Back to summit speeches

Building AI Around Local Languages and Knowledge

Henri Eli Monceau

Permanent Observer of La Francophonie (OIF) to the UN in Geneva and Vienna; Representative for Digital Governance and AI.

So thank you very much for this invitation, for inviting the International Organization of La Francophonie to contribute to this very important discussion. Let me first briefly introduce OIF, because Francophonie sometimes is not known as it really is. La Francophonie is a South-North organization that was founded in 1970 by three African heads of state, the first presidents of the newly independent countries. Today, it comprises 90 member states, associate members, and observers. This includes more than 1,500 languages and cultures.

OIF has been a pioneer in the governance of information and communication technologies, holding its first ministerial meeting on the subject in 1993 and adopting its first action plan in 1995. In particular, it contributed to the establishment of the World Summit on the Information Society, which you know very well. More recently, it has played a key role in ensuring that cultural and linguistic diversity is included among the guiding principles of the Global Digital Compact.

So I serve as the permanent observer to the United Nations in Geneva and Vienna, and as the lead for digital and AI governance for our organization. In this capacity, I oversee the OIF Resource Center for Digital and AI Governance, [unclear]. So our starting point in this discussion should be clear. Artificial intelligence can considerably improve public services. It can help governments analyze complex information, facilitate access to healthcare and education, translate administrative content, support public officials, and make institutions more responsive.

But AI will only serve the public good if public institutions retain the capacity to define what that good means. Today, most states are being invited to adopt systems that they did not design, trained on data they do not control, operating in languages that are not necessarily their own, and reflecting priorities that they did not establish. So the danger is that countries which already depend on external digital infrastructures may develop an even deeper dependency: a cognitive and epistemic dependency.

This is why the debate must go beyond access to technology. We must also ask: Who produces the knowledge embedded in AI systems? Whose languages, experiences, and concepts do these systems reflect? And who retains their capacity to question, correct, and govern them? These questions are particularly important for developing countries and for states that do not have the resources, or perhaps the intention, to compete in the global race to build even larger general-purpose models.

In a recent letter to the UN High Commissioner for Human Rights, I highlighted four interconnected risks. The first is the appropriation and privatization of the knowledge commons. A considerable part of the material used to train large language models comes from knowledge accumulated collectively and often produced, preserved, or funded by public institutions. Society provides the cognitive raw material, while a small number of private actors control the models, access to them, and much of the value they generate.

The second risk is the erosion of human capabilities and shared knowledge. AI systems increasingly do more than provide information. They recommend choices, formulate judgments, and perform intellectual tasks. If public administrations become mere users of external AI outputs without retaining the expertise required to understand and challenge them, the result may be institutional deskilling rather than institutional empowerment. The third risk concerns creativity and intellectual property. Writers, researchers, journalists, artists, and publishers contribute to the vitality of our knowledge and ecosystems.

If their work is absorbed without transparency, consent, or an equitable sharing of value, we may weaken the very cultural and scientific production on which future AI systems depend. The fourth risk is the invisibility of languages, cultures, and knowledge systems. The figures are revealing, and I'm quoting here a very recent study we have ordered. Haitian Creole, for instance, is understood by 95% of Haiti's population, but it appears on fewer than 8% of websites in the country's national domain.

Kirundi is understood by 98% of Burundi's population, yet appears on only around half of 1% of websites in the BI domain. Guarani, spoken by 88% of Paraguay's population, is present on barely a quarter of one percent of websites in the national domain. What is underrepresented online will inevitably be underrepresented in AI. This produces poorer services, greater bias, and ultimately the disappearance of certain communities from the digital representation of the world. Faced with these challenges, our response cannot simply be to help every country consume the same dominant models more efficiently.

I would therefore like to propose that we encourage the development of what I call alternative language models. I use this expression deliberately: these models would be alternative because they would not merely reproduce the architecture, scale, and economic logic of the dominant global models, but they would also be native, rooted in the languages, knowledge, institutions, public-service needs, and data resources of the societies they are intended to serve. An alternative model does not necessarily need to compete with the largest commercial system.

It may be smaller, specialized, and designed for a clearly defined public purpose: supporting teachers in a multilingual education system, helping healthcare workers access locally relevant medical information, translating administrative procedures into national or indigenous languages, assisting agricultural extension services, or enabling public officials to analyze national legislation and public data. Such models could be trained or adapted using carefully curated national and regional datasets. They could combine local resources with open-source technologies and regional computing infrastructure.

Most importantly, their development and governance should involve public institutions, universities, researchers, cultural organizations, and the communities whose languages and knowledge they incorporate. This approach would offer several advantages. It would allow governments to define the problem that AI should address, rather than adapting their priorities to the products already available. It would reduce dependence on opaque systems and enable greater scrutiny of training data, performance, and bias. It would support linguistic diversity and improve the quality of public services for citizens who do not interact primarily in dominant international languages.

It would also help preserve and develop local scientific, administrative, and cultural knowledge while strengthening national and regional skills. And finally, it would create more proportionate pathways into AI. Countries would not need to enter a financially and environmentally unsustainable race for computational scale. Cooperation and specialization could become credible alternatives to imitation. The goal is neither technological isolation nor the fragmentation of the digital world.

It is the capacity to cooperate without becoming dependent, and to participate in global innovation without surrendering the ability to think, decide, and create for oneself. The central question is not whether developing countries will use artificial intelligence. They already do, and increasingly will. The real question is whether they will remain users of intelligence designed elsewhere or become full participants in the production and governance of the knowledge systems that will shape their future. Thank you.