The Gates Foundation has brought together 60 organizations for a five-year effort to help 3.4 billion people use AI in their native languages and voices. Partners include OpenAI, Microsoft, Google and Anthropic. The goal puts a practical question at the center of AI access: whether systems that work well in English can serve people who speak other languages, especially in healthcare and education. Bill Gates argues that widening those benefits must go hand in hand with safeguards and oversight beyond the technology industry.
Why language is an access problem
Early AI systems show a sharp gap between the languages that supplied their training data and the languages people need them to understand. Training data is the material used to develop a model; speech recognition converts spoken words into text, and its error rate shows how often a system gets that conversion wrong.
| Early AI measure | Reported figure |
|---|---|
| Training data drawn from English-language sources | Over 90% |
| Speech recognition error rate for English | Under 6% |
| Speech recognition error rate for Yoruba | Over 60% |
Those figures do not predict performance in every language or application. They do show why adding a language to an AI service is not the same as making that service reliable for its speakers. Gates also points to gaps in local context, including the knowledge needed for healthcare, schooling and farming.
In parts of sub-Saharan Africa with very few doctors relative to patients, an AI assistant that understands a local dialect could help someone describe symptoms and consider whether to seek care at a clinic. That potential depends on the assistant understanding the person accurately. Diagnosis and treatment decisions still belong to medical professionals.
▲ Local-language support in rural healthcare
Language access can matter in other settings too, including translation in U.S. courts and emergency medical services. In each case, performance needs to be evaluated for the language and task at hand rather than inferred from English results.
What the coalition plans to build
The coalition initially sought 10 participating organizations and secured 60. Its work centers on gathering labeled multilingual data—language samples paired with information, such as verified transcriptions, that helps systems learn and be assessed. The partners also plan to establish benchmarks, or shared tests, for comparing how well AI models perform across languages.
This approach shifts attention from the number of people who might receive an AI tool to whether that tool can understand them. Curated speech data in regional dialects is particularly important for developers working on global-health uses. Better data and tests are foundations for access, not a guarantee that an AI assistant will be suitable for a clinical or classroom decision.
The foundation’s global-health work offers a reason for its focus on practical use. Gates has previously worked to build support for polio eradication, and he now places healthcare alongside education as a field where AI’s potential depends on institutions adapting. Teachers and university professors, for example, face questions about how generative AI affects academic integrity and intellectual curiosity. Both sectors need to examine what the tools can and cannot do in their own settings.
Safety cannot be left to technology executives
NVIDIA CEO Jensen Huang has argued for moving quickly to support American innovation while committing not to ship unsafe products. Gates’s position is that AI progress need not stop, but some critical areas may require a slower pace while safeguards are put in place. He identifies severe risks such as bioterrorism and employment disruption, as well as difficult regulatory questions in healthcare and education.
Gates argues that lawmakers should use AI in their everyday work so they can understand its capabilities and failures firsthand. He also calls for politicians, researchers, educators and university professors to take part in decisions about its impact. His concern is that a democratic society cannot function well if those decisions are left entirely to technology executives and company leaders.
▲ Shared oversight of AI deployment
The same reasoning extends across borders. Gates sees the United States and China as competitors in AI development, but says both face risks from biological threats and cyberattacks on essential systems. In his view, neither country will emerge as the sole winner, and cooperation on shared defenses creates benefits for both.
What Korea and other countries can take from the goal
For Korea and other countries, the coalition’s target raises a question for governments and companies: how will they test AI in the languages, voices and public services they intend to support? The reported English and Yoruba results should not be treated as a measurement of Korean-language performance. They illustrate why local evaluation matters before an organization relies on a broad claim of access.
The practical next steps are to test tools directly, help develop well-labeled language data where it is needed, and assess performance for specific uses. Educators can examine classroom effects; health institutions can evaluate proposed tools with clinical professionals, who retain responsibility for diagnosis and treatment. The coalition’s five-year goal may widen access, but its value will depend on whether people can use AI accurately and safely in the settings that matter to them.