I've Been Writing About AI for Two Years. I Was Looking at the Wrong Part of the World.

There is an app in Lagos that can tell you whether your malaria medication is real or counterfeit. It costs sixty dollars. It uses a spectrometer the size of a thumb drive and an AI algorithm trained on the molecular fingerprints of four hundred drugs. A pharmacist holds the device against a blister pack, waits twenty seconds, and the screen tells her whether the pills will treat a child’s fever or do nothing while the parasite multiplies.

The company behind it was founded by a Nigerian pharmacist who, at fifteen, swallowed a counterfeit asthma medication and spent twenty-one days in a coma. He survived, studied at Yale, and built a system that now operates in over five thousand pharmacies across Nigeria, Kenya and Uganda. In 2024, it identified and removed 1.3 million counterfeit medications from the supply chain.

Nobody in the Western AI discourse is talking about this.

They are talking about alignment. They are talking about whether GPT* can reason like us. They are talking about watermarking, about detection, about whether students in Melbourne or Michigan are using Claude to write their essays. I have been talking about these things too. I have spent too long building an intellectual framework around cognitive sovereignty and augmentation and the careful stewardship of human thinking in the presence of increasingly capable machines. I have written about beautiful inefficiencies and the skills that compound. These are, I still believe, real concerns. But I am starting to suspect they are concerns that only make sense from inside a very particular kind of institution, in a very particular part of the world, and that the questions consuming the Anglophone AI conversation may be the wrong questions entirely.

The numbers make the case before the argument does. The World Bank’s 2025 Digital Progress Report found that more than 40 per cent of ChatGPT’s global web traffic now comes from middle-income countries, led by Brazil, India, Indonesia and Vietnam. India alone has a hundred million weekly active users. When Datareportal measured adoption as a proportion of internet users, Kenya led the world. Brazil was second. When you strip away the headlines from San Francisco and the opinion columns from London, the actual centre of gravity for generative AI use is not the Anglophone West. It is the global majority.

And what are they doing with it? Not debating consciousness. Not agonising about academic integrity. In Ethiopia, a company has built an AI app builder that lets users describe applications in Amharic, Tigrinya, Swahili or Hausa and receive production-ready code. The whole thing runs on a phone. Earlier this year it partnered with a mobile money operator to bundle AI creation tools into a consumer product, meaning a young entrepreneur in Addis Ababa can build and deploy a working application without a laptop, without broadband, without English. That sentence should land harder than it does. Read it again. Without a laptop, without broadband, without English.

In India, a government AI platform has processed over a billion translation tasks across twenty-two official languages, including real-time speech translation for millions of religious pilgrims. Across sub-Saharan Africa, AI-powered lending platforms use mobile money transaction histories as a proxy for creditworthiness, extending credit to tens of millions of people who have never held a bank account. In Kenya, an offline smartphone app uses computer vision to diagnose crop disease from a photograph. One farmer using it increased yields by 146 per cent in a single season. In several West African countries, an AI agricultural chatbot operating in fifteen languages has fielded over ten million queries from six million smallholder farmers, sixty per cent of them women, and users report a 24 per cent average increase in income.

None of these are pilot programmes. They are scaled systems, already operating, already changing outcomes.

The World Bank has given this phenomenon a name. They call it “Small AI”: affordable, accessible, context-specific. It runs on phones, works offline, uses smaller datasets, and solves problems that the people building frontier models have never had to think about. The distinction carries an implicit accusation. The billions of dollars flowing into frontier capability, into reasoning benchmarks and constitutional AI, are solving problems for populations that already have ‘functioning’ healthcare, trustworthy pharmaceuticals, agricultural extension services and credit markets. The places where AI could do the most measurable good are the places with the least compute, the least data infrastructure, and the least representation in the rooms where AI governance gets decided.

The scale of what is being ignored should trouble anyone who writes about this technology for a living. The UN estimates that counterfeit medicines kill nearly 500,000 people in sub-Saharan Africa every year. Two hundred and sixty-seven thousand of those deaths are from fake antimalarials. In some countries, more than half of all pharmaceuticals on the market are counterfeit. Meanwhile, four billion people speak languages that major AI systems do not serve well. Of roughly seven thousand living languages, about twenty are considered high-resource for AI training. Hindi, spoken by five hundred million people, is classified as low-resource. A 2025 study in Nature found that ChatGPT recognises only 10 to 20 per cent of sentences written in Hausa, a language spoken by ninety-four million people. The linguistic map of AI capability is a colonial map with a fresh coat of paint.

Set those numbers beside the conversation that dominates the AI commentary ecosystem. Set 267,000 annual deaths from fake antimalarials beside the anxiety about whether a second-year student in Sydney used a chatbot to write a paragraph. I am not saying the second concern is illegitimate. I am saying the first concern is a mass casualty event, and our discourse has treated it as someone else’s problem.

I return to the India AI Impact Summit, held in New Delhi in February 2026, the first global AI summit hosted in the Global South. Its framing was a deliberate rebuke. Where the Bletchley and Paris summits organised around safety and risk, India chose the word “impact”. The tagline, drawn from Sanskrit, translates as “welfare for all, happiness for all”. Rumman Chowdhury, CEO of Humane Intelligence, observed that no credible AI summit in the West would have chosen that tagline. She was right, and the reason she was right is the reason this essay needs to exist. Ninety-one countries endorsed the resulting declaration, more than Bletchley, Seoul and Paris combined. Both the United States and China signed. The summit reframed AI governance around diffusion, adoption and development outcomes rather than containment and control.

The Western AI commentary class, myself included, spent that month writing about reasoning traces and system prompts. The declaration that redrew the political map of global AI governance barely registered in our feeds.

I need to be precise about the accusation I am making, because it includes me. The intellectual frameworks I have spent two years building assume a world that already has the things AI threatens to erode. Cognitive sovereignty assumes you had sovereignty to begin with: access to education, to information, to institutions that could be trusted to know things on your behalf. Beautiful inefficiencies assume the inefficient practices existed and did their formative work. Cognitive credit card debt assumes you possessed the cognitive infrastructure to mortgage.

What about the places where none of that was ever true?

A woman in rural Rajasthan who receives an AI-interpreted chest X-ray for tuberculosis has not had her cognitive sovereignty eroded. She has received a diagnosis that no human radiologist was available to give. A farmer in western Kenya whose phone identifies cassava mosaic disease has not suffered an apprenticeship layer collapse. He has accessed knowledge that his country’s extension service could not deliver to his village. The entire vocabulary of loss that structures the Western AI debate, the anxiety about what we are giving up, does not translate into contexts where the baseline was absence. For billions of people, AI is not a threat to existing capability. It is the first capability they have had.

This reframing does not erase the risks. The same AI tools enabling financial inclusion also enable predatory lending. Open-source Chinese models are filling the vacuum in twenty countries where Western platforms are banned or unavailable, and that carries its own geopolitical weight. Microsoft’s own data shows the Global North-South adoption gap is widening: 24.7 per cent versus 14.1 per cent, with high-income countries hosting 77 per cent of data centre capacity and low-income countries hosting less than 0.1 per cent. The concentration is brutal. High-income nations produce 87 per cent of notable AI models despite being 17 per cent of the global population. The Delhi Declaration, for all its political achievement, establishes no enforcement mechanisms and does not confront the corporate concentration of compute.

These are real problems. But they are problems within the frame I am arguing we have been ignoring. Whether AI credit scoring is extractive or liberatory is a debate worth having. It is only possible, though, if you first acknowledge that AI credit scoring is reaching tens of millions of people and that the Western discourse has barely glanced at it.

The World Bank is preparing its World Development Report 2026 under the title “Artificial Intelligence for Development”. Not risk. Not alignment. Not assessment. Development. The question of what AI does for the 83 per cent of the world’s population living outside high-income countries. The question most of us writing in this space have treated as a footnote, when it is the text.

I have written perhaps a two hundred essays about AI. I have built frameworks, coined terms, earned readers. And looking at this research, I find it difficult to avoid a blunt conclusion: I have been writing about the concerns of the global minority and calling them universal.

An essayist writing about human flourishing from an Australian university should probably know where the humans are. Most of them are not here. Most of them are not reading this. And the AI that matters most to their lives is not the kind that writes essays or passes bar exams. It is the kind that fits in a pharmacist’s hand in Lagos and tells her, in twenty seconds, whether the pills will work.

I do not know how to write about that world with the authority it deserves. I know that I have not been writing about it at all. That is the admission this essay exists to make, and I am not going to soften it with a pivot to what we can all learn from each other. Nor should I. Nor should we.