Why AI can’t transform classrooms until it learns local languages

In classrooms across Ghana, students bend over smartphones, eyes fixed on a screen as their virtual math tutor, Rori, guides them through lessons on fractions and probabilities via WhatsApp. For many, it’s the first time technology has been used to support their learning, offering a glimpse of how artificial intelligence could reshape education in low- and middle-income countries. Similar pilots are emerging across the global south, from AI tutors and chatbots to early-grade reading apps and automated assessment tools. They promise not only personalized instruction for learners, but also support for teachers — from planning lessons, setting tests, and marking homework. For education systems where one teacher may be responsible for 80 children or more, the potential is obvious. The question is whether AI can deliver it. Experts caution that AI’s potential is constrained by structural barriers — overburdened teachers, limited funding, unreliable connectivity, and shifting data-protection rules. Language is a critical but often overlooked challenge. Most AI systems cannot understand the languages millions of children speak. Africa alone is home to roughly a third of the world’s languages — between 1,500 and 3,000, according to the United Nations. However, most classrooms teach in English or French as opposed to the children’s mother tongue, which is a severe barrier to learning. Less than 40% of children in many low- and middle-income countries are taught in their mother tongue; in some regions, the figure rises to 90%. This linguistic mismatch is emerging as a central limiting factor for AI in education in low- and middle-income countries, where learning levels are at their lowest. “Children learn best in their mother tongue, yet most LLMs (large language models) struggle with low-resource languages, and particularly with deciphering children’s voices,” Ben Piper, head of global education at the Gates Foundation, told Devex. Han Sheng Chia, director of the AI for global development initiative and a policy fellow at the Center for Global Development, or CGD, called this “underrepresented language problem” a stark “market failure that must be addressed.” “AI is definitely not going to solve the education crisis, but there are some very specific barriers, which if overcome, could make this different from traditional ed tech paradigms,” he told Devex. This linguistic mismatch is not just academic. If children cannot read or understand instructions — in any language — they cannot meaningfully use AI-based learning tools at all, locking them out of the very innovations designed to support them. Paul Atherton, head of Fab Inc., a nonprofit focused on building AI tools for education in low-resource settings through its flagship initiative, AI-for-Education.org, emphasizes that AI must go hand in hand with literacy. “If we can unlock that AI-enabled personal learning experience, then the benefits are clear,” Atherton said. “But this needs kids to be able to read. If you can’t read, you’re locked out of chatbots.” From promise to practice: Scaling AI in low-resource schools David Evans, senior fellow at CGD, described AI in LMIC classrooms as “still in pilot stage,” and warned that measurable learning gains may not be visible for “another decade.” He also cautioned that AI could worsen inequality. “Historically, ed tech tends to expand inequalities, and there’s no reason to believe AI will be any different,” he said. Again, language is a major factor here — it is students with access to devices, connectivity, and English-language content who will benefit first, while those in rural or linguistically diverse communities risk being left further behind, Evans told Devex. Rori illustrates this gap. Developed by Rising Academies, a for-profit that runs low-fee private schools and supports government schools across Africa, the software is currently delivered only in English. While it would be “ideal” if Rori could be translated into multiple languages, there is little political appetite to adapt AI tutors into local languages at the moment, Shabnam Aggarwal, chief technology officer at Rising Academies, told Devex. “Language-of-instruction policies are shaped by complex historical and political factors, so we align our tools to each country’s chosen medium of instruction, which is currently English,” Aggarwal said. As a result, the tool is only used by older students who are fluent enough to communicate with an English-speaking virtual academic tutor. Rori’s English-only design highlights a broader, systemic problem: most AI tools in education are built for languages with large digital footprints, leaving millions of children in low- and middle-income countries behind. Without extensive, labeled datasets — recordings and transcriptions capturing diverse children, dialects, and contexts — AI cannot reliably recognize, transcribe, or analyze speech, nor communicate effectively in the languages that learners actually speak. However, some tech innovators and donors are beginning to take on the challenge. The Gates Foundation has made it a priority, dedicating millions to partnerships with organizations including Pratham, Wadhwani AI, Fab Inc., and IDinsight, to develop freely available digital tools, such as voice recordings and simple speech-recognition models, that can operate offline and in local languages, so teachers can assess and support every student in the language they understand best. In Uganda, a team of engineers at Crane AI Labs is developing local-language apps and models, including a version of Google’s open-source model, Gemma, that is fine-tuned for Swahili and Luganda. Supported by Fab Inc., the lab is also creating an open-source finetuned model and a phone app to help primary school teachers improve students’ foundational literacy. This work is grounded in a robust data strategy, including the creation of new pedagogical datasets and translated benchmarks for local languages. Still in development, the app will initially focus on Luganda, and later other African languages such as Runyankore, Swahili, and Kinyarwanda, allowing teachers to generate resources such as phonics exercises, decodable stories, and comprehension questions. Critically, it is designed to function entirely offline, making it usable in areas with poor or no connectivity. Crane AI Labs’s co-founder, Kato Steven Mubiru, said the lab’s mission is to ensure the AI revolution does not leave African youth behind. “The African continent has a lot of youth, and many don’t have resources. The only thing you can give them is tools — technology tools — to solve their own problems and create economic value,” he told Devex. On the importance of developing AI in local languages, Mubiru added: “A lot of people face literacy challenges, and that won’t change soon. The education system is very weak. If you want to engage people in villages and connect them to technology, it has to be through local languages.” Incentives and scale Despite interest in addressing language barriers, progress is uneven. Atherton highlighted market realities: “Teaching children to read in rural DRC [Democratic Republic of the Congo] isn’t going to make anyone a billionaire, and the prizes or incentives aren’t big enough for companies to invest heavily.” Even when tools exist, scaling is difficult. Many AI solutions remain in software form and rarely reach classrooms. “Getting them into classrooms is limited by a lack of devices, power, and funder support for standalone interventions,” he said. Small pilots, often costing around $50,000, cannot deliver systemic change. “Software is too expensive to build for that model to really work, and governments have no money,” he added. Mixed results Another challenge is the lack of rigorous evaluations of AI in LMIC classrooms — and the few that exist mostly examine English-language tools, with mixed results. An eight-month study in Ghana, for example, found that students using Rori scored significantly higher — an effect size of 0.36 — after two short sessions per week. This suggests that students who used Rori made significantly faster learning gains than their peers. In addition, a World Bank evaluation in Nigeria testing GPT-4-powered tools in secondary English classes found student gains of 0.23 standard deviations. However, a Wharton School experiment with high-school students in Turkey found that generative AI harmed learning when students became dependent on it. Once AI access was removed, students who had relied on the tools performed worse than those who had never used them. To address this, Fab Inc. has launched an evidence facility with $3.9 million from the Gates Foundation to conduct “desk-based” testing and school-term field trials. Atherton said the aim is faster, actionable insights on what works. Piper from the Gates Foundation added: “We have launched investments such as Benchmarking and Rapid Evaluations, which allow us to lab-test and field-test a wide range of teacher- and student-facing AI tools. “These pipelines generate comparable evidence on pedagogical quality, contextual relevance, and accuracy, giving frontier labs and applied EdTech developers clear visibility into where their systems fall short and what is required for LMIC classroom use,” he told Devex. Teacher adoption: A persistent bottleneck Teacher uptake is another challenge. A study from Sierra Leone showed teachers barely used an AI chatbot designed to support lesson planning and assessment. Many teachers are overworked, undersupervised, and lack incentives to integrate new tools, Evans explained. And so the best way to get teachers on board is to design tools that reduce their workload. “By removing the worst parts of the job, you leave teachers with more energy and space for quality interactions with students,” he said. Rising Academies’ CTO agrees that the biggest potential for AI in classrooms is in supporting teachers, not students. “Investing in student-facing tools has always been the default because it’s easier to measure direct learning outcomes. But there’s so much we can do with teachers and administrators,” Aggarwal told Devex. This is partly because most teachers already have smartphones, which they are often willing to upgrade themselves, cutting out the need for repeated investment in hardware that quickly goes out of date. Rising Academies has piloted several teacher-focused tools, including Tari, a WhatsApp teaching support tool; Learn Lens, a grading assistant with roughly 90% accuracy, according to Aggarwal; and BambiaPlus, a structured pedagogy app for foundational literacy and numeracy. Yet, Aggarwal emphasized that human oversight remains essential. “There’s huge potential to support teachers. AI can provide real-time feedback, professional development, and classroom support — but without a human in the loop, trust and accuracy fall apart.” Piper agreed that teachers need to be at the center of design for AI tools to be effective. But that is currently not the case. He noted that a review of nearly 1,000 pitch decks found that just 3% of AI ed tech ideas put teachers at the center of product design. “We see multiple use cases for AI in education to support teachers, including ones that make the job of teachers easier, support them in their core daily tasks, and streamline the complex tasks of lesson planning, continuous assessment, and providing teachers coaching. In LMICs, the teacher is essential, and we see AI as an opportunity to help them,” Piper told Devex in an email. Beyond classrooms, Evans sees untapped potential for AI in education management information systems, or EMIS. They help ministries identify which schools need support, optimize teacher deployment, or plan new school construction. However, data compliance is a looming barrier. Most AI tools rely on servers in the U.S. or Europe, requiring legal “export” of student data. In offline communities, consent often means staff traveling house to house to collect signatures and submit proof to regulators, Atherton explained. What comes next Atherton sees the next frontier as AI “agents” — models that perform end-to-end tasks autonomously. For example, marking homework, generating lesson plans, and emailing results to teachers. Prototypes exist in areas such as coding and evidence summarization, but they remain experimental — and they face the same foundational barrier: Without robust training in local languages, they will not serve the learners who need them most. Aggarwal stressed the importance of more language data for equity. “If frontier labs want to build more equitable models, they’ll need structured data from global south classrooms. We should be exploring incentive models that make that collaboration sustainable and fair,” she said. “AI can only be an equalizer if it’s actually built for everyone.”