In 1984, the educational psychologist Benjamin Bloom reported that students tutored one-to-one performed about two standard deviations better than students taught in a conventional class: the average tutored student outperformed 98% of the comparison group 1. He called the challenge of achieving that at scale "the 2 sigma problem". Four decades later, large language models have made a personal tutor for every child technically plausible for the first time. The question is whether it works, and for whom.
The scale of the opportunity in India
India's schools enrol around 24.7 crore students 2. Foundational learning has been recovering since the pandemic, but the gaps remain large: in 2024, 23.4% of Class III children in government schools could read a Class II-level text 3. The National Education Policy 2020 calls universal foundational literacy and numeracy an urgent national mission and encourages the use of technology to support teachers and learners 4.
Teachers carry most of this load, often in large or multi-grade classes. Any technology that helps must make their day easier, not add another dashboard.
What the evidence says
The most relevant evidence so far comes from a World Bank randomised evaluation in Nigeria. Over six weeks, secondary-school students used a GPT-4-based assistant in after-school sessions, guided by teachers, to practise English. The researchers estimate learning gains equivalent to 1.5 to 2 years of typical schooling, with benefits that grew with attendance and showed up in end-of-year exams 5.
Three details from that study matter for India:
- Teachers were in the loop. The AI was used in structured sessions with teacher guidance and curriculum-aligned prompts, not as an unsupervised homework shortcut 5.
- Dosage mattered. Students who attended more sessions gained more, which suggests that design and habit matter as much as the model.
- It was low-cost. The programme used a general-purpose AI assistant rather than bespoke software, which makes it relevant for resource-constrained school systems.
One trial does not settle the science. Results from one country, subject and age group may not transfer, and unsupervised chatbot use can just as easily become a way to skip thinking. That is why we treat this as a research question rather than a product claim.
What an Indian AI tutor would need
From our research so far, we think an AI tutor that works in Indian classrooms needs at least five properties:
- Fluency in Indian languages, including code-mixed speech, and the ability to work through voice for younger learners.
- Alignment with state curricula and textbooks, so that practice reinforces what is taught in class.
- Teacher control over what the tutor covers, with simple reports that save time rather than create work.
- Frugal design that works on entry-level Android phones, shared devices and intermittent connectivity.
- Measured, not marketed results: every claim of learning gain should be tested against a fair comparison group.
Our approach
Dark Basalt's AI EdTech vertical is in the research stage. We are exploring vernacular tutoring for foundational literacy and numeracy, teacher co-pilots for lesson planning and assessment, and lightweight learning diagnostics. We intend to run small, carefully measured pilots with schools and NGOs before building anything at scale.
If you lead a school network, an education NGO or a research group and want to pilot with us, we would like to talk.
Working on these problems? Dark Basalt partners with researchers, institutions and investors.
Partner with us