The government of Togo is encouraging its citizens to contribute texts, audio recordings, and translations to create artificial intelligence models that reflect the country's rich linguistic diversity, as reported by the AI for Developing Countries Forum.
This initiative is being viewed not merely as a localized experiment but as a template for how developing countries can leverage technology for cultural preservation and economic significance in the era of digital transformation.
Africa boasts approximately 2,000 languages, yet very few are inclusively represented in AI platforms, as highlighted by experts on The Conversation. Failing to integrate these languages into technological advancements carries the risk of alienating vast populations and widening the technological divide.
The challenge lies in AI's limited ability to accommodate a majority of African languages. This shortfall stems from the disparity between the presence of native speakers and the digital representation of their languages, according to African Business. Research published in the Proceedings of Machine Learning Research reveals that across a range of language models, only 41 African languages are represented, with a mere four—Amharic, Swahili, Afrikaans, and Malagasy—consistently included. More than 98% remain unsupported.
Despite a rapid increase in ChatGPT usage in Africa, as noted by The Economist, even the most advanced language models struggle with African languages, often performing 10 to 30 percentage points worse in performance assessments in comparison to English. Difficulties arise from the multitude of dialects and the phenomenon of "code-switching," where speakers intertwine different languages in conversation.
The core issue, as stated by experts, is a lack of data. Most language training for lesser-represented languages relies heavily on Bible texts, while high-quality audio for speech recognition is limited. To address this gap, researchers are actively recording natural speech patterns.
Moreover, the challenges extend beyond just data scarcity. Infrastructure inadequacies present further barriers, such as the absence of essential tools like dictionaries and glossaries. There are also shortcomings in the technology required to support these languages, including keyboard layouts, fonts, spell-checkers, and tokenisers, which all affect how languages are processed and understood. These factors create issues with spelling variations and tonal distinctions within dialects.
The dominance of texts in LLMs mirrors cultural assumptions prevalent in the Global North, emphasizing individualism and competition, as pointed out by University World News. This reinforces a historical pattern of epistemic marginalization experienced by the Global South.
The lack of representation of African languages in AI models has now become a pressing issue for policymakers and technology leaders across the continent, as stated by Semafor. The World Bank indicates that Africa accounts for just 0.6% of global data-center capacity, despite its significant share of the world’s population.
Currently, Africa’s primary connection to the AI industry has been through its mineral resources like cobalt and copper; however, the ultimate opportunity may lie in enhancing productivity and harnessing the full potential of AI. At present, the continent lacks the necessary infrastructure and energy resources to capitalize on this, leaving meaningful growth linked to AI unlikely.
African tech companies are urged to progress beyond being mere suppliers or consumers within the AI landscape. Instead, there is a call for the continent to be a leader in AI development where the most significant benefits can be realized.
Building AI models from scratch is often costly and resource-intensive. By adopting and tailoring existing models, Africa can accelerate its technological advancements. Togo’s initiative demonstrates that the future of AI in Africa may depend less on the scale of the models created and more on the ability to craft technology that serves the greatest number of people effectively.
