Executive Summary

• Problem: 70% of 10-year-olds in Low-and-Middle Income Countries can’t read. In India, foundational literacy remains critically low, especially in Hindi-speaking regions.
• Intervention: An AI-powered app based on a proven pedagogy, which achieved +0.4 SD learning gains (equivalent to ~2 years of business as usual schooling) in 45-day RCTs involving thousands of government school students, using paper books. The app will guide 15-minute daily home sessions via parents' smartphones.
• Cost-effectiveness: Estimated $5 per person annual income doubling (Charity Entrepreneurship’s bar: $30 or less), with an overall benefit-cost ratio of ~200:1.
• Ask: We’re seeking feedback on our cost-effectiveness model and scaling plan, as well as introductions to contacts at Google, for potential integration with their ReadAlong app.
   

🔗 Links to project materials (these are open access, so please leave comments):

• Detailed report (2 hour read)
• Pitch deck (20 minute read)
• Cost-effectiveness spreadsheet
• Literature review of Ed-Tech in developing countries
• Demo video of app prototype (2 mins)
   

The Case for AI teaching Foundational Hindi Literacy

• Why Literacy? Literacy is the bottom rung on the educational ladder - you need to learn to read before you can read to learn. While the world has steadily improved primary school enrollment rates, hundreds of millions of children attend school without learning key foundational skills.
• Why Hindi? Hindi is the 3rd most spoken language in the world (after English and Mandarin), yet India’s literacy rate (~76%) suggests that Hindi is likely to have the largest number of illiterate native speakers worldwide. In India, over half of Grade 5 students can’t read a Grade 2 text. Hindi’s phonetic nature makes it well-suited to phonics-based pedagogy—and thus to AI-driven instruction.
• Why AI in education? AI bridges two crucial gaps: personalization and scalability. Personalized tutoring is one of the most effective known interventions in education—but often unaffordable due to labour costs. AI tutoring cuts costs by orders of magnitude while preserving key elements of human interaction (voice, feedback, adaptivity). Two studies in majority-world contexts—Rori (Ghana) and Copilot (Nigeria)—showed 0.3 and 0.36 SD learning gains respectively using AI-based instruction. These are significant improvements: the equivalent of 1.5-2 years of business as usual schooling in the developing world. And importantly, EdTech performs best when it builds on existing effective pedagogies—which is why we are using ALfA.
• Why the Accelerating Learning for All (ALfA) program? I work for an Indian educational not-for-profit, which has developed a picture-to-phoneme pedagogy that has proven highly effective in RCTs across North India. My organisation is well connected with state governments and has agreements for ALfA implementation in 35,000 schools in Uttar Pradesh, Himachal Pradesh and Jammu. ALfA’s question-based interactive pedagogy is well-suited for adaptation to an AI app.

The ALfA program in English. Learners are asked three key questions (left) to build from picture to sound to word. Note that this is easier in Hindi, a phonetic language (each letter makes a single sound, unlike English)

• Why home-based education? Once developed, the app could be fruitfully used at home or at school. However, the vast majority of classrooms in India do not have tablets or other devices for children to use. Using existing smartphones at home would avert the need for significant upfront capital to purchase devices. Further, reforming the Indian education system is notoriously difficult - a theory of change that requires transforming teaching habits in the classroom is less likely to succeed.

Importance Tractability Neglectedness Framework

Importance

• Scale: Globally, 750M people cannot read. UNESCO estimates that achieving universal literacy could lift 170M people out of poverty.
• Positive Spillovers: Foundational literacy has large positive externalities—across employment, health and social equality.

Neglectedness

• EdTech ≠ effective pedagogy. Most education apps in India are exam-focused, pedagogically ineffective, or only accessible for paid subscribers.
• Foundational skills are underserved. One of the best existing literacy apps, Google’s ReadAlong, assumes children already know how to decode letters and blend to form words. It doesn’t teach letter-sound correspondence or how to join sounds together in a structured and explicit way—the key step for beginners.
• To our knowledge, there is no free AI tutor for teaching foundational Hindi literacy.

Tractability

• ALfA has a track record. ALfA's paper-based program in govoernment schools has produced average learning gains of 0.39 SD in maths and 0.58 in Hindi across several RCTs covering tens of thousands of students.
• Smartphone access is widespread. Two-thirds of rural primary students had smartphone access in 2021.
• Government appetite is high. The NIPUN Bharat (‘Literate India’) initiative sets a 2026–27 deadline for universal foundational literacy. While India is virtually certain to miss this target, it at least indicates government commitment to the issue.

• A recent World Bank report names three ‘great buys’ in education. This project includes key aspects of all three:

  a) Structured pedagogy ensures content builds progressively.

  b) Differentiated instruction allows each child to move at their own pace.

  c) Mass messages to parents boosts consistency and encourages short daily sessions.

How does the app actually work?

Children receive 15-minute app-based lessons each afternoon/evening, delivered on a parent’s phone, guided by AI. The app uses voice interaction and pictures to teach decoding and blending skills—starting from scratch (demo video). It adapts the difficulty in real time, offering hints and practice until mastery is achieved. AI-generated encouragement and corrective feedback mimic a patient human tutor—without the labor cost.

Scaling Plan

We will soon be seeking philanthropic grants to help complete the app development and start piloting it. In the medium-to-long term, our business plan is to sell the app to private schools, using this revenue to allow us to provide it for free to government school students (who tend to be more disadvantaged).

Another pathway to impact would be to partner with an existing organisation which already has a large reach. For instance, if Google’s ReadAlong App (which already has 40 million downloads) integrated our structured phonics modules in their existing architecture, this could solve both our distribution and funding problems.

Cost-Effectiveness Model

The following table shows the key parameters of our cost-effectiveness model. Please see our spreadsheet for more details (note that figures in the table are rounded to one significant figure, whereas the spreadsheet sometime uses more precise figures)

Expected benefits:         Product(A:G)         = USD 1.6 billion
Expected costs:               H+I                            = USD 8 million
Benefit:Cost Ratio                                             ~ 200:1

In simpler terms (think of it as the four 4s!)

• App cost per user                   $4
• Learning gain                         0.4 SD
• Lifetime income increase   4%
• Working life                             40 years

When applying a discount rate for benefits occurring in the future, this translates to 0.8 income doublings per app user - or a cost of about $5 per income doubling.

I warmly welcome critiques of this analysis  - please let me know what I am missing!

My biggest doubts

While excited by the potential of this project, I have many doubts too - both practical and philosophical. And I acknowledge that several of my epistemic superiors are quite cautious about the prospect of an AI EdTech revolution.

• Can educational apps compete with social media and video games? Jared Hovrath notes that some advanced countries, like Sweden, are now reducing the amount of tech in classes, because they find students are easily distracted. In my own observations of children in slums in India, I see many using smartphones - but primarily for the easy dopamine hits of video games and social media. Could the app be engaging enough to compete in this aggressive marketplace for attention?
• Would a broadly AI-powered education system destroy social skills? Jonathan Haidt brings this concern home powerfully - the amount of time children are spending on ‘social media’ is making them unsociable and antisocial, and ‘AI friends’ could trigger further isolation and inability to relate with humans. Derek Muller argues that there have been many promised techno-educational revolutions - from radio to TV to MOOCs - but ultimately the traditional format of a classroom of peers guided by a (human) mentor remains powerful, because joint quests are more motivating than solo pursuits.

My hope for our app is to ameliorate these concerns by:

• Keeping schools as highly social spaces, with minimal tech for students
• A degree of parental supervision / teacher pressure for students to use the app
• App gamification, so that it’s a fun experience
• Technical fixes, like the app blocking notifications from other apps and incoming calls, so that students can remain focused

Next Steps and Asks

• Build and test: We are still working on an English prototype, the next step will be converting to Hindi. We aim to launch Hindi pilots in late 2025.
• Explore integration with Google ReadAlong: If we can position our app as a “starter module” that feeds learners into ReadAlong’s storybooks, we could scale our impact. If you have contacts at Google (especially the ReadAlong team), we’d love introductions.
• Refine the model: We welcome all feedback on our cost-effectiveness estimates, scaling plans, implementation assumptions, and doubts about the project. Let us know what we’re missing!

Acknowledgements

A huge thanks to my colleagues David Nelson and Edward Brazier, who are developing the app and also made massive contributions to the literature review and theory of change. We’re grateful to educators and early testers for their insights.