Masakhane: The Story of Building Together

By Sakhile Dube 

PART 1 - How Masakhane began building African-language AI 

As humans, language has shaped how we have moved throughout the world. It shapes how we communicate, learn, access services, share knowledge and pass on stories, history and identity.

As technology increasingly mediates those same parts of life, it also shapes whose knowledge can be found, whose voices can be understood and who can fully participate in the digital world.

Eight years ago, African languages were largely absent from the AI systems that were beginning to reshape that world.

For the Masakhane community, building an inclusive future meant building differently.

A problem too big for two people

In 2018, South African machine learning researchers Jade Abbott and Laura Martinus were working on African language natural language processing when they encountered a problem that was difficult to ignore.

There was very little research on African languages, making up less than 0.5% of all studies.

For many African languages, the only open dataset available was JW300, a digital collection of translated sentences, never designed for today’s multilingual AI systems.  Researchers lacked the data needed to build and evaluate models. African academics working in global institutions faced barriers such as difficulty in accessing large datasets for African language research compared to colleagues working on dominant world languages. Startups had little reliable infrastructure on which to build language technologies, and communities themselves had few opportunities to influence how AI systems represented them.

Abbott and Martinus encountered this gap repeatedly. At one NLP workshop, they found themselves counting perhaps two other African language experts in a room of a few hundred people.

“We realised that this problem is so much bigger than two people could ever handle, and it's going to require a whole village to map the ecosystem,” says Jade Abbott, Masakhane co-founder. 

That realisation came during a train ride after the conference, when the two researchers began thinking about what they had taken on.

This was not simply a technical problem. It resulted from a much longer history in which some languages were treated as more valuable, more useful, and worthy of more investment than others. Commercially, African languages were often seen as too fragmented or too small to justify the cost of building the data and tools needed to support them. 

Kathleen Siminyu, current board chair at Masakhane Research Foundation, noted that the gap helped explain why a community-led approach was necessary.

“People were part of volunteer efforts because they cared about their languages,” says Siminyu. 

We build together 

In 2019, at the Deep Learning Indaba in Nairobi, Masakhane, which means "we build together" in isiZulu, was formally launched.

The name became a model for how the work would be done.

Instead of waiting for large-scale funding or a conventional institution to emerge first, Masakhane started as a distributed research community. Researchers, students, translators and enthusiasts could contribute regardless of whether they had a PhD, an institutional affiliation or an established research profile.

“We came to this work as researchers, focused on the higher end of the pipeline,” Siminyu says. “But we quickly realised the datasets needed did not exist. So, we had to build them ourselves.” 

The community created datasets, published research openly, built transparent benchmarks and shared knowledge. And, importantly, the community made room for people who had traditionally been excluded from deciding what AI research should look like.

What began with a shared notebook and a translation model grew into a Discord community and weekly meetings that still bring people across the continent together. 

Some of the work was slow and very detailed. 

Researchers had to find and assemble fragments of translated text from religious publications, government documents and old radio archives. The community came to call this process data archaeology, digging through material that had existed for years but had never been collected in ways useful for language technology.

Every dataset made a language visible.

Hatem Haddad, Masakhane Research Foundation board member (at the time of the interview), says that the absence of African languages in the digital world reflected a much broader structural inequality: ‘The problem was a severe linguistic digital divide, one that grows more consequential as AI gets adopted into systems shaping access to information, services and opportunities.”

That is why Masakhane’s work goes beyond technical dataset building. It is an effort to repair a digital infrastructure gap, one that determines who gets to participate, access  and benefit from the technologies being built around them.

A welcoming community

Moyahabo Rabothata joined the Masakhane community in 2019, while completing her master's degree, bringing both her skills as a data scientist and her passion for using AI to address challenges in education, health, agriculture, and law.

A newsletter forwarded by her course coordinator and Founding member of Masakhane, South African AI researcher Professor Vukosi Marivate, introduced her to Masakhane. Earlier, a friend, Nomonde Khalo, had introduced her to NLP.

But it was the focus on African languages that made the community particularly appealing.

“I was drawn to Masakhane's focus on African languages, because I strongly believe in ensuring that languages are represented in technology and education,” says Rabothata, 

Rabothata describes herself as a relatively quiet member of the community. Over the years, however, she has followed projects, participated in discussions and attended smaller gatherings, including the Deep Learning Indaba.

Rabothata does not measure her involvement simply by what she has contributed. She also values what she has received: knowledge, exposure and a community that she can continue learning from.

Her message to people who might be intimidated by AI or research is simple.

“You don't need to be an expert in coding or anything to start. Masakhane is open to anyone, as long as you are curious, or you have love for African languages or technology; there's a space for you to learn and grow while giving back,” says Rabothata.

This welcoming spirit is central to the organisation.

Everlyn Asiko Chimoto’s story shows how creating space for people to participate can grow into meaningful work that benefits the wider community. Starting with just 8,000 Luhya-English sentence pairs, she developed more efficient ways to train and improve language models while working within the constraints of low-resource languages. Alongside her research, she has consistently shared her knowledge, helping to organise AfricaNLP, teaching practical AI skills, bringing together KenyaNLP, and contributing to frameworks that give African language communities more say in how their data is used.