OpenAI Foundation and Gates Foundation back five-year AI language access push for 3.4bn people

A coalition of 60 organizations wants people who speak languages underrepresented in today’s AI models to be able to use AI in their own language and voice

A coalition of 60 organizations has set a five-year goal to expand AI access for people who speak languages currently underrepresented in AI models

The OpenAI Foundation and Gates Foundation have joined a 60-organization coalition setting a five-year goal to improve AI access for an estimated 3.4 billion people whose languages are currently underrepresented in AI systems.

The group brings together AI companies, researchers, governments, philanthropies and community organizations, with initial signatories including Anthropic, Google, Microsoft, NVIDIA, UNICEF, the World Bank Group and the UK Foreign, Commonwealth and Development Office.

Its focus is not simply translation. The coalition says many of the world’s roughly 7,000 languages lack sufficient data, tools and benchmarks to support strong AI performance, affecting how accurately systems handle dialect, slang, idioms and cultural context.

Voice is a particular priority. Speech can provide a more practical way to interact with AI where typing or text interfaces are less accessible, with the organizations identifying potential applications across education, health, agriculture, financial services and public services.

OpenAI Foundation puts voice AI at the center

The OpenAI Foundation says improving performance in low-resource languages will become a significant focus of its work, beginning with the data and infrastructure needed for voice AI.

Its efforts will concentrate in part on languages that are unlikely to attract enough commercial investment on their own.

Anna Makanju, Head of AI for Civil Society and Philanthropy at The OpenAI Foundation, said on LinkedIn that her team is now “investing in the data and infrastructure needed for reliable voice AI - the most accessible way to use these tools for many.”

She argued that language performance will determine whether more capable AI reaches communities currently underserved by technology.

“It is a remarkable quality of today’s most powerful AI models that they learn from, and allow interaction in, human language,” she wrote. “But none of it can happen unless the models work reliably in the languages those communities speak.”

Makanju has previously worked on multilingual pretraining data, tokenization and improving GPT-4 performance in languages underrepresented in AI training data.

The OpenAI Foundation says capable voice systems could support use cases ranging from patients describing symptoms in their own language to farmers asking spoken questions and receiving information relevant to their crops.

Coalition sets four areas for joint work

The wider group has outlined four areas where participating organizations will contribute resources and expertise.

The first is building shared language data infrastructure under open licenses. The coalition also plans to develop benchmarks for tracking progress, turn language data into models and applications that can be used by a wider range of AI builders, and establish responsible practices around privacy, consent and data sovereignty.

The commitment brings together organizations working at different parts of the AI stack, from dataset development and research through to model building, implementation and public policy.

One example of the access problem comes from BRAC, the Bangladesh-based NGO. Its CEO, Asif Saleh, says limitations in Bangla AI performance prevented the organization from building an AI scribe for health workers.

“For the communities we serve, voice is the only practical way into this technology, so getting it right in languages like ours will decide whether AI is genuinely useful to the 145 million people we reach, or makes very little difference to them,” he says.

The coalition’s detailed governance structure and individual workstreams have not yet been finalized. Participating organizations plan to develop them collaboratively over the coming year.