From Data to Dignity: Olanrewaju Samuel on Afrocentric AI

In an exclusive interview with Black Business Magazine, Olanrewaju Samuel, a PhD student and computational linguist at Stony Brook University, explores the frontier of Afrocentric AI. Focused on African languages, phonetics, and the intersection of language and computation, Olanrewaju is building inclusive AI models that respect local linguistic structures while addressing systemic gaps in technology.

Interview By SK Uddin

I am Olanrewaju Samuel, a PhD student in the Department of Linguistics and affiliated with the Institute of Advanced Computational Science, Stony Brook University. I am a computational linguist with interest in documentation and computation of understudied African languages. Primarily, I focus on the phonology, phonetics and language-music connection entrenched in computational linguistics of these languages.

In my approach, I prefer to work with African language communities or the typical community-based approach to research. Computationally, I work with both machine learning and subregular models. Experimentally, I am interested in applying recent statistical models to investigate phonetic details of the understudied languages that were not readily available in previous impressionistic studies. Given that I focus on documentation, my work also centers on developing benchmark dataset for African languages. In this regard, I am interested in data governance, responsible AI and data ethics.

You describe linguistic datasets as “knowledge systems,” not just collections of words. In simple terms, what is linguistic data, and why is it so foundational to building any serious AI system for African languages?

A linguistics data is a collection of information that provides generalizations that reflect how and why a group of people say or do what they do and in the way they do it. The data can be text, speech or signs. They are knowledge systems because they show important details of the internal structures of the mind and it makes sense of the world around it. Data are, therefore, not mere words but a reflection of the cultural system of its speakers.

The linguistic data is fundamental to building AI systems because it is the centre of why we build AI systems. AI systems are designed to mimic human intelligence. If the data are knowledge systems, then they serve as the initial representation that should be used, ideally, to train the AI.

Linguistics Island has trained and mentored hundreds of native‑speaker linguists and created the first Africa Computational Linguistics Summer School focused on hands‑on skills. How does this kind of community‑based training change who gets to participate in AI and language technology on the continent?

A community-based training program ensures grassroot developments. This means that the effectiveness of knowledge dissemination takes a stronger root in shared interest and passion. A single person can learn but learning together in a group fosters collaborative efforts for more impact.

Since it is a community effort, representations of underrepresented groups are easier to attain, making it possible to get even people in remote places to work on their languages. The help that community members offer to fellow members affect who gets to participate in the development of AI technologies. On the continent, people can walk their ways up to being one of the most sought-after scholars based on their grassroot learning schemes and engagement.

Therefore, a training that focuses on advancing the skillset of community members will ensure that more people participate in the development of technological products for their languages. It reduces marginalization and promotes collaborative growth.

You’ve emphasized that AI ethics in Africa must involve the communities whose languages and knowledge systems are being modeled. What does ethical, community‑centred AI look like in practice when you’re collecting, annotating, and using African language data?

Community-centred AI looks like having every member of the community involved in the development of technology for their own language. Since they are involved in such development, they get to make stronger and pragmatic decisions based on their specific needs. They will not gamble about what to do for they will be well equipped to solve their own problems.

In practice, collecting and annotating dataset in a community looks like having everyone involved and informed about the benefits and consequences of building data for models in their languages. It means that everyone knows what and how to start developing a dataset for their language. It does not require external efforts for they will grow and work from within themselves.

All community members understand how to use the tools. They taught to think through the system and point out potential pitfalls that may not be available to externals.

When you talk about developing “Afrocentric models,” you argue that AI for Africa should reflect the structure, music, and patterns of African languages rather than copying Western templates. What does it actually mean to design and deploy a truly Afrocentric language model?

There are different linguistic features that are peculiar to African languages. These features should have served as parameters for training Afrocentric language models. However, since most of the approaches to training these models are based on fine-tuning a pretrained model, we risk the idea of having a model that speaks and acts like African languages in a western template. To design and deploy an Afrocentric model, we will need to abstain from fine-tuning, understanding our language nuances and using those to build the models.

Among many features of African languages, noun class, tone system, vowel harmony and verb structure stand out in shaping the contribution of linguistic scholarships. A model that will account for these features will be an Afrocentric one. It will speak and behave exactly like an African. The tones in African languages are particularly different from those that are in other continents where tones are also present, like Asian languages. The peculiarities of the tone systems suggest that our languages are not a combination of two things but that of three parts which are consonants, vowels and prosodic features. A true model for African languages will not just summarize and brush through the internal patterns of the languages but show how it leverages these features for a better understanding.

From winning a shared task award at EMNLP 2025 to building a VAR‑based dataset collection system that reduces the need for physical fieldwork, what recent milestones are you most excited about—and how do you see them shifting Africa from “testbed” to active decision‑maker in the AI era?

I am more excited about the involvement of people within the Linguistics Island community in representing their languages as a benchmark dataset for AI models. The reason is because once everyone can make benchmarks and point out where the models are faulty, a true self-awareness has taken place. They will either change how things are working presently or modify the present system, thereby contributing to and fostering a true Afro-centric AI model. With this, they will be at the forefront of decision making instead of fine-tuning and re-tuning what was never designed to work for their knowledge system.

Disclaimer: The views and opinions expressed in this interview are those of the guest and do not necessarily reflect the official policy or position of Black Business Magazine or its affiliates. The magazine is committed to supporting Black entrepreneurs and fostering conversations that promote inclusion and economic empowerment.