Using artificial intelligence is not the same as using it to learn. A peer-reviewed analysis of 14,653 computer-science learners across 166 countries finds that some people who regularly use AI never bring it into their studies. Among everyday users in low-income countries, the adjusted odds of using AI for any learning purpose were about half those of comparable users in high-income countries. The research was published on 8 October 2026 in Frontiers in Education.
That distinction matters as schools and governments increasingly measure digital progress by access to devices, subscriptions or tools. A learner may be comfortable asking an assistant questions but have little idea how to use it to debug code, investigate a concept or evaluate an explanation. The study proposes that the step from everyday adoption to learning use should be measured separately from connectivity and from educational achievement.
What the study actually measured
Researcher Krishnashree Achuthan calls this the access-conversion gap. In the study, effective access is reporting everyday AI use. Conversion means reporting at least one of 20 specific learning-related activities. Seven concern general study support, including summarising text, translation and study planning. Thirteen concern programming, including explaining code, generating tests, debugging and interpreting errors. The count of distinct purposes is called breadth. Importantly, neither conversion nor breadth is a direct measure of learning gains.
The author reanalysed an archived international survey of 18,032 respondents collected during the first half of 2024. The analytic sample comprised 14,653 people who answered the everyday AI-use item, from 166 countries. One person lacked an income-tier classification, leaving 14,652 in that comparison. Every respondent had studied computer science in the preceding year, either formally or through self-education. This is a study of an already engaged computing population, not a representative sample of all students or citizens.
Participants were recruited through social-media advertising, peer referrals and the survey organiser’s channels. In the weighted sample, 83.5% were male, 67.8% were younger than 30, 50.1% were formal students and 30% reported no professional coding experience. Those characteristics matter because digitally engaged volunteers may use AI differently from people who lack affordable connectivity or computing equipment.
The researcher compared income tiers and world regions, adjusting for age, education, coding experience, student status and other characteristics. The analysis used survey weights, country-clustered uncertainty estimates and models accounting for country and region. A two-part statistical model distinguished whether adopters used AI for any learning purpose from the number of purposes reported by those who did. Twenty multiple imputations addressed missing background data. Sensitivity analyses tested whether results depended on the definition of learning use.
Everyday AI adoption was high, but conversion differed
On the weighted measure, 67.3% of the analytic sample reported everyday AI use. Among these selected computing learners, adoption was 59.2% in high-income countries and 76.4% in low-income countries. This does not show that the general population of poorer countries has better AI access. Rather, it illustrates how a survey restricted to connected, motivated learners can look different from population-level digital-divide statistics.
The key contrast concerned people who had already adopted AI. In high-income countries, 5.6% of everyday users reported no learning-related use. In low-income countries, the corresponding share was 10.3%. Thus, non-conversion among adopters was nearly twice as common in the low-income tier. These percentages describe the proportion of adopters who did not report learning use, not the proportion of all people without access.
In the adjusted model, everyday users in low-income countries had an odds ratio of 0.49 for any learning use compared with those in high-income countries (95% confidence interval 0.27 to 0.90; p = 0.022). This means roughly 51% lower odds, not a 51-percentage-point reduction in probability. The lower-middle-income estimate was 0.95 (95% CI 0.75 to 1.19), with no clear difference from the high-income group. The upper-middle-income estimate was 1.63 (95% CI 1.28 to 2.08), indicating higher adjusted conversion odds.
The low-income subgroup was small: only 106 respondents, of whom 82 were everyday users. Its estimate is therefore less precise than the large overall sample might suggest. Robustness checks gave similar central results when survey weights were removed or individual low-income countries were excluded, with leave-one-country-out odds ratios ranging from 0.39 to 0.51. Nevertheless, those checks cannot make the subgroup representative of low-income countries as a whole.
The largest difference was at the threshold
The second part of the model asked whether adopters who already used AI for learning reported fewer different learning purposes. For the low-income tier, the estimated rate ratio for breadth was 0.91 (95% CI 0.79 to 1.04; p = 0.16), which did not establish a difference from high-income users. The lower-middle-income estimate was similarly close to parity at 1.01. Upper-middle-income converters did report a small but statistically significant increase in breadth, with a rate ratio of 1.05 (95% CI 1.02 to 1.08).
This distinction suggests that the low-income disadvantage in the sample appeared mainly at the point of beginning learning use, rather than consistently widening the range of features used after that point. When the threshold was raised to at least two learning features, 19.4% of low-income adopters fell below it, compared with 13.9% of high-income adopters. At three features, the respective proportions were 34.1% and 25.8%. More demanding definitions did not remove the ordering, but feature counts remain behavioural indicators, not examination scores.
What the regional numbers mean for Africa
Sub-Saharan Africa accounted for 1,392 respondents. Weighted everyday AI adoption was 70.6%, while any learning use was 66.2%, leaving a 4.4-percentage-point gap. Among everyday adopters, 6.3% reported no learning use. That was higher than the corresponding rates in Western Europe (5.8%) and East Asia and the Pacific (3.9%). These figures are regional aggregates and cannot be treated as estimates specifically for South Africa.
Study devices may contribute to unequal opportunities, although this analysis cannot establish a cause. Smartphone-only study was reported by 1.6% of the overall sample, but by 5.4% in lower-middle-income countries and 4.8% in low-income countries. Its estimated association with non-conversion was positive but statistically uncertain: odds ratio 1.80, 95% CI 0.51 to 6.36, p = 0.36. A phone can handle chat easily while making it difficult to write, run and inspect a computer program. The data make this a plausible design question, not a proven explanation.
Feeling proficient is not proof of learning
The researcher also examined whether AI learning use was associated with self-rated computer-science proficiency. In the high-income tier, the adjusted relationship was close to zero. Relative to that reference relationship, the lower-middle-income interaction was +0.46 points (95% CI 0.28 to 0.64; p below 0.001). The low-income interaction was larger, at +1.11 points, but its confidence interval included zero (-0.35 to 2.56; p = 0.14). Regionally, the association was approximately +0.56 points in Sub-Saharan Africa, +0.40 in South Asia and +0.39 in East Asia and the Pacific.
The author suggests that AI may complement scarce tutoring or educational support. That is a reasonable interpretation to test, but the survey cannot demonstrate it. Learners who already feel more capable might be more likely to adopt AI, and confidence may increase without an equivalent increase in objective skill. Self-rated proficiency also differs across cultural contexts. The study did not assess students through a common programming examination.
Feature breadth itself correlated only weakly with self-rated proficiency, at a rank correlation of 0.05 overall. The correlation was 0.08 for programming features and approximately zero for general study-support features. This is a useful warning against assuming that checking more AI features automatically means greater educational benefit. Adoption, learning behaviour and demonstrated achievement are different outcomes.
Practical implications for teaching and digital policy
For teachers, the research suggests that a useful question is not simply whether students have an AI account. It is whether they can use the tool while actively reasoning through a task. A structured debugging exercise could ask a student to identify an error, request an explanation, test a proposed fix and describe why the code now works. Such an activity would reveal more about understanding than a one-question survey about AI adoption.
For institutions and policymakers, separate measures of access, adoption, learning use and educational outcomes would help expose hidden barriers. Learners who rely on phones, have little coding experience or lack mentoring may need different forms of support. In the exploratory analysis, coding inexperience appeared in about half of the non-conversion cases, while disengaged learners were found across income tiers. No single condition was shown to be a necessary cause of the gap.
The findings add context to previous Research Today coverage showing that how students use generative AI may matter more for critical thinking than frequency alone. They also complement a semester-long study that found no additional measurable knowledge gain from AI. Different research designs ask different questions, and these results should not be collapsed into a simple claim that AI improves learning.
Limits and what comes next
The survey was collected in early 2024, although the analysis was published in October 2026. Tools, interfaces and study practices may have changed substantially since then. The opt-in sample was disproportionately male and young, and weights cannot remove every recruitment bias. The lowest-income cell was small, the data were cross-sectional and the results do not establish causality. The study’s 20-feature breadth measure assumes all features count equally and has only weak validation against self-ratings, not independently assessed skills.
More convincing evidence would follow learners over time, observe what they actually do with AI, and assess their knowledge through comparable tasks before and after instruction. Interventions could then test whether guided practice, accessible devices or better explanations help people turn routine AI use into productive study. For now, the clearest conclusion is that high adoption can coexist with a meaningful usage gap, even among people who are already digitally connected.
Source Information
Study: The access-conversion gap across countries: how generative AI adoption exceeds learning use.
Author: Krishnashree Achuthan.
Journal: Frontiers in Education, Volume 11.
Publication: 8 October 2026.
DOI: 10.3389/feduc.2026.1946943.
Research type: Peer-reviewed secondary analysis of an international cross-sectional survey.
