The Foundational Learning Crisis

Foundational Literacy and Numeracy (FLN) are the basic skills every child should master by the end of Grade 3: understanding short texts, writing simple sentences, and performing basic arithmetic (UNICEF, 2022). Yet in low- and middle-income countries, about 70 percent of ten-year-olds cannot read and understand a short passage—a figure that rose sharply after COVID-19 (World Bank, 2022). Children who miss these skills early rarely catch up, limiting later learning and increasing the risk of dropout (UNICEF, 2022). Weak FLN ripples through a person’s life and a nation’s economy, constraining skill development, employment, and long-term growth (Obiakor & Newman, 2022).

Sri Lanka reflects this pattern. Although more than 93 percent of the country’s population can read and write (Department of Census & Statistics, 2021), only 6 percent of Grade 3 students meet all the expected foundational literacy benchmarks, and 7 percent meet all numeracy benchmarks[i]. This is concerning because Grades 3 and 4 are when children move from learning to read to reading to learn (Reutzel & Cooter, 2007); falling behind at this stage blocks later progress. Outdated methods of instruction, limited teacher preparation for early-grade pedagogy, and wide disparities among schools all deepen the problem. Pandemic-related closures and the economic downturn have further eroded learning time (World Bank, 2022). Without decisive action, Sri Lanka risks losing a generation of learners and weakening the foundation of its human-capital base.

What Has Been Tried: Strengths and Limits of Traditional Approaches

Efforts to strengthen foundational learning have taken various forms. One popular approach has been creating child-centered classrooms where students learn through play and exploration. Play-based learning allows children to experiment, solve problems, and construct meaning—skills that are difficult to build through rote instruction. Evidence shows that such approaches yield better outcomes in early literacy and numeracy (Ali et al., 2018; Vasoya & Vansdadiya, 2023).

Teacher capacity is another pillar of effective FLN. Well-trained teachers who understand early childhood development and appropriate pedagogy are better equipped to nurture foundational skills. Continuous professional development and coaching—grounded in classroom observation and formative assessment—enable teachers to tailor instruction to individual learning levels (Zulkarnaen & Zulfakar, 2021).

Family and community engagement have also been shown to play a vital role. Reading at home, storytelling, and community literacy programs help children practice and consolidate what they learn in school. Such involvement reinforces motivation and builds supportive environments for early learners (Epstein & Salinas, 2004; Kumar, 2024).

Across these strategies, a recurring principle is assessment-informed instruction: teaching that responds to evidence of where each child stands. The Teaching at the Right Level (TaRL) approach operationalizes this idea by grouping students by competence rather than age or grade and providing targeted remediation. Studies across several countries show that TaRL can deliver rapid gains in reading and numeracy, particularly in classrooms with diverse learning levels (Muammar et al., 2023; Global Partnership for Education, 2023).

While each of these approaches has its strengths, their true value emerges when they work together. A well-trained teacher who observes and assesses students can identify each child’s level, strengths, and gaps. Once these learning levels are understood—following the Teaching at the Right Level (TaRL) logic—the teacher can design lessons using play-based methods, targeted exercises, and formative feedback to meet each child where they are. Parents, in turn, can be engaged through feedback and simple home-based activities such as shared reading and storytelling, reinforcing what happens in the classroom.

Yet when one imagines this ideal system operating at scale, its limits become clear. Success depends on the quality of teacher training, the availability of classroom resources, and the consistency of monitoring—all of which vary widely across schools. These challenges are even more acute in developing and underserved settings. They highlight the need for complementary tools and systems that can sustain personalization and feedback at scale, supporting teachers rather than adding to their burden.

Technology in Foundational Learning: Promise and Limitations

Digital tools have begun to bridge some of these gaps by extending the reach of traditional FLN strategies. Interactive games and simulations make learning more engaging and concrete, allowing children to grasp abstract literacy and numeracy concepts through play (Miller, 2018). Online platforms enable continuous teacher training and virtual coaching at a scale that face-to-face programs often cannot. Mobile applications and low-cost devices bring learning into homes and communities, while digital assessments allow teachers to track student progress in real time and adjust instruction accordingly (Vasoya & Vansdadiya, 2023; Global Partnership for Education, 2023).

Yet, most of these technologies digitize existing practices. They tend to offer the same content to all learners, sometimes rely on consistent connectivity, and depend on teachers already equipped with digital and pedagogical skills to guide students using these tools. As a result, their impact is often uneven and difficult to sustain. These limitations point to the need for a new generation of tools—those that can adapt to individual learners, provide instant feedback, and support teachers in delivering truly personalized instruction.

AI: Extending and Amplifying Proven Methods

Educational psychologist Benjamin Bloom posed a striking question (Bloom, 1984): how can ordinary classroom teaching achieve the same learning gains as one-on-one human tutoring? His study found that tutored students performed, on average, two standard deviations higher than those in conventional classrooms—a difference so large that the average tutored child outperformed 98 percent of peers.

This insight—known as the 2-sigma problem—showed that personalized instruction remains the most effective path to mastery learning, but it demands constant feedback, tailored pacing, and sustained teacher attention; conditions rarely achieved even in OECD classrooms, and almost unattainable in resource-constrained developing contexts. AI offers a way to approximate that level of personalization within the practical limits of real classrooms. By interpreting learner data, adapting tasks dynamically, and providing teachers with timely insights, AI can extend the benefits of individualized tutoring while keeping teachers central to the learning process.

In an ideal world, every teacher could give each child sustained, individualized attention. In reality, time and resources make that impossible in most classrooms.

This is the role AI can play: creating a scalable, relatively low-cost means of providing children in foundational grades with personalized education. Students can access digital learning tools in multiple settings—during school hours in computer labs, through shared devices, or on low-cost tablets lent by schools or community centers. These platforms can deliver personalized exercises and feedback through audio, animations, and gamified learning, sometimes even through conversational AI chatbots that guide students step by step. Importantly, there are numerous ways to adapt these tools to local realities—for instance, by using lightweight mobile applications, offline-first designs, or even messaging platforms such as WhatsApp. Projects like Rori (Henkel et al., 2024) in Africa already demonstrate that AI-assisted learning can reach children effectively over low-bandwidth networks, showing that personalization does not have to depend on expensive infrastructure.

Use of personalized AI for FLN is no longer speculative. In India, Wadhwani AI’s Oral Reading Fluency tool is being used across 30,000 schools in Gujarat (The Hindu, n.d.) to automate reading assessments and help teachers identify struggling students in real time. EkStep’s AXL platform, piloted in close to 5,000 schools in Telangana and Karnataka (Times of India, n.d.), combines AI-driven personalization with teacher support, showing how adaptive learning in literacy and numeracy can scale within public systems. Even in high-resource contexts like the United States, Alpha School’s TimeBack system demonstrates how AI can manage student engagement and optimize learning time (See Annex A for further details).

Beyond Personalization: Other Ways AI Can Strengthen Foundational Learning

While the greatest promise of AI in foundational learning lies in its ability to personalize instruction for each child, its potential in improving FLN reaches further.

1. AI-driven tools can strengthen teacher capacity by serving as mentors and creative assistants—helping educators design lessons, generate visual and audio materials, and adapt instruction to children’s diverse learning styles. Generative models can lower the cost and time of developing engaging, contextually relevant classroom resources [ii].
2. For administrators and policymakers, it can help identify struggling schools, allocate teachers more equitably, and monitor progress against FLN benchmarks [iii].
3. Even communities can benefit from automated communication systems that can deliver targeted messages to parents, reinforcing literacy and numeracy practices at home. Together, these applications expand the ecosystem that supports each learner, ensuring that personalization is not confined to the screen but integrated across the teaching–learning process [iv].

A Way Forward for Sri Lanka: Building, Testing, and Scaling Responsibly

Given the risks of delay and fading momentum, it is pragmatic to begin with a focused, achievable effort that generates evidence and builds institutional capability while addressing a real educational need. Sri Lanka could start by developing and testing an AI-assisted foundational learning platform in English[v], where language tools, content, and infrastructure for AI applications are already mature[vi]. This low-lift starting point would demonstrate how adaptive learning can improve early-grade literacy, while creating space for teachers, technologists, and policymakers to learn together how to design, govern, and evaluate such systems. The roadmap that follows—Build, Test, and Scale—outlines this progression, beginning with a rapid, DPI-ready English prototype that can later be adapted across languages and other use cases.

1. Build: Develop the Core Building Blocks of an AI-Enabled FLN Ecosystem
   The first step is to build a working AI-enabled learning application for early-grade English, designed to fit Sri Lanka’s classrooms and align with the national curriculum. Because English already benefits from mature language technologies, this phase will focus on integrating existing speech-recognition and text-to-speech tools, developing adaptive exercises, and ensuring the platform functions offline where connectivity is limited. The system will be co-designed with teachers and pedagogy experts so that features such as feedback, assessment, and content progression align with classroom realities and proven methods like Teaching at the Right Level. Built on open standards and modular architecture, the prototype will serve as a public digital good that can be extended and localized over time.
2. Test: Pilot, Evaluate, and Refine Through Real-Classroom Experiments.
   The next step is to test the tools in real classrooms before expanding them. Pilots should be small, deliberate, and structured to generate lessons that can guide improvement and future scale-up. They should track both learning progress and practical usability—how teachers and students interact with the technology, what challenges arise, and how the design of the tools can be refined. Pilots should also account for the diversity of Sri Lanka’s classrooms, so that the findings are broadly representative and useful for scaling.
3. Scale: Establish Standards, Build Capacity, and Institutionalize.
   Once the tools have been tested and refined, the next step is to expand their use across schools in a structured and sustainable way. The government’s role will be to set clear standards for pedagogy, data protection, and child safety, and to establish certification processes that determine which tools can be deployed in classrooms. Scaling should move in tandem with teacher training, local-language content development, and reliable access to devices. Evidence from pilots and early deployments should feed directly into policy and funding decisions, allowing lessons from the classroom to shape future design and implementation. 

For long-term impact, AI for foundational learning must be embedded within Sri Lanka’s education policy and treated as part of its core digital infrastructure. The underlying datasets, speech models, and language tools should remain open, secure, and adaptable, so that innovation can continue without duplication. Sustainable financing will require a mix of public investment, donor support, and private participation, with predictable funding for system maintenance, teacher development, and evaluation. Local universities and research centers should be central to this ecosystem, refining language technologies and generating evidence that builds confidence and ensures policy continuity.

Watch this space!

This piece explores one high-leverage application of AI in education—FLN. The broader agenda is much wider. We have not, by design, addressed in depth the many legitimate concerns about AI in classrooms—questions of bias, privacy, , and equity, as well as the growing debate about children’s relationship with technology itself: screen time, attention, and digital wellbeing. Nor have we examined opportunities that extend beyond early-grade learning, or how policymakers might approach AI in K–12 education more broadly. These themes deserve sustained and careful treatment. Future articles will explore these dimensions in depth, and this page will be updated with links and cross-references as that work unfolds.

Authors: Merl Chandana* (Research Manager and Team Lead: Data, Algorithms, and Policy, LIRNEasia), Anish Fonseka (Junior Researcher, LIRNEasia) & Nipuni Habaragamuwa (Researcher, LIRNEasia)

---

[i] This national survey assessed Grade 3 students’ literacy and numeracy skills in Sri Lanka (2021). A sample of 10,600 students was drawn from the national Grade 3 population of 332,231. The literacy tool measured 28 foundational skills (listening, speaking, reading, and writing) in Sinhala & Tamil, while the numeracy tool measured 30 foundational skills (pre-mathematical concepts, numbers, measurement, time, currency, and shapes & space). The full report can be found HERE.

[ii] Early pilots show this potential in practice. In Chile, Elige Educar’s teacher chatbots offered classroom and emotional support with high satisfaction). In the U.S., teachers using TeachFX asked 20 percent more focusing questions after receiving AI-driven feedback. Tools like MagicSchool.ai now generate lesson plans and assessments in minutes, lowering preparation time and expanding creative options. These findings were summarized by the World Bank in this report, on pg. 11, 12-13, and 15 respectively.

[iii] In Chile, UPlanner analyzes student retention risks and supports targeted interventions, with positive results reported by the Pontifical Catholic University of Chile. In Ecuador, the “I Want to Be a Teacher” system uses an algorithmic matching process to assign teachers more equitably, balancing candidate preferences and school needs.

[iv] The READY4K! program in the United States, though not AI-based, showed how simple text messages to parents can improve early literacy outcomes. Similar approaches could be scaled and personalized further with AI—tailoring messages to each child’s progress and home context.

[v] While Sri Lanka does not have directly comparable FLN scores for English, a national assessment conducted by the Ministry of Education in 2022 among Grade 4 students found that more than 50 percent of students scored below 60 marks in English. Report can be found at: https://heyzine.com/flip-book/c3a177531c.html#page/1

[vi] Extending this model to Sinhala and Tamil will require deeper language infrastructure work, including corpus development, phoneme-level annotation, and fine-tuning of ASR/TTS models for local speech patterns. These languages lack the mature datasets and NLP tools available for English, making the technical lift heavier and timelines longer. The English prototype, however, will provide a tested reference architecture—allowing future teams to focus on language adaptation rather than rebuilding core systems.

Annex A: Global Examples of AI-Driven Personalized Learning for Literacy and Numeracy

EkStep AXL – Assisted Language and Math Learning (India)

EkStep’s AXL platform applies AI to the dual challenge of classroom heterogeneity and teacher workload. It combines a teacher assistant, which helps group students by learning level and recommend activities, with a student assistant, a conversational interface that provides level-appropriate exercises in local languages. Pilots in over 5,000 schools in Telangana and Karnataka cover several hundred government schools and are integrated with the national Sunbird/DIKSHA digital-public-infrastructure backbone. Teachers report that AXL simplifies classroom differentiation and that students, especially those in lower grades, engage more confidently with interactive, judgment-free practice. Although independent impact evaluations are still limited, AXL’s design—teacher-centered, open-architecture, multilingual, and privacy-conscious—illustrates how AI can scale proven approaches such as Teaching at the Right Level within public-school systems. Its open-source model also lowers barriers for state adaptation and local content creation.

Wadhwani AI – Oral Reading Fluency (Gujarat, India)

Wadhwani AI’s Oral Reading Fluency (ORF) tool automates the labor-intensive task of assessing how well children can read aloud. Using speech recognition tailored to Indian languages and background-noise conditions, the tool records a student reading a short passage and instantly measures accuracy, speed, and error types. Teachers receive immediate feedback and digital reports, replacing manual checklists that can only be done infrequently. Since its launch, the system has been deployed across all 33 districts of Gujarat, reaching over 30,000 schools and completing more than six million assessments. Early field evidence suggests that frequent, low-burden diagnostics help teachers identify struggling readers faster and focus on remedial attention where it is most needed. While formal evaluations of learning outcomes are ongoing, the Gujarat experience demonstrates how AI can convert a paper-based, irregular process into a routine, data-driven element of classroom practice—an essential foundation for any personalized reading intervention.

Alpha School and the TimeBack System (United States)

Alpha School, a network of private schools in the United States, experiments with AI to compress formal instruction into what it calls a “two-hour school day.” Its proprietary TimeBack system monitors student engagement in real time—detecting when learners switch tasks, pause for too long, or drift off-task—and issues prompts that nudge them back to focus. Combined with adaptive learning apps for math, language, and science, this creates a continuous feedback loop that manages both content and attention. Internal reports claim that Alpha students’ progress at roughly twice the national rate, though independent verification is pending. The approach demonstrates how AI can optimize learning time and maintain motivation, not merely personalize content delivery. At the same time, it raises important questions about privacy, data use, and the cultural fit of attention-tracking tools in more constrained or lower-resource environments.

References

Ali, E., Constantino, K. M., Hussain, A., & Akhtar, Z. (2018). The effects of play-based learning on early childhood education and development. Journal of Evolution of Medical and Dental Sciences, 7(43), 6808–6811. https://doi.org/10.14260/jemds/2018/1528

Miller, T. (2018). Developing numeracy skills using interactive technology in a play-based learning environment. International Journal of STEM Education, 5(1), 39. https://doi.org/10.1186/s40594-018-0131-2

Muammar, M., Ruqoiyyah, S., & Ningsih, N. S. (2023). Implementing the Teaching at the Right Level (TaRL) approach to improve elementary students’ initial reading skills. Journal of Languages and Language Teaching, 11(4), 610–625. https://doi.org/10.33394/jollt.v11i4.7942

Obiakor, T., & Newman, K. (2022). Education and employability: The critical role of foundational skills. RISE Insight, 1–20. https://riseprogramme.org/sites/default/files/2022-11/Education_Employability_Critical_Role_Foundational_Skills.pdf

UNICEF. (2022). Are children really learning? Exploring foundational skills in the midst of a learning crisis. https://data.unicef.org/resources/are-children-really-learning-foundational-skills-report/

The Science Publications. (2023). Journal of Social Sciences, 2023, 92–95. https://thescipub.com/pdf/jssp.2023.92.95.pdf

Wijesinghe, N. (2023, July 31). Direct challenge to Sri Lanka’s educational leaders. FT.lk. https://www.ft.lk/columns/Direct-challenge-to-Sri-Lanka-s-educational-leaders/4-781646

Reading Rockets. (n.d.). How children learn to read: Typical reading development. Reading Rockets. https://www.readingrockets.org/reading-101/how-children-learn-read/typical-reading-development

Department of Census and Statistics, Sri Lanka. (2021). Literacy rate of population aged 10 years and above, by district and sex. https://www.statistics.gov.lk/GenderStatistics/StaticalInformation/Education/LiteracyRateOfPopulation-10YearsAndAbove-ByDistrictAndSex

Ministry of Education. (2023). Assessment on achievement levels of foundational skills in literacy and numeracy of grade 3 students in Sri Lanka. https://noolaham.net/project/1032/103181/103181.pdf

National Education Research and Evaluation Center. (2023, September). National assessment of learning outcomes of the subjects of Mathematics and English of students completing Grade 04 in the year 2022/2023. Ministry of Education. https://edu.cmb.ac.lk/nerec/?p=993

UNESCO. (2020, April 6). Learning through radio and television in the time of COVID-19. UNESCO. https://www.unesco.org/en/articles/learning-through-radio-and-television-time-covid-19

UNICEF. (2023, May 16). 70% of 10-year-olds now in learning poverty, unable to read and understand a simple text. UNICEF Cuba. https://www.unicef.org/cuba/en/press-releases/70-10-year-olds-now-learning-poverty-unable-read-and-understand-simple-text

Bloom, B. S. (1984). The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational Researcher, 13(6), 4–16. https://doi.org/10.3102/0013189X013006004

Henkel, O., Horne-Robinson, H., Kozhakhmetova, N., & Lee, A. (2024). Effective and scalable math support: Evidence on the impact of an AI-tutor on math achievement in Ghana (arXiv preprint arXiv:2402.09809). https://doi.org/10.48550/arXiv.2402.0980

The Hindu. (2023, July 26). AI tool seeks to enhance Gujarati reading skills of students: One step forward towards Indian language LLM. The Hindu. https://www.thehindu.com/education/ai-tool-seeks-to-enhance-gujarati-reading-skills-of-students-one-step-forward-towards-indian-language-llm/article69357873.ece

Times of India (2025, June 16). All govt schools in T to get free tech-enabled coaching. https://timesofindia.indiatimes.com/city/hyderabad/all-govt-schools-in-t-to-get-free-tech-enabled-coaching/articleshow/121868569.cms