The Gates Foundation will spend at least $1 billion over the next two years to make AI more accessible in health care, education and agriculture, Bill Gates writes in the foundation's 2026 Goalkeepers report. The money goes to tools for teachers, clinicians and farm advisers, and to something less visible than tools: adding non-English sources to the large datasets developers train on. Gates argues the window is narrow. The next 12 to 18 months of decisions about how AI is built, funded and deployed will settle whether the main benefits reach the people who already hold the most resources or the people who hold the least.
The argument rests on language. More than 90% of the data the first large language models trained on was English, which means the people who stand to gain most from the technology are barely represented in what these systems know.
Gates puts the split in plain arithmetic. For roughly 1 billion mostly English-speaking users, AI gets better every month. For the other 7 billion, the change is almost imperceptible.
Speech recognition makes it concrete. Leading systems have an error rate below 6% in English. On Yoruba, a West African language, the same system is wrong more than 60% of the time. In a medical emergency, Gates writes, an inaccurate translation can decide whether a person lives.
He does not expect the market to close that gap. Gates calls it a powerful source of innovation and a poor guarantor of equal opportunity: leave development to market mechanisms alone, and the most effective tools get built first for the people who can pay for them.
The report leans on pilots already running. At Penda Health's clinics in Kenya, an AI assistant helps medical staff make diagnoses and calculate drug doses; diagnostic accuracy rose 16 percentage points. In Sierra Leone, schoolchildren who used Gemini Guided Learning improved over an eight-week pilot by the equivalent of 1.7 years of learning. In India, the farmer advisory service MahaVISTAAR AI already reaches more than 740,000 people and costs the state less than 18 cents per user.
Source: the-decoder.com
The billion is a slice of a much larger plan. The Gates Foundation intends to spend about $200 billion before it closes in 2045 — a sum Bloomberg puts at 99% of Gates's remaining fortune.
That proportion is worth sitting with. Half a percent of the total spend-down is going to the thing Gates says will be decided in the next year to 18 months. Either the urgency is real and the allocation does not match it, or the two-year figure is a first tranche and the report is not saying so. A foundation with a 2045 horizon and a 12-month thesis owes the reader an account of how those fit.
The pilots carry their own quiet detail. The most striking number in the report — 1.7 years of learning in eight weeks — comes from Gemini Guided Learning, which is Google's product. The foundation is not building models. It is buying access to other people's, and paying to improve the inputs those models are trained on. That is a defensible strategy, and it is also a dependency, since the terms on which a frontier lab serves 7 billion non-English speakers are set by the lab, not the grantmaker.
Notably absent from the announcement is what happens to the non-English data once it exists. The report describes adding those sources to large datasets so developers can build systems for other languages. It does not say which developers, under what license, or whether the labs that end up training on Yoruba corpora funded by philanthropy owe anything back to the speakers of Yoruba. The gap Gates identifies is a data gap; the remedy he funds is a data transfer, and the ownership question sits underneath it unaddressed.
Gates recently spent nearly 6,000 words accusing the technology industry of deliberately understating AI's risks because too much money is at stake, naming prolonged job losses, AI-assisted biological weapons and the psychological damage of always-available AI companions. Goalkeepers is the other face of the same position: the same technology as an enormous opportunity, provided the benefits are secured by public policy rather than left to chance. Both halves can be true at once. But the billion will be spent inside an industry Gates has said cannot be trusted to describe its own dangers honestly, and the remedy he names — policy — is the one input a foundation cannot buy.