We’re seeking feedback on our cost-effectiveness model and scaling plan
The cost-effectiveness you mentioned is incredibly strong, which made me suspicious. "$5 per income doubling" is high, indeed.
I've worked in software for most of my professional life. Going through this more, I'm fairly skeptical of the inputs to your model.
Good web applications are a ton of work, even if they're reusing AI in some way. I have a hard time picturing how much you could really get for $2M, in most settings. (Unless perhaps the founding team is working for a huge pay gap or something, but then this would change the effective cost-effectiveness).
I don't see much discussion of marketing/distribution expenses. I'd expect these to be high.
The AI space is rapidly changing. This model takes advantage of recent developments, but doesn't seem to assume there will be huge changes in the future. If there are, the math changes a lot. I'd expect a mix of [better competition], [the tool quickly becoming obsolete], and [the employment landscape changes so much that the income doubling becomes suspicious].
You mention the results of academic studies, but my impression is that you don't yet have scientific results of people using your specific app. I'd be very skeptical for how much you can generalize the studies. I'd naively expect it would be difficult to motivate users to actually spend much time on the app.
In the startup world, business models in the very early stages of development are treated with tremendous suspicion. I think we have incredibly large uncertainty bounds (with lots more probability of failure), until we see some more serious use.
Overall, this write-up reminds me of a lot of what I hear by early entrepreneurs. I like the enthusiasm, but think that it's a fair bit overoptimistic.
All that said, it's still very possible it's still a good opportunity. Often in the early stages, people would expect a lot of experimentation and change to the specific product.
Thanks Ozzie! I agree that it's a bit hard to take $5 per income doubling at face value. I was tempted to revise some of my figures (eg reduce the 1/4 probability of 'success') to come up with a more 'realistic' figure (though arguably this undermines the very process of cost efficacy analysis if we post-hoc change the inputs to reach an output we want/find plausible). I fully agree that the uncertainty bars are large. Addressing your points in more detail: 1. It's true, David & Edward have both made massive contributions to the coding for free. That said, both have reported that use of AI has consistently helped them achieve goals faster & easier than they had thought. 2. We were budgeting $3 per user for distribution costs (row 41 on first sheet of spreadsheet) which is roughly what my organisation's paper-based intervention currently costs. But it obviously depends a lot on whether word-of-mouth helps it take off, or if we need to continue promoting it widely. 3. Agreed that the AI landscape is highly uncertain; for this reason we are only modelling out 5 years and not assuming any benefits post that period (perhaps we should reduce this to 3 years). More broadly, I think if/when we do reach AGI, it will require a massive rethink of everything - including most global health & development initiatives - and most cost-efficacy analyses will go out the window. Whether returns to education would increase or decrease in a post-AGI world is a fascinating question. 4. Agreed, we have not yet built a Hindi version of the app; once we have this will be a top priority to measure learning gains. The effect sizes are derived from both others' studies of comparable EdTech and my organisation's paper-based ALfA program. As written above, I have similar doubts about whether we can hold the user's attention with the app - will require significant work. Fair call to be skeptical until we get some results. If/when we do manage to build the app and pilot it, will post the results here. Thanks, Tom
I went back-and-forth on this topic with Claude. I was hoping that it would re-derive my points, but getting it to provide decent criticism took a bit more time than I was expecting.
That said, I think with a few prompts (like asking it what it thought of those specific points), it was able to be useful.
Very promising pitch / TOC. Thank you for sharing it.
I would like to seek your opinion on a couple of thoughts.
a) Have you found / done any study which discusses access to digital platforms such as mobile phones, laptops, computers (with parents or govt schools) with illiteracy among peoples of Hindi speaking states?
I believe delivery of your service (along with the development) will be a crucial element of the success of this intervention.
b) Partnering with existing education focused organizations like Labhya, Teach for India, Child Rights and You, Bachpan Bachao Andolan, etc. at the stage of development may make these interventions substantially more effective.
These organizations, grounded in reality of Indian education system, curriculum, and its limitations can help in developing more directly actionable insights. If you do not intend to scale your ground force for delivery and dissemination, then you can consider becoming their tech partner.
I apologize this suggestion if you have already approached them.
Thanks Shubham! a) yes, agreed that delivery will be crucial (and may well be harder than the development). We haven't done a study ourselves, however are basing our work on several other studies:
the Annual Status of Education Report found that 2/3 of rural students had access to a smartphone. I presume (and am backed by my own experience) that this is higher in urban areas.
Our literature review found several studies of EdTech in India, though most of these involved the researchers/school providing the hardware, rather than using parents' phones.
b) My organisation, DEVI Sansthan, has partnered with Teach for India in the past, and we're aware of Pratham, Labhya, LLF, CSF etc's work. We are ourselves quite well grounded in the frustrating intricacies of the Indian education system, through rolling out the paper-based ALfA program in thousands of schools. But it's a good point that if we involve some other organisations in the development of the program, then they could be very helpful in distributing it (in a sense, this is what we are hoping with Google's ReadAlong).
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.
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):
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)
A discount rate of 3% applied to the income benefits accruing over 40 years
G. Discount for possible failure modes
25%
Low
I.e. a ¾ chance of failure due to significant technical, political & economic barriers
H. Fixed & app development costs
US$ 2m
Medium
Discussion with industry experts, Charity Entrepreneurship’s models for starting a new charity
I. Expected running costs before business model kicks in
US$ 6m
Medium
Calculations on AI & server costs, field team stipends during scaling phase
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.
This is the third in a sequence of posts taken from my recent report: Why Did Environmentalism Become Partisan?
Summary
Rising partisanship did not make environmentalism more popular or politically effective. Instead, it saw flat or falling overall public opinion, fewer major legislative achievements, and fluctuating executive actions.
Public Opinion...
This post presents the executive summary from Giving What We Can’s impact evaluation for 2025. At the end of this post we share links to more information, including the full report and...
Happy to see genuine attempts at this area.
The cost-effectiveness you mentioned is incredibly strong, which made me suspicious. "$5 per income doubling" is high, indeed.
I've worked in software for most of my professional life. Going through this more, I'm fairly skeptical of the inputs to your model.
In the startup world, business models in the very early stages of development are treated with tremendous suspicion. I think we have incredibly large uncertainty bounds (with lots more probability of failure), until we see some more serious use.
Overall, this write-up reminds me of a lot of what I hear by early entrepreneurs. I like the enthusiasm, but think that it's a fair bit overoptimistic.
All that said, it's still very possible it's still a good opportunity. Often in the early stages, people would expect a lot of experimentation and change to the specific product.
Thanks Ozzie!
I agree that it's a bit hard to take $5 per income doubling at face value. I was tempted to revise some of my figures (eg reduce the 1/4 probability of 'success') to come up with a more 'realistic' figure (though arguably this undermines the very process of cost efficacy analysis if we post-hoc change the inputs to reach an output we want/find plausible). I fully agree that the uncertainty bars are large.
Addressing your points in more detail:
1. It's true, David & Edward have both made massive contributions to the coding for free. That said, both have reported that use of AI has consistently helped them achieve goals faster & easier than they had thought.
2. We were budgeting $3 per user for distribution costs (row 41 on first sheet of spreadsheet) which is roughly what my organisation's paper-based intervention currently costs. But it obviously depends a lot on whether word-of-mouth helps it take off, or if we need to continue promoting it widely.
3. Agreed that the AI landscape is highly uncertain; for this reason we are only modelling out 5 years and not assuming any benefits post that period (perhaps we should reduce this to 3 years). More broadly, I think if/when we do reach AGI, it will require a massive rethink of everything - including most global health & development initiatives - and most cost-efficacy analyses will go out the window. Whether returns to education would increase or decrease in a post-AGI world is a fascinating question.
4. Agreed, we have not yet built a Hindi version of the app; once we have this will be a top priority to measure learning gains. The effect sizes are derived from both others' studies of comparable EdTech and my organisation's paper-based ALfA program. As written above, I have similar doubts about whether we can hold the user's attention with the app - will require significant work.
Fair call to be skeptical until we get some results. If/when we do manage to build the app and pilot it, will post the results here.
Thanks,
Tom
I went back-and-forth on this topic with Claude. I was hoping that it would re-derive my points, but getting it to provide decent criticism took a bit more time than I was expecting.
That said, I think with a few prompts (like asking it what it thought of those specific points), it was able to be useful.
https://claude.ai/share/00cbbfad-6d97-4ad8-9831-5af231d36912