Education Technology

A more direct theory-of-change than merely providing access to ICT tools uses technology as a delivery mechanism for specific academic instruction. Such instruction is differentiated both by the content of instruction and the method of delivery. In this section, we review prominent classes of evaluated interventions.

Distance education and video-based instruction

An important class of interventions that pre-dates the computer revolution focuses on televised lessons for students. The approach originated in Italy in the late 1950s with the telescuola programme and spread across Europe, Latin America, Asia, and Africa during the 1960s and 1970s. One of the most sustained and large-scale implementations is Mexico’s telesecundaria programme, which provides distance education for lower secondary students in rural and underserved areas. These are brick-and-mortar schools that combine centrally produced high-quality televised lessons with in-person support from a single adult monitor, offering a form of blended learning. The programme has operated continuously since 1968 and, by 2016, enrolled about 1.6 million students.

These low-tech, standardised, technology-aided models are seen as potentially cost-effective because they reduce delivery costs relative to traditional secondary schools. By substituting multiple subject-specialist teachers with high-quality televised instruction, TV schools can deliver more consistent lessons and mitigate problems related to teacher shortages or absenteeism. However, scepticism is warranted. The one-size-fits-all design (broadcasting the same lessons nationwide) may not align with the learning needs of students in underserved communities who often lag behind grade-level expectations. These concerns are magnified when such models are scaled across diverse and unequal settings, where rigid standardisation risks overlooking local conditions and exacerbating existing disparities.

Two recent papers provide causal evidence on the short- and long-term effects of TV schools on student outcomes. Borghesan and Vasey (2024) estimate the short-term impacts of TV schools on learning outcomes by using an instrumental variable design to address selection into these schools, exploiting students' distance to a TV school relative to a traditional secondary school. They report large average effects on test scores after just one year of attendance, about 0.36σ in math and 0.23σ in Spanish. Fabregas and Navarro-Sola (2025) investigate the long-run effects using the staggered rollout of these TV schools for identification, and find substantial gains in educational attainment and adult earnings, explained only modestly by shifts in labour force participation or sectoral composition. The authors provide evidence that most students who attended the TV-schools would likely have otherwise remained out of school, indicating substantial extensive-margin effects on enrolment. Nevertheless, the estimated returns to an additional year of TV school education are broadly comparable to those from traditional secondary schools, with cost-benefit calculations of the TV schools of at least 3:1.

Johnston and Ksoll (2022) report results from an experiment that provided remote live instruction to students in 70 randomly selected primary schools in Ghana. After two years of treatment, they report significant positive effects of the treatment on student achievement, measured using the Early Grade Math and Reading Assessments (EGRA and EGMA). A challenge with interpreting these effects as resulting from technology alone, however, is that treated schools were also subject to greater monitoring and follow-up, facilitators faced high-powered monetary incentives, and the programme also featured changes to the instructional material as well as the hiring of teachers. Finally, Naik et al. (2020) also report positive effects from a large-scale evaluation of a satellite education programme in Karnataka in students from Grades 8-10.

Three recent papers make much greater progress in highlighting the sensitivity of video-based interventions to how they are implemented in public school systems.

Beg et al. (2022) provide evidence on the sensitivity of incorporating video-based content in regular instruction using two RCTs in Pakistan. In the first experiment, teachers in Grade 8 were provided a tablet with pre-loaded videos illustrating core material from the curriculum and classrooms were equipped with a large LED screen. The authors report gains of about 0.26σ from the intervention after a year, which are visible both in independent and official tests. In a parallel experiment, similar material was developed and pre-loaded on tablets that were then given to students in Grade 6 for individual use and self-led learning. In this arm, they find that test scores declined by 0.4σ in the treatment group (as opposed to control schools).

Second, de Barros (2023) studies an ambitious attempt to implement a blended instruction programme where secondary schools in Haryana, India, were provided infrastructure upgrades (including tablets for teachers and TVs), an application with video materials, accompanying workbooks, and related teacher training. de Barros (2023) evaluates this using a large and well-powered RCT covering 240 schools and ∼24,000 students. The study randomly assigned these 240 schools into three arms: (i) a pure ‘Control’ group, (ii) the full programme with all components described above, and (iii) a low-tech variant that only featured workbooks and teacher training, but not the associated ICT components. After 11 months, students in both treatment arms did worse than the Control group, but this difference was larger for the ICT arm, where the negative effect (-0.16σ) equalled about half relative to the progress in the control group; the equivalent decline in the Workbook only (no ICT) arm was 0.07σ and not statistically distinguishable from zero.

The evidence in Beg et al. (2022) and de Barros (2023) provides particularly compelling examples of the importance of carefully integrating the use of any technology in classrooms and schools. Beg et al. (2022) reiterates the core theme of Section 3.1 that the mere provision of ICT tools to students along with educational content may, as in this case or Malamud and Pop-Eleches (2011), lead to worse outcomes. However, they found encouraging effects from empowering teachers with technology. Yet, de Barros (2023) finds substantial negative effects from a similar intervention featuring videos and ‘smart’ classrooms. A plausible explanation that could reconcile these findings is the effect of the interventions on classroom interactions and regular teaching. de Barros (2023) shows that the ICT intervention severely reduced instructional quality, including in classroom management, feedback, clarity and the monitoring of student learning.[1]1. See also Berlinski and Busso (2017) who present short-term effects of an intervention in Honduras which promoted ‘active learning' methods in a multi-armed trial a subset of which incorporated intensive use of technology. They report negative effects for all treatment arms, and especially so in the treatment arm that provided students individual laptops for use.  Beg et al. (2022) do not have similarly detailed (or directly observed) data on classroom practices to permit a direct comparison but it seems likely that their classroom intervention did not suffer such effects; teachers in video-enabled classrooms reported an increase in student demand for feedback.[2]2. An important contextual difference is also that de Barros (2023) targets teaching in higher secondary grades while Beg et al. (2022) target Grade 6 students. Both students and teachers in secondary schools are more selected, and regular teaching is also much more closely linked to high-stakes exams. These features could result in the productivity of regular classroom instruction being very different across the two studies.

The final paper demonstrates the possibility of delivering such models at larger scales. Bianchi et al. (2022) study a large-scale reform in China that used satellite and multimedia technology to connect high-quality urban teachers with over 100 million rural primary and middle school students. Between 2004 and 2007, the government (i) equipped rural schools with TV and DVD sets containing recorded lessons, (ii) installed satellite systems to deliver updated lectures and online materials, and (iii) built computer classrooms where students could follow central lessons and learn computer skills. Importantly, the authors stress that the lectures were designed to integrate into the standard curriculum followed by rural schools, and that substantial effort had gone into training teachers on the intervention protocols and avoiding disruptions in the classroom. Using the staggered rollout for identification, the authors find long-term gains of 0.18σ in math, 0.23σ in Chinese, and 0.85 additional years of schooling. Students exposed to the programme were more likely to enter cognitively demanding occupations and earned higher wages than comparable peers.

Overall, these results highlight the need for (i) rigorous evaluations of EdTech interventions in schools, even when they superficially resemble interventions that were successful in other contexts and (ii) a careful monitoring of the effect of interventions on default classroom practices and interactions.

Computer-aided learning

That computers might be more effective if paired with specific curricular applications was realised relatively early. Even the OLPC evaluations mentioned above had frequently included bundles of educational software and training for both students and teachers. In this section, we focus on interventions where the use of technology for academic instruction was not merely enabled by providing access to technology, but actively mandated as an integral part of the programme.

This category includes interventions that differ substantially in scope, content, and implementation. In reviewing the evidence, we differentiate these instructions along two main features: (i) whether they displaced core subject instruction or supplemented it, and (ii) whether they delivered uniform content to all students or provided tailored instruction. As we discuss below, both features have been predictive of impact across studies.

Interventions supplementing instructional time in core subjects

Most evaluations of computer-aided instruction (CAI) examine initiatives that provide instruction outside regular classroom hours for the targeted subjects. This includes programmes delivered after school or during lunch breaks, initiatives that replace computer classes or extra-curricular activities, and CAL interventions that focus on homework and other forms of practice outside the classroom.

It is important to distinguish such programmes for both conceptual and practical reasons. First, they typically supplement regular classroom instruction, thus increasing the total amount of instructional time for the targeted subjects. While evaluations of such programmes are informative about whether they can help raise achievement levels, they generally do not typically identify whether equivalent gains could be achieved solely by expanding conventional (non-technology-based) instruction by the same duration. Second, by leaving default classroom instruction unchanged, these programmes minimise the potential for negative effects, such as those found by de Barros (2023). This feature is attractive for programme implementation but also limits the potential for fully integrating EdTech into everyday schooling.

Banerjee et al. (2007), in a seminal study, evaluated two interventions aimed at raising primary school achievement levels, one of which directed students to play educational games targeting basic math skills. Specifically, the computer-assisted learning programme offered grade 4 children two hours of shared computer time per week, during which they played games involving solving math problems whose difficulty level responded to their ability to solve them.[3]3. One hour was scheduled during regular class time and the other after school.  The programme, randomised across classrooms, had sizable treatment effects of ∼0.4σ, with larger initial gains observed for students in the bottom-third of the baseline test score distribution.

Muralidharan et al. (2019) present further evidence of the promise of personalised CAL programmes. They evaluate a software called Mindspark which provided differentiated instruction closely targeted to students’ actual achievement levels. Using baseline data from the programme, the authors document that such personalisation likely solves an important constraint, since most students in their sample were performing well below grade level competence. The intervention was delivered in after-school centres serving students from public middle schools in Delhi. They find sizable intention-to-treat (ITT) test score gains of ∼0.36σ in math and 0.23σ in Hindi language after 4-5 months of treatment. These gains exceed the productivity of default classroom instruction per unit of time, suggesting that the effects reflect genuine improvements in productivity, rather than merely additional instructional time.

In subsequent work, Büchel et al. (2022) aim to disentangle these two channels in a multi-arm experiment conducted in schools in El Salvador with four groups: (i) additional teacher-led classes; (ii) additional CAL classes monitored by a supervisor; (iii) additional CAL classes instructed by a teacher; and (iv) a pure control group. The CAL instruction was provided using an offline version of content from Khan Academy and, while not explicitly adaptive, was tailored ex ante to students' achievement levels, thus providing a degree of personalisation. Instruction focused on math and was during twice-weekly additional classes. They find large ITT effects of >0.2σ in the groups assigned to CAL instruction. Students assigned to teacher-led classes with CAL improved about the same as students assigned to supervisor-led classes with CAL. While all treated groups did better compared to students without the additional instruction, those assigned to teachers without CAL actually did worse than those with teachers or supervisors with CAL, suggesting the CAL-supported instruction to students was actually better than the instruction offered by teachers alone. CAL lessons also appeared to be less sensitive to student ability, class size or teacher pedagogical knowledge than non-CAL lessons led by teachers.[4]4. A related experiment with a similar design is Ma et al. (2024) who also randomise students in China to receive either (non-tech) additional workbook-based sessions or CAL-based sessions. They do not report finding gains on average from either treatment (relative to a pure control). The difference between the CAL treatment and the workbooks only arm is always positive but not statistically significant at conventional levels. Several other experiments in China, however, do seem to find positive effects of CAL, at least in math, when delivered as supplementary instruction (Mo et al. 2013, 2014, Lai et al. 2013, 2015). See Rodriguez-Segura (2022) for further discussion.

Finally, Bettinger et al. (2023) present results from a large-scale experiment in Russia which focuses on the use of CAL for homework. The study compares two alternative treatments featuring ‘low’ (45 minutes) and ‘high’ (90 minutes) dosage of CAL with a pure control. They report positive treatment effects of CAL for math and, in some specifications, for Russian language, but no meaningful difference between the effect of high versus low dosage.[5]5. They interpret the equivalence of gains across the low- and high-dosage arms as demonstrating diminishing returns to CAL. An important caution, though, is that the intervention focused solely on non-personalised practice items completed as homework. It is an empirical question whether similar diminishing returns would be observed in other types of CAL programmes.

Interventions substituting instructional time in core subjects

While supplementary uses of CAL have typically proven promising in improving student achievement, evidence on CAL as a replacement for regular instruction is both more scarce and more mixed. This is concerning, since integrating EdTech use into school-based is often the most feasible way to scale in public schooling systems, especially in LMICs where computer/tablet ownership at home is scarce.

In an early contribution, Linden (2008) presented results from two experiments in India. The first experiment used a CAL software to supplement class time; in the second, he used CAL to replace regular in-class instruction. The content was not personalised or adaptive but, rather, focused entirely on the regular curriculum. As in the broader literature, he documents significant positive effects for supplementary CAL, especially for initially lower-achieving students. However, the model where students were ”pulled out” of regular instruction to participate in CAL sessions shows very large negative effects relative to default classrooms. Importantly, the schools covered in the experiment were described as high-performing institutions run by a motivated NGO. Thus, in this case, the use of CAL was actively worse than regular classroom instruction.

More recently, two experiments have evaluated the use of mature CAL software on student achievement when delivered as a direct substitute for classroom instruction.

In Brazil, Ferman et al. (2019) evaluated an intervention that provided instruction through Khan Academy in 157 primary schools. The programme replaced one classroom period per week with a session in which students worked on the Khan Academy platform in math instead of receiving regular math classroom instruction. They find no effect of the treatment on math scores, with point estimates very close to zero. They hypothesise, and present some suggestive evidence, that the effects were partly a result of inadequate implementation.

Much more encouraging evidence is reported by Muralidharan and Singh (2025), who evaluate an attempt to adapt Mindspark, the software evaluated in Muralidharan et al. (2019), for use in public school classrooms, directly displacing regular subject teaching. The intervention was run for three years in primary and middle school grades in government schools in Rajasthan, India. The intervention set up dedicated Mindspark labs in treatment schools, and students were assigned to spend three classroom periods each per week studying Math and Hindi. This represents a displacement of 25-50% of scheduled instruction in these subjects. After 18 months, the core experiment showed ITT effect sizes of ∼0.2 σ in both math and language, representing a substantial increase in instructional productivity over business-as-usual in the control group. These gains were similar in magnitude across the achievement distribution and by gender and socioeconomic status. The authors also show that distance relative to curriculum narrowed in treatment schools over this period. The intervention initially relied on an external resource person (lab-in-charge) in each school to maintain the computer lab and provide support to teachers and students  during the first two academic years, but this support was severely reduced in the third year. Consequently, usage declined and, while the programme continued to have positive effects, its productivity was lower than in the period with dedicated supervision.[6]6. Positive gains are also reported by Araya et al. (2025), who evaluate a programme integrating a math learning platform into regular instruction in 24 primary schools in Santiago, Chile. Supported by external coordinators, the programme replaced part of classroom teaching and raised math scores by 0.27σ, while also improving students' attitudes toward technology and learning. Cardim et al. (2023) evaluate the ProFuturo programme in Luanda, Angola and find improvements in teaching practices, attendance, and student engagement, but learning gains are limited to Science (∼0.07σ),the most popular subject on the platform, with no effects on other core subjects.

The results of Muralidharan and Singh (2025) suggest that substantial gains from adopting CAL in schools might be possible over the entire span of elementary education. However, realising these gains will likely require a substantial period of adaptation in order to avoid classroom disruption. They also suggest that (i) data on student usage, automatically logged by many digital learning programmes, might be exceptionally useful for both monitoring programme implementation and, ex post, understanding the relationship with estimated effects on achievement; and (ii) effectively scaling the use of CAL will require very careful consideration to the complementarities between adult supervision and computer use.[7]7. An important challenge in all attempts to scale CAL in government settings is ensuring effective programme implementation. Mo et al. (2020) present evidence from China showing that, whereas a CAL programme implemented by NGOs improved student achievement, it did not when implemented by government authorities following their default administrative procedures.

Several additional trials are currently ongoing to evaluate the use of CAL software in class hours in LMICs, and we expect the evidence base in this area to grow substantially in the near future. For instance, in India, large-scale trials are currently underway to evaluate Convegenius in Andhra Pradesh and Khan Academy in Uttar Pradesh, both influenced in part by the earlier work of Muralidharan and Singh (2025) in Rajasthan. Preliminary results from the trial of Convegenius also appear extremely promising, with gains of ∼0.4σ after two years of implementation in middle schools.[8]8. We do not describe the results in detail since the research paper is not yet available in the public domain. The core results cited here reflect public results dissemination by the research team, led by Prof. Michael Kremer, in September 2025 and widely reported in the Indian media. Preliminary results on a trial of Khan Academy in India using external lab-in-charges are also reported to be very promising, while another ongoing randomised experiment suggests that attempts to train existing teachers to adapt their instruction to using Khan Academy in Puerto Rico are less successful. This review will be updated after these studies are in the public domain.  Taken together, these results indicate that, as a class of interventions, personalised computer-adaptive learning might be very promising for improving achievement in LMIC education systems, especially with appropriate support mechanisms in place.

Remote tutoring on smartphones and other devices

Mobile phone use has expanded very rapidly in developing countries, with ownership being near-universal across households in many LMICs. This widespread access has expanded the set of viable education interventions that can be delivered directly to children or their parents. In richer countries, in Europe and Latin America, it is additionally possible to deliver these interventions through tablets and laptops, which are more broadly available.

The most intensively studied set of such interventions is phone-based tutoring, especially in periods of school closures during the COVID-19 pandemic. An early contribution was Angrist et al. (2022), who first experimented with providing phone-based support to students in Botswana. About 4500 students were assigned to either (i) a pure control group, (ii) weekly SMS-based practice problems, or (iii) weekly SMS messages combined with phone calls. They document learning gains of 0.12σ. Following these encouraging results, the authors partnered with multiple organisations to replicate these results in five additional countries — Kenya, Uganda, India, Nepal and the Philippines — and continued to find similarly encouraging effects from the phone call treatment (Angrist et al. 2023).

Contemporaneous to Angrist et al. (2022, 2023), several research teams independently set up experiments that provided remote tutoring to children: see, e.g. Gortazar et al. (2024) in Spain, Carlana and La Ferrara (2024) in Italy, Albornoz et al. (2025) in seven Latin American countries (Argentina, Brazil, El Salvador, Guatemala, Mexico, Paraguay and Peru), Ojha and Yadav (2023) in India, and Zoido et al. (2024) in El Salvador. Across these studies, it appears that remote tutoring has a positive effect on student achievement. However, not all experiments have been as positive. Crawfurd et al. (2023) present null results of a large RCT of phone-based tutoring in Sierra Leone, regardless of whether the lessons were delivered by public-school teachers or private-school teachers. Similarly, Schueler and Rodriguez-Segura (2023) document no effects on math performance after a month of tutoring and negative effects following students’ return to school.

Taken together, this evidence suggests that phone-based tutoring may represent a very promising way to improve student achievement in LMICs. However, there is still substantial uncertainty on multiple grounds. First, even for nominally similar interventions, there are very substantial differences in effect sizes (as measured by internally standardised test scores) that cannot be explained by observed sample characteristics (Angrist et al. 2023). It is possible that they are explained by differences in implementation quality, but we are not aware of systematic documentation, collection of data, or a careful classification of what was transmitted and how across interventions.

Second, the bulk of the evidence was generated during COVID-19 school closures, an environment highly atypical in multiple dimensions. Teachers often had spare capacity to provide remedial instruction during school closures; parents and children were more likely to be home together during lock-downs (which is relevant, since many children are unlikely to have independent phone access); and the absence of regular teaching may have affected the measured degree of student achievement, as control-group students were effectively seeing only a skill depreciation without receiving any further inputs.

Third, most of the evidence comes from one-off interventions of relatively short duration. As a consequence, it is unclear whether these effects would continue to accumulate over time if the treatments were sustained over a longer period. Student and tutor engagement may drop off over time, short audio calls may simply be unable to provide sufficient scaffolding for skills and more advanced learning beyond basic remediation.

Fourth, additional uncertainty results from the reliance of several studies on phone-based assessments (see, e.g. Crawfurd et al. 2023, Angrist et al. 2023). These were often the only feasible choice during the pandemic. While the internal psychometric properties of these tests are often promising, and phone-based assessments may prove to be a cost-effective way for monitoring learning in broader settings, the concern here arises from the overlap in the mode of administration between the intervention and the assessment. In an environment where students are not accustomed to answering math or language questions on the phone, repeated exposure to a phone-based intervention may induce differences between treated and control students solely arising from familiarity with the testing medium rather than the genuine learning gains.[9]9. This point has broader relevance for EdTech studies. For instance, while tablet-based tests are a widely accepted and validated testing metric, these are not suitable for studies evaluating interventions where the treatment group has substantial engagement with tablets/laptops but the control group does not. To avoid conflating treatment effects and familiarity with the mode of administration, it is better to collect learning outcomes through traditional pen-and-paper or oral tests (Muralidharan et al. 2019).

Uncertainty along these dimensions mainly reflects the recency of this mode of intervention as an area of mainstream interest. Overall, the evidence thus far is promising, and many of these current unknowns about the limits and scalability of these models will be resolved as further experiments are conducted, ideally at large scales, over longer durations, and in post-COVID settings.[10]10. It is likely that new evidence on this topic will emerge quickly. E.g. Carlana and La Ferrara (2024) present evidence from Italy that the specific tutoring programme they evaluate continued to be effective in post-pandemic settings. Similarly, the most recent version of Angrist et al. (2023) briefly discuss positive results from attempts to scale up phone-based tutoring in Karnataka, a state in India.

Text- and WhatsApp-based interventions for learning at home

While the previous section mostly related to personalised tutoring delivered over phone calls, the spread of mobile telephony has also enabled other forms of educational interventions.

One such model is to centrally deliver instructional material, without the involvement of tutors, directly to students or their parents. A clear example is the transmission of instructional material through SMS or, more recently, WhatsApp. This model is appealing on many grounds: (i) the lack of tutors makes it cheaper and logistically simpler, making it easier to scale; (ii) the spread of mobile phones, including smartphones, has greatly expanded the reach of such interventions and reduced concerns of social exclusion compared to a decade ago; and (iii) especially with the growing proliferation of internet-enabled devices, which allow for interventions platforms such as WhatsApp, video applications, and chatbots have dramatically increased the range and interactivity of interventions beyond what could previously be delivered through SMS or interactive voice response systems (IVRS).

Two recent examples of such interventions — and of the need for further evaluations in this space — are presented by Arteaga et al. (2025) and Keskar et al. (2025) in India. Both target parents, intending to provide material to support cognitive stimulation of young children. Evidence that interventions focused on such cognitive material can be highly effective comes from a large literature that studies home-visiting programmes in multiple contexts, which consistently find large gains in child development. The primary challenge to scaling such efforts, however, has been the requirement of a substantial workforce of community volunteers or childcare support staff to conduct home visits with fidelity.

Arteaga et al. (2025) evaluate a programme in Uttarakhand, India, that delivered automated phone calls offering child-rearing advice to caregivers of 6-to-30-month-old children. The recordings aimed to create awareness of child-rearing best practices to support child development. The content was organised into 18 modules, which build on lessons from effective home-based programmes for supporting early childhood development.[11]11. Specifically, the modules were: (a) the importance of early years of development; (b) embedding talk, care, and play into everyday life; (c) using art as a medium for learning; (d) setting up the home for learning; (e) managing screen time; (f) enabling learning through expeditions; (g) building an emotional bond; (h) creating an emotionally secure environment; (i) caring for parental well-being; (j) managing difficult behaviour; (k) narrating stories and having conversations; (l) supporting abilities through growth periods; (m) fostering deep and secure sibling bonds; (n) understanding nutritional relationships; (o) learning independence, empathy, and responsibility; (p) promoting physical development through play; (q) imparting experiential learning; and (r) a review of important concepts.  Each module required participants to complete four calls at a day and time of their choosing. The programme was implemented by a local NGO, Dost (meaning ‘friend’ in Hindi), which ordinarily delivers this programme to 100,000 parents in four states. To sustain take-up, Dost also conducted ‘live’ (i.e. non-pre-recorded) calls to keep caregivers engaged with the intervention. Take-up was high: by the end of the study, the average caregiver had completed nearly 31 calls and had been exposed to almost 69 minutes of programme content. Despite this engagement, 10 months after the randomisation the authors report no improvements in child outcomes, and a potential deterioration in caregiver outcomes.

Keskar et al. (2025) similarly focus on an intervention designed to help parents support the cognitive development of preschool-aged children (3-6 years). The programme, developed by Rocket Learning, an Indian non-profit, was already being delivered to over one million children at the time of the study (and about four million now). It is centred around WhatsApp groups which include parents of children enrolled in ICDS centres (anganwadis, which are public centres to support early childhood development) and the preschool worker. The intervention sends multiple activities to parents to do with their children and encourages them to post videos or messages signalling completion. The authors evaluated the programme in Maharashtra state over two successive cluster-randomised trials, randomised at scale in a population of ∼67,000 children. In the first trial, they found large and significant increases in programme usage and reported parent-child interactions, but no detectable effects on student achievement after nine months of treatment. Importantly, by this point, the base intervention had been scaled to a reported ∼1 million students.

Informed by these results, Rocket Learning then adapted the programme design to increase intensity, sending more frequent and personalised messages to parents regarding their child’s learning progress, as well as providing preschool workers with more content, guidance, and peer-support activities. This latter variant significantly improved both usage and also child cognitive outcomes, with gains of 0.12–0.2σ over both business-as-usual ICDS centres and the ‘base’ intervention, which continued to show no effects on cognitive outcomes. The additional components were cheap (∼USD 1.2 per child per year) but appear to have been critical for actually affecting child outcomes. In subsequent periods, these iterations have reportedly been added as default in Rocket Learning operations (although not yet subject to further RCTs at scale).[12]12. For a good example of the possibility to improve effectiveness through repeated iterations on a base programme, see Angrist et al. (2025). It is likely that EdTech interventions, with a continuous generation of usage metrics, will be particularly well-suited to such A/B tests.

Bloomfield et al. (2025) provide complementary evidence from Uruguay on a multi-component, phone-based parenting intervention combining human interaction and technology-enabled tools. Their randomised evaluation covered 1,360 low-income families with children aged 0–3 enrolled in the Uruguay Crece Contigo programme. The intervention integrated weekly teleoperator calls with WhatsApp text and audio messages, a chatbot offering 24/7 guidance, and an AI-based feedback system that analysed uploaded parent–child conversations to promote language stimulation. After eight months, treated families showed higher access to social benefits (∼0.30σ), increased parental engagement in cognitively stimulating activities (∼0.19σ), greater knowledge of language development (∼0.15σ), and reduced stress (∼0.20σ) relative to controls. These findings underscore that coupling scalable digital outreach with sustained human support can yield meaningful improvements in early childhood environments and parental well-being in LMIC contexts

The results in Arteaga et al. (2025), Keskar et al. (2025) and Bloomfield et al. (2025) illustrate three patterns that are likely to be important for EdTech more broadly than the specific setting. First, even interventions with a well-validated theory-of-change might lose substantial efficacy in the translation from in-person implementation models to EdTech-based implementation. Second, in the absence of efficacy trials, neither usage nor widespread adoption by governments is necessarily a reliable marker of likely welfare impact. Third, however, for programmes that do clear the logistical and institutional barriers to scaling, there may be very high returns to iteratively developing and evaluating low-cost adaptations that might lead to improved efficacy.[13]13. For an example of a successful text-based intervention that promoted uptake of a proven intervention, see Gallego et al. (2025). See also Angrist et al. (2025) who report A/B tests across many refinements of phone-based tutoring to discover variants that are either more effective and/or more cost-effective.

For full reference list see the end of the conclusion chapter.