AI also needs to speak the language of those it still cannot listen to
The OpenAI Foundation and the Gates Foundation backed a five-year initiative to expand access to artificial intelligence in underrepresented languages, reaching approximately 3.4 billion people speaking around 7,000 languages that are poorly covered by these systems today, according to EdTech Innovation Hub's report on September 24 in a note by Emma Thompson.
The note does not specify the amount committed by each foundation.
The scope of the initiative is not framed solely as an access issue to a tool, but as an infrastructure challenge.
Anna Makanju, head of artificial intelligence for civil society and philanthropy at the OpenAI Foundation, described the priority in those terms: investing in the data and infrastructure necessary to develop a reliable voice artificial intelligence, considering that this is the most accessible way for many people to use these tools.
Voice as a gateway
This choice between voice and text is one of the design decisions that deserves to be highlighted.
A voice interface can eliminate two relevant barriers: literacy in written language and the need for a keyboard that supports the corresponding alphabet.
Precisely for this reason, the technological decision has an inclusion dimension: bringing AI to populations that are currently outside of the systems does not only mean translating an existing interface, but designing the interaction around the real conditions of its users.
The problem starts before the model
The precedent that changes the reading is how the current models were built.
The systems learned mainly from text available on the internet, a corpus in which a handful of languages concentrates the most published material.
The result was not necessarily an explicit decision to exclude, but a consequence of the unequal availability of data.
The distinction matters because it also defines where the solution lies.
The problem is not simply corrected with more computing capacity. It requires collecting and structuring linguistic data that today do not exist, or that have an insufficient digital presence.
In other words, before an AI can respond correctly in a language, someone needs to build the information infrastructure that allows that language to be represented.
Latin America has a concrete gap
For Latin America, the initiative addresses a documented gap: the region has hundreds of indigenous languages with active speakers and a minimal digital presence.
In addition, public policy for bilingual intercultural education operates with materials that are difficult to convert into trainable data formats.
Therefore, a financially supported effort to collect, digitize, and structure that material has clear beneficiaries in countries like Mexico, Peru, Bolivia, and Guatemala.
The point is especially relevant for education. If a significant portion of AI content and tools continues to depend on languages with greater digital representation, the promise of educational customization and access will necessarily have unequal reach.
The important data is not just how many people it will reach
From the perspective of next+, it is advisable to hold two simultaneous readings.
The first is about access: expanding linguistic coverage is a prerequisite for the educational promises of artificial intelligence to reach those who need them most.
Additionally, five years represents a relevant horizon for data and infrastructure work that requires time and, according to available information, is currently not being sufficiently funded by other actors.
The second reading is one of governance.
Whoever collects and structures the data also largely defines the conditions of its subsequent use.
And here there is a question that the announcement does not resolve: under what license will the resulting corpora remain?
For an educational authority in Latin America, the answer is not an administrative detail.
It is a question that is advisable to ask before incorporating their own materials into such an initiative and not afterwards, particularly when it comes to educational content, community knowledge, and linguistic heritage.
The inclusion of new languages in artificial intelligence does not solely depend on the models being capable of processing them. It depends on who builds the data, how they are structured, who can use them, and under what conditions.