NEWS / SEP.2026

The Gates Foundation brings together 60 signatories for underrepresented languages in AI

The Gates Foundation announces a commitment by 60 signatories to improve access to AI in underrepresented languages. The five-year goal targets an estimated population of 3.4 billion speakers, with particular attention to voice and control over local data.

Anis AyariSeptember 21, 2026 · 4 min

The Gates Foundation wants to open up AI to 3.4 billion people

The Gates Foundation announced on September 21, 2026, in New York, a collective commitment by 60 signatories to narrow AI’s linguistic divide. The partners have set themselves five years to enable an estimated population of 3.4 billion speakers of underrepresented languages to use AI in their own language and through voice. Google, Anthropic and OpenAI Foundation are among the participants, AP reports.

The estimate of 3.4 billion speakers, cited by the signatories, describes the population targeted by the goal, whose access to AI varies. The announcement sets a collective five-year ambition, without presenting a new service or measured progress on this scale.

Mistral, Senegal’s Ministry of Telecommunications and Digital Affairs and Rwanda AI Scaling Hub also appear on the official list of signatories. The coalition intends to coordinate language-related work that has been underway for years, bringing together organizations that produce resources and those that build tools.

When language hinders use

The Gates Foundation’s press release explains that dialects, expressions and cultural context can change the meaning of a sentence. A word-for-word translation can therefore lose essential information. Voice is also useful when typing or text-based interfaces are impractical.

On September 21, OpenAI Foundation announces its participation and plans to start with the data and infrastructure needed for voice AI. It is particularly targeting languages that risk attracting too little commercial investment to have sufficient resources.

In this text, OpenAI Foundation shares an account from Asif Saleh, executive director of BRAC. He explains that in Bangladesh, an AI-assisted note-taking project for healthcare workers ran up against the limitations of models in Bangla. This account concerns a task involving documentation of healthcare workers’ work.

Shared resources for local tools

The joint commitment provides for responsible collection with consent and the opening up of non-proprietary data. Shared infrastructure under open licenses is intended to make these resources available to tool developers. Their reuse must coexist with privacy protection and data sovereignty.

Earlier funding covers several sectors

The announced breakdown allocates approximately 40% to education, 40% to health, 10% to agriculture and 10% to digital foundations. These proportions describe a planned allocation. The two-year multisector funding program and the five-year language commitment have different scopes and durations.

Regarding this funding program, AP reports criticism from Jonathan Shaffer, a sociologist at the University of Vermont. He argues that additional data does not guarantee more equitable tools when healthcare infrastructure is inadequate. The joint commitment also acknowledges that access depends on connectivity, electricity and affordable devices.

An incorrect transcription or an insufficiently represented dialect can distort a request, and then the advice produced by an assistant. Control over the reuse of voices and local knowledge will also depend on clear consent rules and community participation in governance.

A school, public service or health center could request reproducible public tests in its target languages, dialects and use cases before selecting a tool. Licenses and consent procedures would also make it possible to check who can reuse the data and under what conditions.

The Gates Foundation expects the coalition’s structure, governance and areas of work to be developed collectively over the coming year. This timetable concerns its organization, without setting a firm date for an initial delivery of data.