Most large language models are trained on data that skews heavily toward English. Research on widely used training corpora, including CommonCrawl, has documented that English alone accounts for roughly half of all source text, with fewer than ten languages making up the bulk of the remainder. Africa’s approximately 2,000 languages and the Arab world’s dozens of spoken dialects together represent a marginal fraction of that data, leaving hundreds of millions of users poorly served by AI systems now embedded in banking apps, healthcare platforms, public services and content moderation pipelines.
The practical consequences are concrete. A customer service chatbot in Nairobi that struggles with Sheng (the Swahili-English hybrid widely spoken across the city) fails to retain users and depresses digital adoption. In Northern Nigeria, content moderation tools with no understanding of Hausa idioms routinely miss harmful material circulating on major platforms. These are not edge cases: the populations affected number in the hundreds of millions across the continent and in MENA.
A new generation of language-first AI ventures
African and MENA entrepreneurs have identified this structural gap as both a business opportunity and, for many founders, a matter of equity. Among the most recognized initiatives is Masakhane, a pan-African research collective founded in 2019 that has mobilized hundreds of researchers across the continent to build open-source natural language processing datasets and models for African languages. Its outputs have been adopted by academic institutions and commercial developers in more than 30 countries, according to the organization’s published contributor data.
Lelapa AI, a Johannesburg-based startup, has pushed further into the commercial market. The company develops large language model applications specifically tuned to Southern and East African languages, targeting enterprise clients in financial services and telecommunications. In a 2023 interview with Disrupt Africa, founder Pelonomi Moiloa described the company’s approach as “culturally grounded AI,” assembled not by scraping random internet text but through direct collaboration with language communities.
In North Africa and the Gulf, the challenge takes a different shape. Modern Standard Arabic is widely written but rarely spoken in daily life in the same register across the Arab world: a Moroccan and a Yemeni user share a script but may communicate in mutually unintelligible dialects. The Mohammed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi has made multi-dialectal Arabic modelling a research priority and has extended that work to African Arabic-speaking communities in Sudan, Somalia and Mauritania, creating an analytical bridge between the two regions.
Cross-regional capital opening new pathways
The financial landscape has shifted alongside the technical one. Gulf-based cloud providers and venture funds have started treating African AI startups as strategic portfolio bets. Abu Dhabi’s G42 announced partnerships with African institutions in 2023 and 2024 to extend cloud infrastructure for AI research, reducing the compute barrier that had constrained startups working on data-intensive language model training.
Funding volumes in African AI remain modest by global standards but are accelerating. Partech’s 2023 Africa Tech Venture Capital report identified AI and machine learning as one of the fastest-growing subsectors in African tech investment, with activity concentrated in South Africa, Nigeria, Egypt and Kenya. Magnitt’s MENA data shows comparable growth from 2021 to 2023, with natural language processing and voice recognition consistently appearing among funded product categories.
The cross-regional dynamic is commercially relevant in ways that are not always visible. A startup that builds strong tools for Swahili and Hausa can, with targeted adaptation, address the Swahili-Arabic code-switching common in coastal East Africa, or the Hausa-Arabic bilingualism widespread in Northern Nigeria and Niger. Gulf-based investors experienced with the difficulties of Arabic dialect diversity increasingly recognize that African language AI founders are solving structurally analogous problems with transferable methods.
Language data governance and the long-term stakes
Language AI raises governance questions that extend well beyond any individual product launch. Training a model on community language data requires decisions about who owns that data, who benefits commercially from its use, and how cultural misrepresentations are identified and corrected. Kenya and South Africa have both updated data protection legislation in recent years, drawing on GDPR principles adapted to local contexts. How these frameworks apply to AI training data, particularly for languages spoken across multiple jurisdictions, remains legally unsettled.
Within the Masakhane community and at the African Union’s Digital Transformation Strategy secretariat, advocates have called for policy frameworks that treat language data as cultural heritage rather than purely as a commercial input. Their argument is structural: as AI tools become embedded in credit scoring, public service delivery and content moderation, the languages those tools understand will determine who participates fully in the digital economy.
African and MENA startups building language-first AI are working on questions that extend into data governance and cultural representation. Decisions about which languages AI systems understand will directly affect digital access for hundreds of millions of users, from credit scoring to public service delivery. Whether they secure the capital, regulatory frameworks and institutional partnerships needed to scale remains a question that investors, governments and multilateral bodies across both regions are beginning to address.