The Gates Foundation put a billion dollars behind AI's missing users

01

WHAT ACTUALLY CHANGED

The Gates Foundation announced on September 15 that it plans to spend at least $1 billion over two years on AI access and applications. Reuters reports that this sits inside the foundation's previously announced plan to spend $9 billion annually. The new figure is a future spending commitment, not a new endowment, a completed grant total, or evidence that one billion dollars has already reached implementers.

The announced budget gives the promise a rough shape. The Associated Press reports about $400 million for education, $400 million for health, $100 million for agriculture, and $100 million for digital foundations. Examples include teacher tools, clinical support, drug and vaccine research, farmer advice, and data for underrepresented languages. Those categories identify priorities, but the cited materials do not yet provide a complete grant list, country allocation, delivery schedule, procurement record, or outcome framework.

The 2026 Goalkeepers report sets out three conditions for useful access: systems should work in every language people speak, use local data and fit local realities, and arrive with investment in people and affordable access. The report says more than 90 percent of the data used to train early large language models came from English-language sources. It argues that local teams should help decide how data is collected, stored, protected, shared, and evaluated.

The report presents early examples from clinics, classrooms, and farms in Kenya, Sierra Leone, India, and the United States. It describes improved diagnostic accuracy, learning progress, and a growing government farm service. These examples are useful leads, not a universal effectiveness verdict. The report selected them to illustrate possibility, and each result still needs its own study design, baseline, comparison, follow-up period, cost accounting, and independent scrutiny.

Bill Gates also told Reuters that governments are not sufficiently prepared for AI's effects on jobs, security, and personal behavior. That warning is broader than the foundation's spending plan. Philanthropy can fund experiments and public goods, but it cannot create labor policy, clinical liability, education standards, privacy rights, or democratic oversight on behalf of every country.

02

WHY THIS MATTERS

Language is infrastructure, not decoration. A system that performs well in English can mishear a medical phrase, misunderstand a crop name, or miss the social meaning of a request in another language. Building a dataset is only the first step. Local speakers need authority over collection, consent, licensing, sensitive terms, dialect coverage, evaluation, correction, and the conditions under which the data can be reused.

Local fit changes the answer. A recommendation can be medically correct but unavailable in the local clinic, agriculturally sound but unaffordable to the farmer, or educationally polished but mismatched to the curriculum. Tools should be tested against the choices people actually face, the resources they can obtain, and the harms created when a confident answer is wrong.

Frontline workers need leverage, not a software handoff. A nurse, teacher, or agricultural adviser should be able to inspect the source, reject a suggestion, record an error, and escalate a difficult case. Training cannot mean a launch-day webinar followed by permanent dependence on an outside vendor. The budget should build local technical teams, professional judgment, maintenance capacity, and institutions that can stop a failing system.

A billion-dollar headline can hide an evaluation problem. Counting model calls, registered users, translated languages, or distributed devices says little about whether diagnosis improved safely, learning lasted, harvest losses fell, or workloads became manageable. Every project needs a public baseline, a meaningful comparison, disaggregated results, cost per useful outcome, known failures, and evidence after the novelty period ends.

The foundation's partnerships with major model companies can unlock compute and engineering talent, but they can also deepen dependence on firms that set prices, availability, model behavior, and data terms. Public-interest funding should buy portability, documented interfaces, exportable records, durable local access, and a credible exit plan if the provider changes course.

03

WHERE IT COULD HELP

• Publish a grant register with recipient, country, amount, purpose, model and infrastructure providers, data terms, milestones, evaluation design, conflicts of interest, and final results
• Create language evaluations with local speakers across dialect, accent, literacy, code-switching, medical vocabulary, agricultural vocabulary, and high-consequence misunderstandings before a system reaches the public
• Give local institutions decision rights over data collection, consent, storage, licensing, access, correction, deletion, and reuse instead of treating communities as raw material for a distant model
• Measure outcomes that matter to the worker and the person served, including error severity, referral quality, time saved, workload shifted, learning retained, farm income, accessibility, cost, and who is excluded
• Require portability and an exit plan so records, prompts, evaluation sets, documentation, and operational knowledge can move when a vendor raises prices, removes a model, changes its terms, or stops supporting a language