1. Introduction

The mass deployment of generative artificial intelligence in late 2022 transformed, in less than four years, the relationship between schools and computational knowledge. What had long been a promise confined to educational computing laboratories became an everyday infrastructure: students drafting with language models, teachers planning with conversational assistants, and school systems piloting adaptive tutors. However, the scientific literature agrees that this adoption is not evenly distributed: AI can function simultaneously as both cause and remedy of digital inequality (Božić, 2023), and its potential to narrow gaps depends on social, economic, and institutional conditions that are not met by default (Hussein et al., 2025; Peters & Tukdeo, 2025).

The question guiding this review is not whether AI will reach classrooms —it is already there— but who learns with it, under what conditions, and with what outcomes. UNESCO warned as early as 2021 that integrating AI into education required normative guidance to avoid deepening existing inequalities (Miao et al., 2021), and by 2023 it observed that the speed of generative AI had outpaced states' regulatory capacity, leaving "data privacy unprotected and institutions unprepared" (Miao & Holmes, 2023). In this context, the concept of the educational gap requires theoretical updating: measuring access to devices and connectivity is no longer enough; what matters is the differential capacity of students and teachers to understand, use, and create with AI systems.

This article reviews the state of the art on educational gaps in AI learning in classrooms. Three research questions are posed: (1) What types of gaps does the recent literature document, and through what mechanisms are they reproduced? (2) What empirical evidence exists on their magnitude across social and geographic contexts? (3) What pedagogical, curricular, and regulatory responses does the available research propose? The expected contribution is twofold: to offer an analytical typology useful for doctoral research in education and technology, and to synthesize actionable implications for school systems that, like those in Latin America, face the incorporation of AI with heterogeneous resources.

2. Conceptual framework: From the digital divide to the AI divide

Research on digital inequality classically distinguishes three levels: the first-level divide (access to infrastructure), the second-level divide (skills and usage competencies), and the third-level divide (differential conversion of use into tangible outcomes). Recent literature shows that this conceptual architecture remains operative in the face of generative AI, but with substantive shifts. Zhang et al. (2026), in a study of university students from a digital divide perspective, demonstrate that socioeconomic status and digital and AI literacy predict ChatGPT activities, educational outcomes, and trust: conventional digital inequality "persists and evolves" with generative AI. Along the same lines, Hargittai et al. (2026) document significant inequalities in knowledge —not merely access— about ChatGPT, associated with educational attainment and prior exposure to AI information.

In response to these findings, the concept of AI literacy has consolidated as the central construct of the second-level divide. Faruqe et al. (2021) proposed a competency model articulating knowledge, skills, and attitudes toward AI; Carolus et al. (2023) extended the model toward competent interaction with conversational AI systems; and Choi et al. (2025) synthesized definitions, competencies, and challenges of AI literacy in education, highlighting that teacher preparedness is the most frequently cited bottleneck. For specific populations, Atias and Mawasi (2025) systematically reviewed AI literacy programs targeting children and youth, revealing considerable paradigmatic dispersion: there is no consensus on what it means to "know AI" at each educational stage, which hinders both teaching and measurement of the gap (Mauermeister et al., 2025).

In synthesis, the framework guiding this review conceptualizes the AI divide as a multilevel phenomenon that inherits the structure of the digital divide but incorporates specific mechanisms: the cost of advanced models, the hegemony of English in training data and interfaces, and the demand for new interaction competencies (prompt formulation, critical verification, bias awareness) that school curricula do not yet systematically integrate.

3. Methodology

A narrative state-of-the-art review was conducted with systematic selection criteria. The corpus comprises twenty-six works published between 2021 and 2026: articles in peer-reviewed journals (e.g., Computers and Education: Artificial Intelligence, New Media & Society, Teaching Education, TechTrends, Journal of Academic Ethics, Education Sciences, Contemporary Education Dialogue), chapters from academic publishers (Taylor & Francis, IGI Global), preprints under review (SocArXiv, Preprints.org, arXiv), international conference proceedings (EDULEARN), and UNESCO institutional reports widely cited by the literature. Inclusion criteria were: (a) explicit focus on AI, education, and inequality or gaps; (b) publication in the 2021-2026 period; (c) verifiable metadata (authorship, year, publication venue, DOI or URL). The analysis followed a thematic synthesis procedure: each source was coded according to the type of gap documented, the proposed causal mechanism, the population studied, and the recommended responses.

4. State of the art: A typology of five gaps

4.1. First gap: Access and infrastructure

The access gap persists as the enabling condition of all the others. Tripathi et al. (2025), in a systematic review on rural education, show that limited connectivity and device scarcity maintain the urban-rural fracture even when AI tools capable of personalizing learning exist: without minimal infrastructure, the promise of AI fails to reach the populations that could benefit most. Latin American evidence points in the same direction: Valdivieso and González (2025) document that in Salvadoran higher education, socioeconomic factors determine who can integrate generative AI tools into their academic work, so that the technology reproduces prior hierarchies rather than leveling them.

A new phenomenon within this gap is subscription stratification: the most capable models operate under paid schemes, turning full access to AI into a positional good. Caporarello and Trabskaia (2026) describe this "dark side" of AI-driven education: in Global South universities, access to paid AI programs becomes a differentiator among institutions and students, deepening the North-South divide. The first gap is therefore no longer measured solely in megabits per second, but in the hierarchy of models each student can access.

4.2. Second gap: AI competencies and literacy

Even when access is guaranteed, the capacity to use AI in an informed manner is unequally distributed. Hargittai et al. (2026) show that effective knowledge about ChatGPT varies systematically with users' educational level and social capital; Zhang et al. (2026) confirm that AI literacy mediates between socioeconomic status and the educational outcomes obtained with generative tools. In terms of the conceptual framework, the second gap turns AI literacy into a new form of cultural capital: those who possess it extract educational benefits; those who do not face risks of uncritical dependence or exclusion.

The literature agrees that this gap begins before university. Atias and Mawasi (2025) find that AI literacy programs for children and youth are heterogeneous, scarce, and rarely articulated with official curricula; Choi et al. (2025) add that the lack of shared operational definitions hinders program scaling. In response, UNESCO published in 2024 the AI Competency Framework for Students, with twelve competencies across four dimensions —human-centred mindset, ethics of AI, AI techniques and applications, and AI system design— organized into three progression levels: understand, apply, and create (Miao & Shiohira, 2024). The existence of the framework, however, does not guarantee its adoption: the distance between the global standard and national curricula is itself a manifestation of the gap.

4.3. Third gap: Differential use and appropriation

The third gap concerns the conversion of access and competencies into real educational outcomes. Evidence shows that students with stronger prior trajectories use AI more strategically —to verify, contrast, and elaborate— while students with less preparation tend toward substitutive uses that can erode learning (Nguyen, 2025; Zhang et al., 2026). Katona and Gyönyörű (2025) provide counter-evidence from an intervention: AI-based adaptive programming education directed at socially disadvantaged students succeeded in narrowing the digital divide when the pedagogical design was deliberate and scaffolded. The lesson is that appropriation is not a property of the tool but of the instructional design that frames it.

This gap also has an ethical face. Kayyali (2025) warns that privacy, algorithmic bias, and academic integrity differentially affect students according to age, geography, and socioeconomic background, and Nguyen (2025) shows that the absence of clear pedagogical principles amplifies disparities in learning opportunities among groups. Differential appropriation, in sum, distributes not only benefits unequally: it also distributes risks.

4.4. Fourth gap: Teacher training

None of the preceding gaps can be closed without prepared teachers, and the literature is unanimous in diagnosing a structural deficit at this point. Guan et al. (2025) studied pre-service teachers and found that their preparedness for AI-integrated education depends on still-insufficient attitudes, self-efficacy, and AI literacy, with direct implications for their professional identity. Kundu et al. (2025), in Indian schools, show that teacher AI literacy is simultaneously cognitive, pedagogical, ethical, and contextual, and that professional development programs rarely address all four dimensions. Caspari-Sadeghi (2026) documents the preceding link: teacher educators themselves lack holistic AI literacy curricula, so the deficit is transmitted generationally within the training system. It is unsurprising, then, that the specialized literature describes the integration of AI into teacher training as an urgency (Bekdemir, 2024).

In response, UNESCO published the AI Competency Framework for Teachers, which defines fifteen competencies across five domains —human-centred mindset, ethics of AI, AI foundations and applications, AI pedagogy, and AI for professional development— with three proficiency levels: acquire, deepen, and create (Miao & Cukurova, 2024). The framework offers a global reference standard; its effective implementation, however, demands sustained investment in pre-service and in-service training that most school systems have yet to program.

4.5. Fifth gap: North-South curricular governance

The fifth gap operates at the systemic scale: states' capacity to regulate, curricularize, and finance educational AI. Miao and Holmes (2023) observed that generative AI was advancing faster than regulatory frameworks and proposed a human-centred governance agenda including age limits, data protection, and requirements for providers. Peters and Tukdeo (2025) warn that without a dedicated research agenda for the Global South (AI4ED), the region will oscillate between imported utopian and dystopian frames, without endogenous evidence production. Okolo (2023) underscores that inclusive AI governance in the Global South requires institutional capacities that go beyond connectivity: regulation, data, talent, and participation.

Cross-cutting dimensions operate within this systemic gap. Language is among the most decisive: Finardi (2025) shows that AI in higher education tends to reproduce the hegemony of English, with implications for linguistic diversity and the inclusion of Global South knowledges. Rurality, already noted, and the gender inequalities documented in technology access and confidence complete a landscape in which the AI gap is, above all, a reconfiguration of pre-existing inequalities under new technical conditions (Hussein et al., 2025; Valdivieso & González, 2025).

5. Synthesis of the state of the art

Table 1 summarizes the resulting typology, the dominant causal mechanism documented by the literature, and the responses proposed for each gap.

Table 1. Typology of educational gaps in AI learning, mechanisms, and documented responses (2021-2026)
GapDominant mechanismRepresentative evidenceProposed responses
1. Access and infrastructureConnectivity, devices, and paid subscriptions hierarchize access to modelsTripathi et al. (2025); Caporarello & Trabskaia (2026); Valdivieso & González (2025)Rural infrastructure investment; institutional access to models; educational licensing
2. Competencies and literacyAI literacy operates as new cultural capital linked to socioeconomic statusHargittai et al. (2026); Zhang et al. (2026); Atias & Mawasi (2025)AI literacy curricula; UNESCO student framework (Miao & Shiohira, 2024)
3. Use and appropriationStrategic vs. substitutive uses according to prior trajectory; instructional design as moderatorKatona & Gyönyörű (2025); Nguyen (2025); Zhang et al. (2026)Pedagogical scaffolding; explicit ethical and pedagogical principles (Kayyali, 2025)
4. Teacher trainingCognitive, pedagogical, ethical, and contextual deficit in teachers and teacher educatorsGuan et al. (2025); Kundu et al. (2025); Caspari-Sadeghi (2026); Bekdemir (2024)Pre-service and in-service training based on the UNESCO teacher framework (Miao & Cukurova, 2024)
5. Curricular governanceRegulation slower than technology; dependence on Global North frameworksMiao & Holmes (2023); Peters & Tukdeo (2025); Okolo (2023); Finardi (2025)Endogenous research agendas (AI4ED); rights-centred regulation; linguistic diversity

6. Discussion

The state of the art supports a synthetic thesis: the AI gap in classrooms exhibits a triple specificity relative to the classic digital divide. First, it is inherited: the social determinants of digital inequality —income, schooling, geography, gender— also predict access to, competencies in, and appropriation of AI (Hargittai et al., 2026; Zhang et al., 2026). Second, it is amplified by properties specific to generative AI: subscription stratification (Caporarello & Trabskaia, 2026), the linguistic hegemony of English (Finardi, 2025), and the demand for critical interaction competencies that school does not yet teach (Atias & Mawasi, 2025; Choi et al., 2025). Third, it is institutionalized: when education systems fail to train teachers or update curricula, inequality ceases to be a contingency of the technology market and becomes a product of the system itself (Caspari-Sadeghi, 2026; Miao & Holmes, 2023).

This reading nuances the metaphor, frequent in popular discourse, of AI as a "great equalizer". The reviewed evidence suggests instead a conditioned equalization: deliberate interventions, with explicit pedagogical design and a focus on disadvantaged populations, do succeed in narrowing specific gaps (Katona & Gyönyörű, 2025); but the spontaneous diffusion of the technology tends to reproduce and deepen prior hierarchies (Božić, 2023; Hussein et al., 2025). The decisive variable is not the technology but the institutionality that frames it: access policies, teacher training, explicit curricula, and protective regulation (Miao et al., 2021; Miao & Holmes, 2023).

For doctoral research, this typology opens at least three fronts: the need for specific measurement instruments for the AI gap in basic education (Mauermeister et al., 2025); the longitudinal study of AI literacy effects on educational trajectories; and applied research on teacher training models in low-resource contexts, particularly in the Global South and in languages other than English (Peters & Tukdeo, 2025; Finardi, 2025).

7. Implications for education policy and practice

Four implications follow from the review. First, access must be conceptualized broadly: guaranteeing connectivity and devices is necessary but insufficient; systems must ensure institutional access to capable AI models, preventing the quality of the available model from depending on family income (Caporarello & Trabskaia, 2026). Second, AI literacy must be incorporated into official curricula with explicit progressions, for which UNESCO's student framework offers a reference architecture adaptable to national contexts (Miao & Shiohira, 2024). Third, teacher training is the lever with the greatest multiplier: existing competency frameworks (Miao & Cukurova, 2024) will only produce effects if translated into pre-service and in-service programs with time, mentoring, and assessment (Bekdemir, 2024; Kundu et al., 2025). Fourth, governance must protect diversity —linguistic, cultural, and of local knowledges— so that educational AI does not operate as a vector of homogenization (Finardi, 2025; Okolo, 2023).

8. Limitations

This review has limitations inherent to its design. It is a narrative review with systematic criteria, not an exhaustive systematic review with a PRISMA protocol; the corpus, while representative of the 2021-2026 debate, does not exhaust available production. The empirical literature concentrates on higher education and a reduced number of countries, so inferences about basic education and Latin America require caution. Part of the most recent evidence comes from preprints under review, whose results should be considered preliminary. Finally, the field evolves rapidly: the conclusions should be read as a snapshot of the state of the art as of August 2026, not as a closure of the debate.

9. Conclusions

Educational gaps in artificial intelligence learning are not an accidental side effect of a new technology, but the updated expression, under new technical conditions, of structural educational inequalities. The 2021-2026 literature converges on a typology of five interdependent gaps —access, competencies, appropriation, teacher training, and curricular governance— crossed by dimensions of gender, language, and rurality. Against them, the state of the art offers both diagnoses and instruments: competency frameworks for students and teachers, evidence on effective interventions, and a rights-centred governance agenda. The open practical question is whether education systems —particularly those of the Global South— will manage to turn these instruments into sustained policies before the AI gap settles as a new layer of educational inequality. The answer does not depend on technology, but on institutional decisions that have yet to be made.

References

  1. Atias, O., & Mawasi, A. (2025). Conceptualizing AI literacies for children and youth: A systematic review on the design of AI literacy educational programs. Computers and Education: Artificial Intelligence. https://www.sciencedirect.com/science/article/pii/S2666920X25001316
  2. Bekdemir, Y. (2024). The urgency of AI integration in teacher training: Shaping the future of education. Journal of Research in Didactical Sciences. https://www.researchgate.net/publication/386107830
  3. Božić, V. (2023). Artificial intelligence as the reason and the solution of digital divide. Language Education and Technology Journal. http://www.langedutech.com/letjournal/index.php/let/article/view/53
  4. Caporarello, L., & Trabskaia, I. (2026). Global North-South divide: The dark side of AI-driven education. In EDULEARN26 Proceedings. https://library.iated.org/view/CAPORARELLO2026GLO
  5. Carolus, A., Augustin, Y., Markus, A., & Wienrich, C. (2023). Digital interaction literacy model–Conceptualizing competencies for literate interactions with voice-based AI systems. Computers and Education: Artificial Intelligence. https://www.sciencedirect.com/science/article/pii/S2666920X22000698
  6. Caspari-Sadeghi, S. (2026). AI literacy for teacher educators: A holistic curriculum for capacity-building in higher education. Frontiers in Education. https://doi.org/10.3389/feduc.2026.1745768
  7. Choi, W. C., Chang, C. I., Choi, I. C., Lam, L. C., Leong, K. I., & Ng, S. I. (2025). Artificial intelligence (AI) literacy in education: Definition, competencies, opportunities and challenges [Preprint]. Preprints.org. https://www.preprints.org/frontend/manuscript/11759de052eff1261d384254db19cefc/download_pub
  8. Faruqe, F., Watkins, R., & Medsker, L. (2021). Competency model approach to AI literacy: Research-based path from initial framework to model. arXiv. https://arxiv.org/abs/2108.05809
  9. Finardi, K. R. (2025). AI in higher education: Implications for diversity of languages and inclusion of knowledges from the global South. In Applied research in business and education. Taylor & Francis. https://doi.org/10.4324/9781003739920-12
  10. Guan, L., Zhang, Y., & Gu, M. M. (2025). Pre-service teachers preparedness for AI-integrated education: An investigation from perceptions, capabilities, and teachers' identity changes. Computers and Education: Artificial Intelligence. https://www.sciencedirect.com/science/article/pii/S2666920X24001449
  11. Hargittai, E., Lokaj, G., & Hess, L. (2026). Digital inequality in ChatGPT awareness and knowledge. In Connective AI. Taylor & Francis. https://doi.org/10.4324/9781003477587-7
  12. Hussein, S., Qureshi, S. S., & ul Emman, S. K. (2025). Analyzing how AI can both exacerbate and help overcome digital inequalities in education, especially in underserved regions. The Critical Review of Social Sciences Studies. http://thecrsss.com/index.php/Journal/article/view/437
  13. Katona, J., & Gyönyörű, K. I. K. (2025). AI-based adaptive programming education for socially disadvantaged students: Bridging the digital divide. TechTrends. https://doi.org/10.1007/s11528-025-01088-8
  14. Kayyali, M. (2025). Ethical implications of generative AI in education: Privacy, bias, and integrity. In Transformative AI practices for personalized learning. IGI Global. https://www.igi-global.com/chapter/ethical-implications-of-generative-ai-in-education/377145
  15. Kundu, A., Bej, T., & Mondal, R. (2025). Exploring teachers' AI literacy: Cognitive, pedagogical, ethical, and contextual insights from Indian schools. Teaching Education. https://doi.org/10.1080/10476210.2025.2570227
  16. Mauermeister, S., Biere, L., & Steinhardt, I. (2025). Measuring the digital divide among students caused by artificial intelligence [Preprint]. SocArXiv. https://osf.io/preprints/socarxiv/e824j
  17. Miao, F., & Cukurova, M. (2024). AI competency framework for teachers. UNESCO. https://doi.org/10.54675/ZJTE2084
  18. Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. https://doi.org/10.54675/EWZM9535
  19. Miao, F., & Shiohira, K. (2024). AI competency framework for students. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000391105
  20. Miao, F., Holmes, W., Huang, R., & Zhang, H. (2021). AI and education: Guidance for policy-makers. UNESCO.
  21. Nguyen, K. V. (2025). The use of generative AI tools in higher education: Ethical and pedagogical principles. Journal of Academic Ethics. https://doi.org/10.1007/s10805-025-09607-1
  22. Okolo, C. T. (2023). AI in the Global South: Opportunities and challenges towards more inclusive governance. Brookings Institution. https://internationalpolicybrief.org/wp-content/uploads/2024/02/ARTICLE-9.pdf
  23. Peters, M. A., & Tukdeo, S. (2025). Beyond the utopic/dyspotic frames: Towards a research agenda for AI in education and development (AI4ED) in the Global South. Contemporary Education Dialogue. https://doi.org/10.1177/09731849241298118
  24. Tripathi, A., Yadav, V., & Kumar, S. (2025). Leveraging artificial intelligence for rural education: A systematic review of transforming learning opportunities and bridging the urban-rural divide [Preprint]. Preprints.org. https://www.preprints.org/frontend/manuscript/a6e02ac0561289ccf3b0dd2225026145/download_pub
  25. Valdivieso, T., & González, O. (2025). Generative AI tools in Salvadoran higher education: Balancing equity, ethics, and knowledge management in the global south. Education Sciences, 15(2), 214. https://www.mdpi.com/2227-7102/15/2/214
  26. Zhang, C., Rice, R. E., & Wang, L. H. (2026). College students' literacy, ChatGPT activities, educational outcomes, and trust from a digital divide perspective. New Media & Society. https://doi.org/10.1177/14614448241301741