The Responsible AI Lab (RAIL) on September 3, 2025, convened the second lecture in its Responsible AI Series, which brought together students, researchers and faculty. The lecture was delivered by Mr. Sheriff Issaka, CEO of the African Languages Lab and PhD fellow at the University of California, Los Angeles, who spoke on “Building AI Solutions That Work for Africa: A Technical Deep-Dive into Deploying Multilingual AI in Resource-Constrained Environments.”
He began the lecture by situating AI within Africa’s demographic and linguistic realities. With over 1.4 billion people and over 2,000 languages, Africa is diverse and underserved by digital technologies. “Even if these numbers are outdated, 95 per cent of African languages are still not represented in today’s AI systems. Around 88 per cent are critically endangered in digital form”, he said. At the same time, the continent’s youth bulge and growing digital adoption present opportunities. He also reported that recent surveys show that countries such as Kenya, Egypt, Nigeria, and South Africa already exceed the world average in AI usage among internet users. “The potential is there,” he stressed, “but only if we build technologies that work for our contexts.”
A central theme of the lecture was that massive AI models are impractical for Africa. “We cannot compete on trillion-parameter models. Our focus must be on small language models, under 10 billion parameters, designed to run on mobile devices and offline,” he said. He explained that such models were better suited to African users who often access the internet through mobile phones and in areas with limited connectivity. Among the strategies he described were transfer learning, which adapts global models for African languages, and clustering related languages such as Akan dialects to share training efficiencies. He also spoke about data augmentation and synthetic data generation to expand scarce resources, alongside edge computing and offline-first design to ensure AI systems work close to users. Simple but powerful methods such as SMS and USSD could also extend access to AI tools to people with basic feature phones.
He then presented the African Languages Lab’s flagship tool, Mansa MT, a multilingual engine that delivers enterprise-grade translation in 20 African languages, including Hausa, Swahili, Yoruba, Amharic, and Zulu. “Our solutions do better than Google Translate for these languages,” he noted proudly. The system already serves more than 300 million speakers and is projected to support 40 languages by the end of 2025.
He also touched on the ethical dimension of AI and reminded participants that accuracy scores mean little if systems fail in the real world. “You can score 99 per cent on benchmarks, but if your model fails in the field, what’s the point?” He referenced recent tragedies, including a teenager’s suicide linked to chatbot interaction, and raised everyday examples such as facial recognition systems that struggle to identify black faces.
The discussion after his presentation turned to questions of data ownership, open access, collaboration and policymaking. Amina Salifu, a PhD candidate at RAIL, asked about the sources of training data. Sheriff explained that the lab relies on open datasets, partnerships with language service providers, and data collection efforts.
Prof. Jerry John Kponyo, PI and Scientific Director of RAIL, added that questions of ownership must also consider the communities that generate the data. “When we develop solutions, they must benefit the people whose languages we are working with. Otherwise, ownership is hollow.”
Evans Kofi Agyei, a member of the African Languages Lab, reinforced these points by explaining the harsh realities of data collection. “Students will not digitise data for free. Lecturers will not share resources for free. Without funding, progress stalls,” he said. He recounted how the team had trained more than six models for Twi but still found none good enough for deployment. He noted that Ghanaian language data often comes from religious texts, limiting the domains in which models can perform well.
The issue of speech versus text also drew strong reactions. Mr. Abraham Lambon spoke about the difficulty of raising children in a multilingual household where neither parent’s language is being passed on. He asked how AI might help preserve local languages in such settings. Mr. Issaka answered candidly: “For Ghana, the value lies in speech, not text. Written resources are too scarce to build on. But almost every language has a radio station. That’s where the data is.” He stressed that the future of Ghanaian language AI may depend on speech technologies rather than text-based systems.
The debate widened to questions of local versus external ownership. Prof. Kponyo reflected on the risk of external funders shaping the direction of African AI by controlling the resources. “Most of these initiatives are funded from outside the continent. “That creates a problem when they decide to commercialise the results.” Prof. Kponyo noted.
Kwame Acheampong, a participant, added, “We have data in places like the Akan Department. The problem is collaboration. Too often, AI models are built without linguists, which risks misrepresentation.
Participants suggested the creation of an African Dataset Repository, validated by linguists across countries, to allow shared access while maintaining standards.
In his concluding remarks, Mr. Issaka urged participants to see themselves as builders, not merely consumers of AI technologies. Prof. Kponyo closed the session by stressing the importance of strong foundations and collaboration. “It will take partnerships across academia, industry, government, and communities. We must pool our efforts instead of working in isolation and remember that responsible AI is not just about technology but about people. That is how we ensure Africa is not left behind,” he stated.