate, which reaches 10.3% in the low-income tier compared to the 5.6% in the high-income tier. The two measures describe the same learners from different angles, and neither contradicts the other.

Mean breadth peaks in the middle tiers and falls at both ends, so intensity and adoption answer to different forces. This first look already reframes the question. Access does not separate the tiers in this population, so the object of study is conversion, the step from reaching a tool to using it for learning. The digital-inequality tradition reached this conclusion earlier, finding that access is a precondition rather than a benefit and that inequality migrates to use (Samuelson and Litzler, 2016; Yu et al., 2018).

A rapidly growing literature documents how learners use generative AI in computing and higher education (Chan and Hu, 2023; Kasneci et al., 2023; Prather et al., 2023; von Garrel and Mayer, 2023), including using its effects for programming skills and self-efficacy (Yilmaz and Karaoglan Yilmaz, 2023), the drivers of adoption (Ivanov et al., 2024), and its misuse (Lee et al., 2024). This study adds a question those works rarely ask, namely how such use is distributed across the world and where it fails to take hold.

The regional estimates in Table 6 show a wider spread. Adoption ranges from 52.5% in North America to 74.6% in South Asia, and the decoupling rate is highest in Sub-Saharan Africa at 6.3 percent, above the Western European figure. Breadth is widest in South Asia at 5.5 features and narrowest in Sub-Saharan Africa at 4.96, which warns against reading breadth as a simple marker of advantage. These regional differences are not artefacts of a few large countries since the intervals resample whole nations.

Table 6. Weighted estimates by world region.

Within the lower-middle-income tier alone, the decoupling rate runs from 3.2% in East Asia and the Pacific to 10.5% in Latin America, and breadth from about 4.8 to 6.4 (Figure 3). Because these regions hold income constant, the variation is not an income effect, which is why region enters the models as a fixed effect on a country's random intercept.

4.2 Adoption across context (RQ1)

The first question asks whether adoption varies across context once individual characteristics are held constant. Table 7 reports the design-weighted logistic model. Relative to the high-income tier, the adjusted odds of everyday adoption are higher in every other tier, and multiple imputations over the full sample leaves this unchanged. Figure 4 plots the tier odds ratios. A multilevel model with a country random intercept and region fixed effects reproduces the pattern geographically, with higher odds in East Asia and the Pacific, South Asia, Sub-Saharan Africa, and Latin America relative to Western Europe. The country random-intercept standard deviation is 0.34, so countries within a region still differ.

Table 7. Adoption model, design-weighted logistic regression with country-clustered errors.

The result holds across model specifications that make different assumptions, which is the robustness argument. Among the already-connected, adoption does not fall in poorer settings, but this describes the engaged population (Robinson et al., 2015). The finding that matters here is negative in a useful way. Access does not separate the tiers in this sample, so it cannot be what drives the gap, and conversion becomes the place to look.

The covariates behave as expected and reinforce the reading. Higher education and student status raise the odds of adoption, while older age bands lower them, and coding experience is flat. None of these absorbs the tier effect, so the gradient is not a by-product of who populates each tier. The multilevel estimate adds that countries within a region still differ by roughly a 1.4-fold change in the odds for a one standard deviation shift, so the regional labels summarise heterogeneous countries rather than uniform blocks. The model specifications rest on different assumptions yet agree, which makes the result robust.

This apparent tension with the population evidence resolves once the two populations are distinguished. General-public studies sample everyone, so the first-level divide dominates and adoption tracks income (Liu and Wang, 2024). This study samples the already engaged, for whom that divide is largely behind them. The two accounts are therefore complementary. One describes who reaches these tools across the world, the other what happens once reaching them is no longer the question.

4.3 Conversion across context (RQ2)

The second question asks where conversion breaks down. A two-part model separates whether an adopter converts to any learning use from how widely converters then use the tool. Table 8a reports both parts. In the first part the odds of converting at all are sharply lower in the low-income tier, at 0.49 relative to the high-income tier, while the upper-middle tier is higher and the lower-middle tier is level. In the second part, breadth among converters, the tier rate ratios sit within a few percent of one, and none is distinguishable from the high-income tier.

Table 8a. Conversion among adopters, two-part (hurdle) model.

An adopter in a low-income country is about half as likely to cross into any learning use, yet one who does cross uses the tool about as widely as anyone else. The practical consequence is direct. Closing the gap is about helping more learners cross into learning use, not about deepening the use of those already there, which is why the decoupling indicator becomes the outcome the conditions analysis explains.

The result does not depend on where the threshold is set. Table 8c repeats the decoupling comparison at higher feature cut-offs. The low-income tier has the highest decoupling at every level, from a single feature through three, and the ordering across tiers is preserved. Defining learning use more strictly raises every rate but leaves the gradient intact.

The device result points at the same threshold from another direction. In many economies smartphones are the dominant and often the only route online, so a share of learners study from a handset alone (Tsetsi and Rains, 2017). A phone suits messaging and search more than the write, run, and debug loop that studying code requires (Pearce and Rice, 2013). A learner can therefore reach an assistant for a quick answer yet struggle to fold it into practice, which is the access-conversion gap in concrete form. The device and threshold results describe the same barrier from two sides.

The low-income conversion result is robust. Table 8b collects the checks. The effect holds unweighted, at an odds ratio of 0.49, so weighting does not drive it. Removing any single low-income country in turn leaves it between 0.39 and 0.51 and significant throughout, so no one country drives it. The effective sample size is 4,454 against a design effect of 2.09. Pooling the low and lower-middle tiers returns a null, which reflects the lower-middle tier sitting near parity rather than testing the low-income effect, so the tiers are kept separate.

Table 8b. Robustness of the low-income conversion result.

Table 8c. Decoupling by tier at stricter learning-use thresholds.

4.4 AI use and proficiency (RQ3)

The third question asks whether the association between AI use and self-rated proficiency differs across context. The substitution view predicted a weaker association where the gap is wider, since shallow use should not build competence. Table 9 shows the opposite. The association is near zero in the high-income tier and grows in the lower tiers, reaching a significant half point in the lower-middle tier and a large but less precisely estimated shift in the low-income tier. The regional estimates in Figure 5 place it largest in Sub-Saharan Africa, South Asia, and East Asia, and the Pacific and near zero or slightly negative in Western Europe.

Table 9. Proficiency model, learning use interacted with income tier.

The association is largest in Sub-Saharan Africa at about 0.56 points, then South Asia at 0.40, East Asia and the Pacific at 0.39, all of which are significant. It is weaker in North America and Latin America and near zero or slightly negative in Western Europe, Eastern Europe, and Central Asia. The regions where AI use, tracks proficiency most strongly are those where formal computing education is least established, and the regions where it does not are those with dense institutional support.

One reading consistent with this is complementarity under scarcity. Where tutors, mentors, and strong local programmes are fewer, an assistant occupies a larger share of a learner's support, so its use co-moves more strongly with how competent they feel. About half a point in the lower-middle tier, significant at p < .001, and about 0.56 points across Sub-Saharan Africa are large for an observational association, and the low-income tier estimate is larger still, though its interval includes zero. It aligns with a growing experimental literature in which generative AI raises the performance of less-experienced and lower-skilled users most, compressing the skill distribution (Brynjolfsson et al., 2025; Noy and Zhang, 2023). If that mechanism carries into learning, an assistant may matter more where formal instruction is scarcer, which is the opposite of the fear that these tools chiefly help the already advantaged.

The association is not a subgroup artefact. The interaction between learning use and gender leaves it unchanged, and the interaction is not significant (b = 0.14, p = .31), and the interaction with age is likewise not significant (b = 0.05, p = .63). Decoupling is also similar across gender, at 4.5% for women and 4.6% for men. The contextual pattern is therefore not a gender or age effect in disguise.

Proficiency here is self-assessed, so the association may reflect confidence as much as competence, and a learner who uses an assistant may feel more able without being so. The design cannot separate the two, which is why the result is offered as association. The offloading literature warns of the reverse risk, that fluent AI output can substitute for the effort learning requires and leave an illusion of competence (Lee et al., 2025). The present data do not show that signature, since understanding-oriented use dominates. Self-rated proficiency is not an objective measure, and the association between AI use and proficiency is not evidence that use improves learning.

As a validity check, self-rated proficiency correlates with coding experience, at a rank correlation of 0.21, and with education, at 0.14, both stronger than its weak tie to breadth. The experience-proficiency association holds in every region, ranging from about 0.07 to 0.31, with only two of eight regions differing significantly in slope. The self-rating therefore tracks the anchors a proficiency measure should, which supports its use, while the mild cross-regional variation is a reason for the caution already noted.

4.5 What widens the gap?

A model of the decoupling indicator, and a configurational reading, locate the conditions behind a wide gap. Table 10a reports both, and no single condition is a strong predictor. The income gradient on this indicator is indicative, at an odds ratio of 1.79 for the low-income tier. Smartphone-only reliance points the same way, but its interval is wide and it does not reach significance, since smartphone-only study is rare. Language match shows no association, and the feared offloading signature does not appear (Molenaar, 2022) since feature use skews toward understanding.

Table 10a. Conditions behind the wide gap, decoupling model and configurational analysis.

The device pattern is clearer in the raw rates than in the model. Smartphone-only study is uncommon overall, at 1.6%, but concentrated in the lower tiers, at 5.4% in the lower-middle and 4.8% in the low-income tier against well under 1% in the upper tiers. The learners most likely to study from a handset alone are therefore those where the gap is widest, even though the small number of such cases leaves the regression coefficient imprecise (Pearce and Rice, 2013; Tsetsi and Rains, 2017).

The configurational analysis points to inexperience. It is the most common condition among wide-gap cases, present in about half, while low income and smartphone-only reliance are each present in only a small fraction. No single condition is necessary, and the wide gap is reached through several low-consistency routes, the widest by coverage combining low education with inexperience. Because the outcome is rare, consistency is bounded low, so the analysis is read as a map of routes. The conversion estimate in Table 8a and the decoupling estimate in Table 10a are model different outcomes on different samples. Table 8a models converting to any learning use among adopters, while Table 10a models the rarer decoupling outcome under a separate specification, so a strong conversion effect and a weaker decoupling coefficient are consistent.

The typology adds a population reading. Table 10b reports the latent-class fit. The information criterion continues to fall from three through six classes, so the six-class solution is chosen for interpretability rather than by a statistical minimum, with entropy near 0.77 indicating acceptable separation. Profiling the six classes shows that decoupling is concentrated in a single low-breadth class, which draws from every income tier in broadly similar proportions, so the disengaged are a minority present across the income spectrum rather than a low-income group.

Table 10b. Latent-class analysis, fit across solutions.

One demographic pattern is distinct. Men are somewhat more likely to adopt AI in the first place, at an odds ratio of 1.26 (p = .04), and the gender difference in adoption varies by tier, being sharper in the upper-middle tier. This adoption gap does not carry into conversion; decoupling is the same for women and men, so it belongs to who reaches these tools.

4.6 The construct and its scope

These results place the study within the digital-inequality tradition. The variance decomposition sets a ceiling on how much context can matter. Table 10c reports all three levels. Region and country together account for about 4.6% of the variation in proficiency, and that contextual share sits mainly at the country level, at 4.1%, with region adding little. The rest, about 95%, is at the level of the individual learner. The contextual share is small but not trivial, and it is enough to bias inference if ignored, which is why the models cluster and nest. Most of what separates learners is therefore individual, the statistical counterpart of the typology results above.

Table 10c. Three-level variance decomposition of self-rated proficiency.

The construct's validity is partly established and partly assumed, which we state plainly. Breadth correlates positively with self-rated proficiency among adopters, though weakly, at a rank correlation of 0.05, so it carries a small convergent signal. The measure also treats the 20 features as equal and additive, and it is not yet validated against an external learning outcome. Breadth is therefore a defensible but provisional operationalisation, which the limitations revisit.

The 20 features divide into seven general-support features and 13 programming features. Table 10d reports the split. The low-income conversion effect is robust for any learning use, at an odds ratio of 0.49, but it reaches significance for neither subscale on its own. This is the signature of a threshold effect, since the gap concerns crossing into any learning use rather than the kind of use, and the small low-income cell limits power once the outcome is subdivided. The weak tie between breadth and proficiency is carried by the programming features, at a rank correlation of 0.08, while the support features are unrelated to proficiency. Some uses therefore track proficiency more than others, which the equal and additive treatment of features does not capture.

Table 10d. Feature-subscale split.

The analysis yields several propositions that together define the access-conversion gap and its correlates. A usage divide persists among the connected, since the conversion of adoption into learning use diverges across context. That divide operates at the threshold of first learning use. What gates the threshold is less income than a set of enabling conditions. No single condition is necessary. Several different combinations produce a wide gap, and inexperience is the condition that appears most often. And where formal support is thinner, AI use is more strongly associated with proficiency, which points to complementarity under scarcity.

5 Implications

The evidence points to conversion, not access, so the useful levers are the conditions that let access become learning. Distributing more tools would not be enough since the learners in view already reach them. That shift spreads the burden across several actors since conversion depends on the devices learners hold, on how tools are built, and on how learning is taught. For educators and institutions, the practical lesson concerns the threshold. The gap opens at first learning use, so the task is to help a learner move from asking an assistant a stray question to using it inside their own practice. In computing this means teaching the write, run, and debug loop with an assistant present, and modelling the explanation, error-fixing, and debugging uses that the data associate with conversion. Assessment either helps or hinders here. Tasks that reward finished answers push learners toward shallow use, while tasks that reward reasoning and iteration reward the deeper use that builds skill (Ahangama, 2026). Framing an assistant as a tutor to be questioned is a small curricular move with a large bearing on conversion.

For tool providers and platform designers the device pattern is suggestive. Smartphone-only study does not reach significance in the model since such cases are rare, but it is concentrated in the lower tiers, so an interface that assumes a keyboard and a code editor may quietly exclude the learners where the gap is widest (Tsetsi and Rains, 2017). Mobile-first designs, low-bandwidth and offline modes, and support for local languages are therefore plausible levers (Kolil and Achuthan 2023), offered as hypotheses to test. Features matter as much as form. Tools that scaffold understanding (Ottappurakkal et al., 2025; Radhamani et al., 2016), that explain, question, and walk a learner through an error, are the ones this study associates with conversion, while tools tuned only to produce finished output invite the shallow use that does not build competence.

For governments and funders, the main question is what gets measured. National digital-skills statistics that count adoption will report success where a usage divide persists, so monitoring should track breadth of use and the conversion gap alongside access. Targeting follows from the typology. The disengaged exist in every tier. Programmes aimed at whole national tiers will reach many who need no help and miss many who do, so targeting the minamal-use profile directly is important. The equity case is strong. If AI use matters most for proficiency where other support is thin, then failing to help learners cross the threshold is costliest for those with the most to gain, which makes closing the gap an equity aim and not only a matter of efficiency.

6 Limitations

The study has several limitations. The survey is opt-in, recruited through social-media advertising, peer referrals, and the vendor's own channels, so it samples computing learners reachable that way. The published weights calibrate the country distribution to an external benchmark, which is what supports the cross-country and cross-tier comparisons made here, though they cannot remove selection into the survey. The findings therefore describe engaged computing learners, and the adoption pattern is a feature of that population. Proficiency is self-rated, so the third-question results are associational and open to reverse causation. The income and region coding are analyst decisions, the lowest-income cell is small, and the configurational analysis is preliminary because the rare outcome bounds its consistency. Further, the eight-region scheme groups diverse countries together as per the World Bank groupings, so within-region differences are absorbed into the regional averages and not separately estimated.

Three measurement and design limits deserve emphasis. The two-part model conditions on adoption and does not correct selection into adoption, so a selection process correlated with conversion cannot be ruled out. Self-rated proficiency is also open to cross-cultural response styles since respondents in different countries may use the same scale differently (Davidov et al., 2014). Anchoring vignettes are the standard correction for this, but secondary data cannot supply them. Formal measurement-invariance checks are likewise unavailable here (Vandenberg and Lance 2000), so the third-question results should be read with that caution. Breadth treats the 20 features as equal and additive and is validated only weakly, against self-rated proficiency.

Two further limits concern how far the findings extend. The survey is cross-sectional, so every association is contemporaneous and no direction of effect can be established. The construct was developed abductively, from the pattern in the data to the theory that explains it. Replication on other samples, a calibrated fuzzy-set analysis, and an external measure of proficiency are recommended as future work.

The composition of the sample is skewed towards male and younger learners. The assistant landscape is fixed to the survey period, so tool-level readings are tied to a moment in a fast-moving field and may not hold as models and interfaces change. These are reasons to treat the findings as a well-supported account of a particular population at a particular time, and to seek replication as both the tools and the users continue to shift.

7 Conclusion

Across 14,653 computer science learners in 166 countries, this study locates the divide in generative AI not in access but in use. It names that divide the access-conversion gap, the step from reaching a tool to using it to learn. Adoption is high and roughly even across income tiers in this connected population. Yet adoption and learning use come apart, and most where resources are thinnest. Reading the gap instead of adoption changes what the data say. A low-income tier that looks well served on adoption, at about 76%, proves least able to turn it into learning. It decouples at 10.3%, against 5.6 in the high-income tier.

The two-part model shows where the gap sits. It lies at the threshold of first learning use, where adopters in low-income countries are about half as likely to cross, at an odds ratio of 0.49. It does not open in the breadth of use beyond it, which is even across tiers. What gates the threshold is a set of facilitating conditions, not income alone. Inexperience is the most common condition among wide-gap cases, and smartphone-only study is concentrated in the poorer tiers. Region and country together explain only about 4.6% of the variation in proficiency. The gap is therefore a within-context phenomenon that national averages hide. Against a substitution worry, AI use is most strongly tied to proficiency where formal support is scarcest. That points to complementarity under scarcity and raises the cost of failing to help these learners cross the threshold.

These results move the question from who can reach these tools to who can turn reaching them into learning. For policy, the target shifts from distributing devices to building the conditions of conversion. It shifts too from national averages to the disengaged minority present within every tier. Hence, next gains in equity will depend on how the usage divide is addressed.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. The author acknowledges the use of generative AI for language editing, grammar checking, and improving the flow of the text. The author reviewed and revised all AI-generated suggestions for the final content of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feduc.2026.1946943/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Ana María Pinto-Llorente, University of Salamanca, Spain

Reviewed by: Pedro Miralles Sánchez, University of Murcia, Spain

Szonja Jenei, Széchenyi István University, Hungary