1. Introduction and problem

At innovation fairs and in policy documents, artificial intelligence “enters” the classroom as if the classroom were already electrified, cabled and updated. The chatbot “personalizes.” The dashboard “diagnoses.” The assistant “accompanies the rural teacher.” That grammar treats the network, the device, energy and maintenance as a technical background, not as the object of inequity itself. The kindergarten in a low-density locality, the multigrade school, the telebachillerato or Telesecundaria are not, however, a delayed version of the connected urban classroom. They are another material condition. Where there is no stable electricity, there is no cloud. Where bandwidth cannot sustain a video call, there is no streaming language model. Where the only computer works poorly and there is no technician within hours of travel, there is no “AI innovation”: there is a story that presupposes what the territory does not have.

The problem is not hostility to the tool or nostalgia for the blackboard. It is the confusion of three objects. The first is an empirical finding: what is observed when urban and rural household connectivity, electricity and internet for pedagogical purposes in schools, principals’ perception of insufficient network, and experiences of technologies that operate with little or no connectivity are measured. The second is a still-current normative or policy framework: what UNESCO (2023) prescribes on technology in education “on our terms,” the AI guidance for policy-makers (Miao, Holmes, Huang and Zhang, 2021), the guidance on generative AI (Miao and Holmes, 2023) and the OECD Starting Strong VII report (2023a) on early childhood in the digital age. The third is a pedagogical inference: what should not be sold as a solution —AI that “closes the digital divide,” that “brings quality to the rural,” that “personalizes where teachers are missing”— if the source does not measure it. Mixing them produces a pedagogy of the cloud: AI is announced for all children and is implemented, in fact, for those who already have fibre, a tablet and support.

The thesis of this article is restrictive. Without connectivity, energy and technical support, educational AI does not arrive late: it arrives as inequality. Where the source does not measure gap closure, it is not claimed. Where it does not measure improvement of rural learning by a language model, it is not promised. Where it does not measure substitution of teacher scarcity, it is not invented. The gap is infrastructural: what material conditions of the kindergarten and the basic school make possible —or impossible— any “AI innovation.” Left outside as an axis are gaps in “learning AI” as a subject, multilingualism, privacy, inclusion by disability, parental mediation, UNESCO teacher competence, well-being, integrity, pedagogical documentation, student AI literacy, play/PopBots and adaptive tutors. The object is infrastructure: low connectivity, rurality and territorial inequity.

2. State of the art: the network as a presupposition, not a detail

Three strata should be separated. The first is territorial: what is known about connectivity of urban and rural households and communities, particularly in Latin America and the Caribbean. The second is school-level: electricity, computers and internet for pedagogical purposes on the premises, and the perception of those who lead the school. The third is the algorithmic shift: what educational AI tools presuppose —almost always cloud, account, device and bandwidth— and what exists, verifiably, at the opposite pole: low-technology systems and offline-first platforms. The Starting Strong VII report (OECD, 2023a) is used here as an early childhood framework, not as a duplicated axis of “empowering the digital child”: its relevant lesson is that digitalisation in early childhood education and care (ECEC) is not neutral with respect to equity.

In the territorial stratum, ECLAC (2020), Ziegler et al. (2020), the Inter-American Dialogue et al. (2022) and Mexico’s ENDUTIH (INEGI, 2025) converge on an urban–rural and income gradient: nominal access is not meaningful connectivity, and the national average does not describe the rural household. The figures are detailed in case 1. It is a finding of access, not of AI.

In the school stratum, UNESCO (2023) and the UIS-GEM Scorecard (2024) situate electricity and pedagogical internet as SDG 4 inputs, not as a luxury. OECD (2020, 2023c), Rieble-Aubourg and Viteri (2020) and Echazarra and Radinger (2019) shift the focus to the school site: principals’ perception, connection quality, last-mile cost. The figures are detailed in case 2. None of these sources measures a chatbot in the kindergarten.

In the AI stratum, Miao et al. (2021), Miao and Holmes (2023) and the UNESCO Recommendation (2021) formulate humanist frameworks: equity, human oversight, age limits. They do not measure rural deployment or “gap closure.” Soletic and Kelly (2022) show Latin American digital policies concentrated on connectivity and, even so, insufficient. Kabugo (2020) and UNESCO (2023) document, by contrast, what operates with little network: offline-first platforms and low-intensity media (radio, television, Telesecundaria). That contrast —cloud versus broadcast— is the state of the art this article reads. The vacuum is not the absence of AI promises: it is the scarcity of evidence that those promises operate where the network is missing.

3. Review method

A critical narrative review was conducted, not a meta-analysis. The purpose was not to estimate a homogeneous effect size —non-existent among a rural connectivity index, a household survey, a principal questionnaire and an offline-platform case study— but to articulate what happens when educational AI presupposes an infrastructure that the kindergarten and the basic school, in many territories, do not have. Inclusion: (a) connectivity, electricity, devices or technical support in early childhood or basic education, with attention to rurality and low connectivity; (b) privilege 2021–2026, with still-current frameworks (ECLAC, 2020; Ziegler et al., 2020; OECD, 2020; Rieble-Aubourg and Viteri, 2020; Echazarra and Radinger, 2019; Miao et al., 2021); (c) international-organisation report, official survey, journal or working paper with verifiable DOI or institutional URL; (d) at least one case of technology that operates with limited connectivity, so as not to reduce the object to a denunciation of the divide. Themes already treated in this series were excluded as an axis (AI curriculum, multilingualism, privacy, inclusion, parental mediation, UNESCO competence, well-being, integrity, documentation, student literacy, play/PopBots, adaptive tutors).

The search was executed on 21 August 2026 (around 21:03, America/Mexico_City) on UNESCO, OECD, ECLAC, IICA, IDB, Inter-American Dialogue, INEGI, Crossref and journals with DOI. Each source was verified against at least one of those pages. Articles on “offline AI” in journals of unconsolidated editorial quality and essays of percentage gains without an early-childhood protocol were not retained as nuclear. Nuclear cases: (4) territorial connectivity; (5) the school as infrastructure; (6) what already works without the cloud. OECD (2023a) saturates the digital early-childhood framework; it is not used as a case of AI in the rural kindergarten.

The analysis distinguished three enunciative statuses. Empirical finding: what is observed in the survey, the index or the case study. Normative or policy framework: what an organisation prescribes or recommends. Pedagogical inference: the translation to the kindergarten and the low-connectivity basic school, marked as such. The limits are those of any narrative review (section 9).

4. Case 1. The territory before the model: households, rurality and meaningful connectivity

ECLAC (2020), in COVID-19 Special Report no. 7, documented that, around 2019, one third of the inhabitants of Latin America and the Caribbean had limited or no access to digital technologies. Empirical finding: 67% of urban households connected to the internet versus 23% of rural ones. In Bolivia, El Salvador, Paraguay and Peru, more than 90% of rural households had no connection; even in Chile, Costa Rica and Uruguay, only about half of rural households were connected. Across quintiles, 81% of households in the highest quintile had a connection, versus 38% of the first and 53% of the second, on average in twelve countries. Forty-two percent of those under 25 had no connection. Status: a household finding, not of early-childhood classrooms and not of AI. Converting “23% rural” into “AI will bring quality to the peasant kindergarten” would be an illegitimate inference: the source measures connectivity and affordability baskets, not language models.

Ziegler, Arias Segura, Bosio and Camacho (2020) shift the indicator from “having internet or not” to meaningful connectivity —quality, not only coverage— and construct rural and urban indices (SRCi and SUCi) for 24 countries, with 39 interviews with public and private actors and data to September 2020. Empirical finding: 71% of the urban population with meaningful connectivity services versus 36.8% of the rural (a 34-point gap). Excluding Brazil for its population weight, the mean of rural inhabitants without meaningful connectivity rises to 75%. The regional figure the authors and the IDB underline is at least 77 million rural inhabitants without internet of minimum quality. The study also cites the CIMA-IDB figure that, in eight of ten countries analysed, fewer than 15% of rural schools had sufficient bandwidth or speed. Status: a finding of network quality and of the rural school as a critical point. It does not measure AI. It does not claim that a generative assistant compensates for the absence of megabits.

The Inter-American Dialogue, the IDB and the World Bank (2022) name “complex zones” those of low density, distance or isolation, and record that, according to ECLAC, in 2020 67% of urban households had connectivity versus 23% rural. Policy-synthesis finding: most countries in the region do not meet meaningful-connectivity conditions for educational purposes under the Alliance for Affordable Internet definition; 4G penetration is put at 42%; 55% of people do not access daily connectivity. The report is not a kindergarten trial. It is a last-mile map: incentives for private investment, public investment, technological options (including satellite) and ministerial coordination. Status: a finding of an infrastructure agenda. Pedagogical inference, marked as such: any “educational AI” that requires daily use, sufficient data and an affordable device stumbles on exactly those four deficits.

In Mexico, ENDUTIH 2024 (INEGI, 2025) estimates 83.1% internet users among the population aged six and over, and 73.6% of households with the service. Territorial empirical finding: 86.9% of users in the urban sphere and 68.5% in the rural. The states with the lowest share of connected households were Chiapas (50.7%), Oaxaca (55.5%) and Guerrero (58.9%); the highest, Mexico City and Sonora (84.4%). Status: a household survey, not a kindergarten or AI sample. Soletic and Kelly (2022) recall that, in 2018, 46% of girls and boys aged five to twelve —some 31 million— lived in households without a connection. A finding of a cohort in early-childhood and basic-education age. It is not a finding that AI reaches them.

Cumulative status. Empirical finding: connectivity in Latin America is an urban–rural, income and state gradient; “meaningful connectivity” is scarcer than nominal access; tens of millions of rural inhabitants and a majority of rural households in several countries do not have the network that a cloud AI service requires. It is not a finding of child learning or of personalization. Pedagogical inference: announcing educational AI “for the region” without disaggregating territory is announcing a good that, with these numbers, will be realized mainly where a connected household, a device and data already exist. That is not delay. It is selection.

5. Case 2. The kindergarten and the school as infrastructure: electricity, pedagogical internet and what principals report

UNESCO (2023), in Technology in education: A tool on whose terms?, sets the material floor. Global empirical finding: around 2021, almost 9% of the world population lacked electricity; one in four primary schools does not have it. Worldwide, 40% of primary, 50% of lower-secondary and 65% of upper-secondary schools were connected to the internet; 47% of primary schools had computers for pedagogical purposes. During COVID-19 closures, at least 500 million students —31% worldwide, mostly the poorest (72%) and rural (70%)— could not be reached by remote learning; 91% of countries used online platforms, but these reached only a quarter of students. The cost of basic digital learning in low-income countries and of connecting all schools in lower-middle-income countries would add 50% to the SDG 4 financing gap. Status: a finding of inputs and of mass exclusion. The report does not measure AI in early childhood education. Its question —a tool, on whose terms?— is the frame with which this article reads promises of “AI for all.”

The UNESCO Institute for Statistics and the GEM Report (2024) make internet connectivity in primary schools the eighth SDG 4 benchmark indicator. Monitoring finding: it is one of the two indicators with the lowest rate of national targets submitted (32%); richer countries report more and are closer to universal coverage; two in three low-income countries have no data, and none of them meet their national benchmark in the published cut. Status: a finding of data scarcity, not of AI efficacy. Pedagogical inference, marked as such: one cannot govern a “school AI innovation” on an indicator that many systems still do not measure. Opacity is part of inequity.

Rieble-Aubourg and Viteri (2020), in IDB CIMA Brief no. 20, summarised Latin America and the Caribbean’s preparedness for online learning at the start of the pandemic. Diagnostic empirical finding: inequity in access to technology, connectivity and digital resources marks the region; the SIGED project survey showed that most countries lacked baseline digital conditions (school connectivity, platforms, virtual tutoring, resource packages and a repository). Household internet was put at 77% in the region versus 96% in the OECD. Status: a finding of systemic preparedness, not of kindergarten and not of a generative model. Soletic and Kelly (2022) confirm, in six countries (Argentina, Chile, Colombia, Costa Rica, Mexico and Uruguay), that policies concentrated on providing connectivity and equipment and that, even so, they did not guarantee equitable access; in rural areas, stable electric power and networks remain a challenge. Rieble-Aubourg and Viteri (2020), with PISA 2018, stress that having internet is not equivalent to quality sufficient for schoolwork. Status: a finding of “having internet” versus “being able to use it to teach.” A data-consuming chatbot does not operate on any connection: it operates in the subset that also has a device, an account, network quality and pedagogical mediation.

OECD (2020), in Making the Most of Technology for Learning and Training in Latin America, contributes principals’ perception and the school as a node. Empirical finding: around 51% of principals in the region report insufficient internet as a barrier to quality instruction, versus 17% in the OECD average; around 43% report shortage or inadequacy of digital technology for teaching. In Colombia, Mexico and Peru, the school is an internet provider for more than 20% of rural students. More than 41% of rural students in Peru have access to a desktop, laptop or tablet only at school; in Mexico, 27%; in Colombia, 20%. Status: a finding of school-site dependence and perceived quality, in PISA secondary, not in early childhood. It does not measure that a language model improves that quality. OECD (2023c), in Shaping Digital Education, updates quality inequalities: gaps in fast connection between rural and urban schools of 30% in Mexico, 28% in Costa Rica, 44% in Romania and 21% in Slovenia (PISA 2018); gaps in computers connected between advantaged and disadvantaged schools of 43% in Colombia and 42% in Mexico. In OECD systems, the share of principals who report that instruction is hindered by insufficient internet is, on average, 7 points higher in rural schools than in city schools, and exceeds 40 points in Colombia, Mexico, Italy and Alberta (Canada). Status: a finding of school geography. Echazarra and Radinger (2019) qualify: not every rural school lacks infrastructure, and performance gaps often shrink after controlling for socioeconomic status; the cost of delivering education in remote areas is nonetheless higher. That nuance prevents romanticizing “the rural” as a void and treating cloud AI as automatic compensation: the source does not measure it.

OECD (2023b), PISA 2022 Volume II, describes a post-disruption school world, not a kindergarten: on the OECD average, at least three in four students reported few problems of access to a device or the internet in remote learning. It is cited so as not to extrapolate the “connected OECD student” to the three-year-old in a locality without a network. The Starting Strong VII report (OECD, 2023a) does speak of early childhood: digitalisation creates opportunities and risks for ECEC; equity is transversal; the report itself recognises limits of the evidence base. A policy framework, not a rural-AI trial. Pedagogical inference: if ECEC must respond to digitalisation with equity, the first datum is not the language model, but whether the centre has light, a device and an adult who can use it without an unstable cloud.

Cumulative status. Empirical finding: electricity and pedagogical internet are not universal; in Latin America, the perception of insufficient network is three times the OECD average; rural–urban and public–private gaps in Mexico and Colombia are of tens of points; a relevant fraction of rural students only touch a computer at school; the pandemic showed that online platforms do not reach those who live far away and are poor. It is not a finding that AI improves the rural kindergarten. Pedagogical inference: an “AI in early childhood education” programme that does not condition its deployment on energy, bandwidth, working devices and technical support is not a national programme. It is a programme for the already-connected subset.

6. Case 3. What already works without the cloud —and the illegitimate leap toward educational AI

UNESCO (2023) documents a century of low-intensity technologies that did reach territories where broadband does not. Synthesis empirical finding: interactive radio instruction is used in about 40 countries. Mexico’s Telesecundaria programme —televised lessons, in-class support and teacher education— is associated with a 21% increase in secondary enrolment. During COVID-19, 70% of 101 distance-education projects in crisis used radio, television and basic telephony; Mexico expanded Telesecundaria content to all levels. Online platforms, used by 91% of countries, reached a quarter of students; the rest depended on low technology, paper and mobile phones. Status: a finding of reach, not of “AI quality.” GEM does not claim that the teleclass is equivalent to a generative tutor. It claims that, when the network is missing, what scales is not the cloud. Pedagogical inference, marked as such: citing Telesecundaria as proof that “technology reaches the rural” does not authorise replacing that technology with a language model that demands exactly what Telesecundaria did not (symmetric broadband, a cloud account, remote compute).

Kabugo (2020) studied the use of open educational resources on Kolibri —Learning Equality’s platform designed to teach with technology without continuous internet— in ten Ugandan government secondary schools, with usage logs and interviews with 25 teachers and 100 students, during COVID-19 closures. Empirical finding: teachers used Kolibri open resources to sustain teaching in contexts of scarce space, materials and science and mathematics teachers; the author proposes a model of efficient use from a discourse analysis of that use. Kolibri is offline-first: libraries of lessons, assessments, books, games and simulations that are installed and updated when there is a network and used when there is not. Status: a secondary-school case, not kindergarten; open resources and a platform, not a generative language model; Uganda, not Latin America. It does not measure that “AI closes the gap.” It does not measure algorithmic personalization. Its value for this article is negative and precise: there is evidence that offline-first design allows some digital practice where the cloud does not reach. That evidence does not transfer, as such, to a 2026 chatbot.

UNICEF Latin America and the Caribbean (n.d.) describes Giga —an initiative with the ITU— to map and connect schools. Status: a programme framework, not an AI evaluation. The Inter-American Dialogue et al. (2022) insist on fitting electricity and equipment. Miao et al. (2021), Miao and Holmes (2023) and UNESCO (2021) are frameworks: equity, age limits, human oversight. None measures a Latin American rural kindergarten with a local language model. “Offline AI” candidates in journals of unconsolidated editorial quality (2025–2026) are not retained as nuclear cases.

Cumulative status. Empirical finding: radio, television and offline-first platforms have a documented trajectory of reach in low-connectivity contexts; Kolibri has a situated case of teacher use; Giga is a mapping and connection agenda, not a learning outcome. It is not a finding that generative AI operates in the kindergarten without a network. Pedagogical inference: the leap “if Kolibri works offline, then an on-device model will personalize rural early childhood education” confuses genres. A repository of open resources installed on a local server is not a language model. A language model that fits on a device still requires energy, maintenance, periodic updating, technical literacy and, in early childhood education, a mediating adult. Where those inputs are missing, “educational AI” is not delayed: it is absent, and its relative absence concentrates the benefit in the already-connected territory.

7. Inferential framework: without a network, AI does not delay, it selects

The framework that follows is this article’s pedagogical inference, anchored in the cases and the verified frameworks. It is not a new international standard. It distinguishes five tests. If a proposal of “AI for early childhood and basic education in rural or low-connectivity contexts” does not pass them, it is not deployed as an equity policy.

7.1. Test of energy before the model. UNESCO (2023) measures that one in four primary schools has no electricity. Ziegler et al. (2020) and the Inter-American Dialogue et al. (2022) situate the rural school as a critical point of bandwidth and minimum conditions. Inference: an AI system —in the cloud or on the device— is a consumer of energy and maintenance. Where there is no stable light, there is no “innovation.” Promising AI to the kindergarten without electrification is a sequence violation, not a shortcut of modernization. This article does not claim that electrifying closes learning gaps: it claims that, without energy, the AI discourse is cosmetic.

7.2. Test of meaningful connectivity, not of the access point. ECLAC (2020) and Ziegler et al. (2020) distinguish having a signal from having quality. Soletic and Kelly (2022) recall that only a third of regional school connections reach quality for the task. Inference: a generative assistant that “works with internet” does not work with the intermittent internet of a school that shares a narrow channel. Calling “AI school” a school with a modem that opens email is an abuse of language. Where the source does not measure latency, symmetry and data, this article does not invent pedagogical viability of a language model.

7.3. Test of the device and of support, not of the software. OECD (2020) documents that a high fraction of rural students access a computer only at school, and that principals report shortage of digital technology. Inference: educational AI is not a file that is “downloaded.” It is an assemblage of hardware, accounts, updates and someone who repairs. Without territorial technical support, the device becomes furniture. It is not promised that an on-device model eliminates that dependence: Kabugo (2020) shows use of an offline platform with teachers and an installation ecosystem, not magic autonomous hardware.

7.4. Test of not substituting low-technology evidence with the cloud promise. UNESCO (2023) shows that radio and television reached those the platform did not; Telesecundaria is associated with an enrolment increase. Inference: that evidence legitimates reach policies with low-intensity media. It does not legitimate replacing them with a chatbot. Nor does it legitimate the slogan that AI “personalizes where teachers are missing”: no nuclear source measures that personalization in rural kindergartens. Miao et al. (2021) and Miao and Holmes (2023) require a human-centred approach and age limits. A child in early childhood education is not the autonomous user of a language model, even where there is fibre.

7.5. Test of inequality as design, not as delay. The 500 million unreached (UNESCO, 2023), the 77 million rural people without meaningful connectivity (Ziegler et al., 2020) and the ENDUTIH gap between Chiapas and Mexico City (INEGI, 2025) do not describe a calendar in which “AI will arrive later.” They describe a map: if AI is deployed under the presupposition of the network, it concentrates where the network already is. Inference: the risk is not rural backwardness, but selective innovation presented as universal. OECD (2023a) asks for transversal equity in ECEC: an urban broadband pilot is not an “AI in early childhood policy.”

Operationally, the framework admits: (a) mapping electricity, meaningful connectivity, devices and support before buying AI; (b) low intensity and offline-first platforms as what the evidence of reach supports; (c) on-device AI only where energy, maintenance and adult mediation are assured, as a measured trial. It rejects: (d) calling equity a deployment in the connected subset; (e) selling AI as gap closure, rural quality or teacher substitute without measurement; (f) treating the low-connectivity kindergarten as a delayed version of the cloud classroom.

8. Discussion

Three tensions organise the discussion. The first is between coverage and quality. ECLAC (2020) and INEGI (2025) show gains in access; Ziegler et al. (2020) and Soletic and Kelly (2022) show that nominal access is not meaningful connectivity. Generative AI sits at the extreme of highest network demand. Inference: each coverage point that is not quality can be narrated as “ready for AI” without being so. GEM (UNESCO, 2023) warns that the total cost of technology is underestimated. This article does not oppose book to model. It opposes sequence: without a material floor, AI spending is spending in the territory that already had a floor.

The second tension is between reach and personalization. Radio, television and Kolibri (UNESCO, 2023; Kabugo, 2020) have evidence of arriving. Algorithmic personalization has, in this corpus, null evidence for the low-connectivity kindergarten. A child not reached is not “not yet personalized”: they are excluded. Rieble-Aubourg and Viteri (2020) and OECD (2020) showed that the region was not prepared for online learning; UNESCO (2023) showed who was left out. Repeating the gesture with AI —platforms first, territory later— reiterates the design of exclusion, now with more technical prestige.

The third tension is between early childhood and connected secondary. PISA and TALIS speak of 15-year-old students (OECD, 2020, 2023b, 2023c). ECEC has another density: care, play, adult mediation, kindergartens with less staff. OECD (2023a) asks for equity; Miao and Holmes (2023), age limits. Inference: the low-connectivity kindergarten is not the place for a chatbot pilot. It is the place to ask whether there is light, an adult and a device that does not depend on a top-up the family cannot pay. Treating the three-year-old in a rural locality as an “AI user” is an error of educational level and of infrastructure.

If the object is early childhood and basic education in low connectivity, the system does not need a model that presupposes the network. It needs the authority to say that an innovation that only runs where fibre already exists is not an equity policy. Inequality is not the time the rural “takes” to copy the urban: it is the result of designing the innovation for the urban. Keeping radio, television, offline open resources and electrification as what the evidence supports, and reserving the name “educational AI” for assemblages that measure their material conditions, is the path coherent with the verified sources.

9. Limits

This review is narrative. It does not apply a full PRISMA protocol or estimate combined effects of “AI on rural learning”: that variable does not appear measured in the nuclear corpus. ECLAC (2020), Ziegler et al. (2020) and ENDUTIH (INEGI, 2025) are household and territorial reports, not kindergarten trials. UNESCO (2023) and the UIS-GEM Scorecard (2024) are global syntheses. Rieble-Aubourg and Viteri (2020) and Soletic and Kelly (2022) diagnose preparedness and policies, not AI. OECD (2020, 2023b, 2023c) speaks mainly of PISA secondary and of principals; Echazarra and Radinger (2019) use PISA 2015 and TALIS 2013. Kabugo (2020) is a Ugandan secondary case, not a Latin American early-childhood trial. Telesecundaria, in GEM, is an enrolment datum. Miao et al. (2021), Miao and Holmes (2023), UNESCO (2021) and OECD (2023a) are frameworks, not 2026 evidence of a language model in the rural kindergarten. UNICEF (n.d.) describes a connection programme. No Latin American trial of generative or on-device AI in low-connectivity early childhood education that measured learning, gap closure or teacher substitution was located with the same degree of verification. That absence is a vacuum, not proof that “AI works without a network.” The inferences of section 7 are threshold hypotheses, not evidence of national implementation. This article does not claim that AI closes the digital divide, brings quality to the rural or personalizes where teachers are missing: no nuclear source measures those promises.

10. Conclusions

In early childhood and basic education, artificial intelligence in the face of low connectivity is not resolved by announcing innovation. It is resolved by asking what infrastructure —energy, meaningful network, devices, support— is a condition of any tool presented as “educational AI.” Three families of verified evidence support the argument. Households and territories in Latin America exhibit an urban–rural and income gap that meaningful connectivity has not closed (ECLAC, 2020; Ziegler et al., 2020; Inter-American Dialogue et al., 2022; INEGI, 2025; Soletic and Kelly, 2022). Schools are not an abstract enclosure: one in four primary schools worldwide lacks electricity, a minority has quality pedagogical internet, and in Latin America principals report network insufficiency far above the OECD average, with especially wide gaps in Mexico and Colombia (UNESCO, 2023; UNESCO Institute for Statistics and GEM, 2024; OECD, 2020, 2023c; Rieble-Aubourg and Viteri, 2020; Echazarra and Radinger, 2019). What documentedly reaches low-connectivity territories is, so far, low intensity —radio, television, Telesecundaria— and, in situated cases, offline-first open-resource platforms, not generative AI in the cloud (UNESCO, 2023; Kabugo, 2020). AI and early-childhood frameworks require equity, human oversight and age limits; they do not authorise treating the disconnected kindergarten as a residual market of innovation (Miao et al., 2021; Miao and Holmes, 2023; UNESCO, 2021; OECD, 2023a).

The restrictive thesis holds. Without connectivity, energy and technical support, educational AI does not arrive late: it arrives as inequality. Where the sources do not measure that AI closes the digital divide, brings quality to the rural or personalizes where teachers are missing, this article does not invent it. The risk is not that the rural kindergarten “is left without a chatbot.” It is that an innovation designed for fibre is narrated as national policy while it is realized only where fibre already existed. Childhood in low connectivity does not need to be promised a model. It needs that what presupposes a network it does not have not be called equity.

Editorial Laboratory of NEXTECH.IA / Engineer Mitre

References

  1. Economic Commission for Latin America and the Caribbean. (2020). Universalizar el acceso a las tecnologías digitales para enfrentar los efectos del COVID-19 (COVID-19 Special Report no. 7). ECLAC. https://www.cepal.org/es/publicaciones/45938-universalizar-acceso-tecnologias-digitales-enfrentar-efectos-covid-19
  2. Inter-American Dialogue, Inter-American Development Bank, and World Bank. (2022). Conectividad educativa en zonas complejas: Desafíos y objetivos para una agenda regional. Inter-American Dialogue. https://thedialogue.org/wp-content/uploads/2024/04/Informe-Tecnico-.pdf
  3. Echazarra, A., and Radinger, T. (2019). Learning in rural schools: Insights from PISA, TALIS and the literature (OECD Education Working Papers no. 196). OECD Publishing. https://doi.org/10.1787/8b1a5cb9-en
  4. Instituto Nacional de Estadística y Geografía. (2025). Encuesta Nacional sobre Disponibilidad y Uso de Tecnologías de la Información en los Hogares (ENDUTIH) 2024 [Press release]. https://www.inegi.org.mx/contenidos/saladeprensa/boletines/2025/endutih/ENDUTIH_24.pdf
  5. Kabugo, D. (2020). Utilizing Open Education Resources to enhance students’ learning outcomes during the COVID-19 schools lockdown: A case of using Kolibri by selected government schools in Uganda. Journal of Learning for Development, 7(3), 447–458. https://doi.org/10.56059/jl4d.v7i3.465
  6. Miao, F., and Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. https://doi.org/10.54675/EWZM9535
  7. Miao, F., Holmes, W., Huang, R., and Zhang, H. (2021). AI and education: Guidance for policy-makers. UNESCO. https://doi.org/10.54675/PCSP7350
  8. Organisation for Economic Co-operation and Development. (2020). Making the most of technology for learning and training in Latin America. OECD Publishing. https://doi.org/10.1787/ce2b1a62-en
  9. Organisation for Economic Co-operation and Development. (2023a). Empowering young children in the digital age (Starting Strong). OECD Publishing. https://doi.org/10.1787/50967622-en
  10. Organisation for Economic Co-operation and Development. (2023b). PISA 2022 results (Volume II): Learning during – and from – disruption. OECD Publishing. https://doi.org/10.1787/a97db61c-en
  11. Organisation for Economic Co-operation and Development. (2023c). Shaping digital education: Enabling factors for success. OECD Publishing. https://doi.org/10.1787/bac4dc9f-en
  12. Rieble-Aubourg, S., and Viteri, A. (2020). CIMA Brief #20: COVID-19: Are we prepared for online learning? Inter-American Development Bank. https://doi.org/10.18235/0002303
  13. Soletic, Á., and Kelly, V. (2022). Políticas digitales en educación en América Latina: Tendencias emergentes y perspectivas de futuro. IIEP UNESCO and UNICEF. https://unesdoc.unesco.org/ark:/48223/pf0000381837
  14. UNESCO. (2021). Recommendation on the ethics of artificial intelligence. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000381137
  15. UNESCO. (2023). Global education monitoring report 2023: Technology in education — A tool on whose terms? UNESCO. https://doi.org/10.54676/UZQV8501
  16. UNESCO Institute for Statistics and Global Education Monitoring Report. (2024). SDG 4 scorecard: Progress report on national benchmarks on school internet connectivity. UNESCO. https://www.unesco.org/gem-report/en/publication/2024-sdg-4-scorecard-progress-report-national-benchmarks-school-internet-connectivity
  17. UNICEF Latin America and the Caribbean. (n.d.). La educación en la era de la conectividad. https://www.unicef.org/lac/la-educaci%C3%B3n-en-la-era-de-la-conectividad
  18. Ziegler, S., Arias Segura, J., Bosio, M., and Camacho, K. (2020). Conectividad rural en América Latina y el Caribe: Un puente al desarrollo sostenible en tiempos de pandemia. Inter-American Institute for Cooperation on Agriculture. https://repositorio.iica.int/handle/11324/12896