Giving everyone AI won't make things fair

Researchers gave 640 small-business owners in Kenya free access to GPT-4, one of OpenAI's models, through WhatsApp, and tracked their businesses for five months. On average, the AI advice changed nothing. Underneath the average, it split the group in two. Owners whose businesses were already doing well gained about 15 percent. Owners who were struggling ended up with an 8 percent drop in revenue. Everyone had the same tool at the same price of zero, and the gap between them widened.

The part that stays with me is why. The researchers found no real difference in the number or general quality of the questions the two groups asked. What separated them was which problems they chose to bring to the machine. The stronger owners asked about problems the AI could actually help with. The struggling owners brought their hardest ones, the kind that would stump any adviser, human or otherwise. The study is still a working paper, so hold the exact numbers lightly, but the pattern matches what I watch happen every week in my own work.

Last week I argued that not using AI is a luxury. This is the follow-up, because the obvious fix, giving everyone access, turns out to be the easy half. Access is genuinely being solved: OpenAI launched ChatGPT Go in India at 399 rupees a month, about $4.50, made it free there for a year from November 2025, then took it worldwide at $8 a month in January 2026. The price of the model is quickly becoming the smallest part of the problem. Guidance, knowing what to ask and having someone show you, is the half nobody budgets for.

Handing over a tool has never converted into use on its own. The GSMA, the industry body for mobile network operators, counts about 3.4 billion people who do not use mobile internet, and only around 300 million of them lack coverage. The other 3.1 billion live under a working signal and stay offline anyway, held back by some mix of cost, skills, and having nobody nearby who can show them the point. Cost is still real: the UN's telecoms agency puts a basic 2GB data bundle at 4.2 percent of average income in Africa, against 0.4 percent in high-income countries. Even the quality of access is uneven. A preprint released in late June, one author's early work shared ahead of peer review, measured that African languages need a median of 1.88 times as many tokens as English to say the same thing to GPT-5. Tokens are the units these systems read and bill by, so the same subscription buys less AI in Sesotho than it does in English.

All of it points the same way as the Kenya result: you can put the tool in everyone's hands and still watch the gap grow, because the scarce ingredient is knowing what to do with it. The strongest evidence for closing that gap comes from classrooms.

In Edo State, Nigeria, students spent six weeks in after-school sessions with an AI tutor and teachers there to support them. The World Bank reported gains "equivalent to nearly two years of typical learning in just six weeks", with improvement across the board and girls, who had started behind, appearing to gain the most. In Ghana, an AI maths tutor called Rori, used in two 30-minute weekly sessions with teachers present and phones the schools provided, produced roughly an extra year of learning over eight months, at a marginal cost of about $5 per student. Both are early results rather than settled science. But notice the shape they share: scheduled sessions, a teacher in the room, and a narrow, defined job for the tool. Nobody was handed a chatbot and wished good luck.

So the practical question is who supplies the guidance, and I think it splits three ways.

Companies are furthest ahead on access and furthest behind on this. Buying licences is the easy half of a rollout. For my small-business clients in South Africa, a chat window has become, in effect, the entire marketing department: it drafts the ads, reads the analytics, and writes the landing page copy. Getting each of them there took deliberate setup, sitting down to work out which problems to hand it, building the first prompts around their actual business, and being blunt about the tasks it fumbles. If you run a company, that setup work is a training line in the budget, and it belongs right next to the subscription line.

Schools and governments should copy the boring part of Estonia's plan. Its AI Leap programme gives 20,000 students in grades 10 and 11 free AI learning tools from September 2025, and it trained 3,000 teachers before the students got access, with 38,000 more students joining in 2026. Rwanda is running a bigger bet with Anthropic and the African training organisation ALX: a learning companion called Chidi that guides students with questions rather than answers, deployed across more than 200,000 learners, alongside training for up to 2,000 teachers. Those are launch numbers rather than measured outcomes, so watch that space rather than celebrate it yet. But Ghana's $5 per student per year removes the standard affordability objection. Training the teachers and scheduling the sessions is the expensive part, and even that is nowhere near moonshot spending.

Individuals can lift the lesson straight out of the Kenya study. The owners who gained were the ones who brought the machine problems it could plausibly help with, and worked outward from there. That was my whole arc. I had never written code; AI tools taught me app development and then ad platforms, and the trick the entire way was scope, asking for the next small thing I could verify instead of dumping my hardest unsolved problem on it and scoring it on the miss. If these tools feel useless to you, my honest first question is what you have been asking them.

Access will keep getting cheaper without anyone campaigning for it, because $8 plans and free tiers are the market doing what markets do. Guidance has no market default. It gets built into a deployment on purpose, the way Estonia built it into a school calendar and the way I build it into every client engagement, and each rollout that skips this step will keep reproducing the Kenya result: the same tool for everyone, and the strong pulling further away.