Nutrition has no app

Ask a room of engineers which cause for children they would fund and most of them say education, and most of them mean something with a screen in it. I understand the instinct because I had it. Education looks like an information problem, and information problems are the ones we know how to solve: content, delivery, access, scale. Nutrition looks like a logistics problem with no software in it, so it does not come up. That is not a judgment about which matters more. It is a bias about which is legible to people who build software, and it is worth naming, because I fund both and the second one took me longer to see.

The evidence points the other way

India's school meal programme, the Midday Meal Scheme, now PM POSHAN, is one of the largest feeding programmes in the world, and because states adopted it at different times, economists could measure what it did to learning. Chakraborty and Jayaraman's study, first circulated as IZA discussion paper 10086, found that children with up to five years of exposure to the programme in primary school improved their test scores by roughly 10 to 20 percent. That is a learning result from a feeding intervention, and it is of the same order as results people quote for well-designed teaching interventions.

The reason is not mysterious. A hungry child does not learn, whatever is on the screen, and stunting in early childhood has effects on cognition that no later curriculum repairs. The National Family Health Survey tracks the scale of that problem. Stunting among children under five was 48 percent in NFHS-3 in 2005 to 2006, 38.4 percent in NFHS-4 in 2015 to 2016, and 35.5 percent in NFHS-5 in 2019 to 2021, as reported in the NFHS-5 national report, which carries the earlier rounds for comparison. The trend is downward and the level is still more than a third of all children.

Put the two findings together and the picture is plain. The bottleneck for a large share of Indian children is material before it is informational. Food, transport to school, fees, a place to sleep. An app does not move any of those, and a child who has all of them learns well from a blackboard.

The software-shaped hole

I think of the bias as a software-shaped hole in how technologists choose causes. We look at a problem, and if there is a hole in it the shape of software, we see the problem. If there is not, we look past it. Education has the hole: content can be digitised, delivery can be scaled, progress can be measured. Nutrition does not: food has to be bought, cooked, transported and served, by people, every day, and the only software in it is a spreadsheet.

The check I use on myself is two questions. First: is the bottleneck for this child a material input or an information input. Second: if it is material, is anyone in my circle funding it. Technologists reliably get the first question wrong, because we answer it by looking for the hole rather than by looking at the child. And when we get it right, the second question usually has an uncomfortable answer, because the people we know fund what we fund.

Why the instinct exists

The bias has a respectable-sounding defence, and it is worth taking apart. The defence is comparative advantage: fund what you understand, because you can judge it, improve it, and tell when it is failing. An engineer can read a learning app's retention curve and cannot read a kitchen's supply chain, so the engineer should fund the app.

The defence confuses two roles. Expertise matters when you are building the thing. It matters much less when you are paying for it, because money is the most fungible input there is, and the organisations that run feeding programmes have their own experts. What a donor owes a cause is not understanding of its internals but honest judgment about whether it is the bottleneck, and that judgment is easier for nutrition than for software, not harder, because the evidence is older and the outcomes are measured in centimetres and test scores rather than in engagement.

The other half of the instinct is about feedback. Software gives a donor something to look at: a dashboard, a number of users, a screenshot from a classroom. A meal gives a donor a receipt. People fund what reports back to them, and the things that report back are the things built by people like us. That is a fact about donors, not about children, and the correction is to demand the same reporting from the material programmes: how many children fed, on how many days, against how many enrolled. Good feeding programmes have those numbers. The donor has to ask.

Where the bias is right

I do not want to overstate this, because the bias is sometimes correct and knowing when is the whole point. There are children whose material needs are met and whose bottleneck really is information: a good teacher, a book in their language, a device on which to practise, an adult who can explain a concept the school did not. For those children the software-shaped intervention is the right one, and it scales in a way a kitchen does not. Digital access is one of the four things my own giving is organised around, alongside education, nutrition and care, and I do not apologise for it.

The bias is wrong when it is applied first, before the first question has been asked. The order matters. Ask whether the child is fed, housed and in school before asking what they should learn, and fund in that order, because the later rungs do not hold without the earlier ones. That is not a discovery. Every teacher in a government school knows it. It is a correction to how people like me choose, and it needs restating because our instincts run the other way.

What this changed in practice

Two things, both small and both permanent. The first is an order. My own giving is organised around education, nutrition and care, digital access, and opportunity, and for a long time I treated those as four parallel lines. Now they are a sequence: a child who is fed can go to school, a child in school can use a device, a child with a device and a school has a path somewhere. Funding further down the sequence before the earlier rungs are covered is funding a monument.

The second is a test. I no longer evaluate a cause by whether I can imagine building something for it. That test flatters my skills and says nothing about the child. And when I do fund something with software in it, I ask what material input it assumes: the device assumes electricity and a data plan, the app assumes a child who has eaten, the tutoring programme assumes transport home after dark. If the assumption is not funded, the software is a monument to it.

None of this is an argument against technology in education. It is an argument about order and about honesty. The feeding programme raised test scores by a tenth to a fifth without a line of code, and the children it reached were the ones an app would have missed entirely, because you cannot download lunch. Fund the lunch. Then build the app, if there is still a hole for it.