AI for Africa by Africans: How cultural diversity can be attained in AI globalization

Currently, many Large Language Models (LLM) and AI applications have proven to be a reserve for tech giants, with only a few beginners succeeding in this space. 

In terms of language diversity, Africa is yet to be well represented in AI products. 

For instance, OpenAI’s ChatGPT supports 50 languages including English, Spanish, French, German, Chinese, Japanese, Arabic among others. But none of African dialects. 

On the other hand, Google’s Bard supports 40 languages, including 9 Indian languages and Swahili - which is a popular language among Africans. 

In its Interim Report titled Governing AI for Humanity, the United Nation’s AI Advisory Board highlighted the divide that exists in the globalisation of AI, and the need to ensure access and benefits go hand in hand, especially for the Global South region which include Africa.  

“Even outside the Global South, taking advantage of AI will require efforts to develop local AI ecosystems, the ability to train local models on local data, as well as fine-tuning models developed elsewhere to suit local circumstances and purposes,” the Advisory body states. 

Through inclusion and enabling equitable access to technology, the use of current systems will play a crucial role in ensuring language diversity in the AI era. 

AN AI FOR AFRICA

There is one sure way of ensuring diversity in AI, this is through representation in training data. Data is the lifeblood of AI systems, and directly impacts how well AI reflects cultural diversity.

A lack of adequate representation in training data leads to more and more marginalisation of cultures that are underrepresented by a technology that is reshaping the world.

Imagine an AI application that can translate a Swahili video, and recreate an output in Dholuo or Kikuyu. Will this be possible? Yet there is no database of the Kikuyu and Dholuo languages? 

In unpacking the possibility of an Africa-oriented AI system, I had a discussion with Michael Michie, the Executive Director Nia za Akili Zetu Foundation. 

He explains that a significant barrier in language diversification is scenarios where required data does not exist in digitised formats. 

“Collaboration with cultural experts and community representatives is essential. Their insights can help ensure the AI is sensitive to cultural nuances and avoids perpetuating biases,” Michie explains. 

“Fostering diversity within the AI development teams themselves is crucial. This allows for a broader range of perspectives to be considered during the design and development process,” he adds. 

COMMUNITY MANDATE IN CULTURAL DIVERSITY

To enable the availability of data to support African-centred AI, communities hold a crucial role since they are the custodians of their language and cultures. 

“It is important to engage with them while working on more culturally inclusive AI, especially where there is little to no data available to train AI models on,” Michie explains, adding that their involvement, and clearly showing how the communities benefit to ensure they are not exploited. 

In addition to the communities, a diverse human workforce such as data trainers, who are at the centre stage of AI development is also key to realizing this goal. 

THE DRIVING FORCE 

As the lifeline of AI systems, data plays a crucial role in ensuring cultural diversity. Lack of diversity in any training data results to clear bias in AI systems, whose output will clearly indicate underrepresentation of certain communities and cultures.

“This coming to light more proficiently with generative AI tools is a clear indication that without data that is culturally balanced, we risk continuing to marginalize underrepresented cultures. The data used to train AI systems needs to be culturally balanced,” Michie remarks. 

Significant effort should therefore be made in getting and using data that is representative of underrepresented cultures. 

“The desire of the communities to share their culture and even how they would gain from such an exchange as they are the custodians of the culture itself should be utilized.”

THE DIVERSITY RISK 

In a bid to ensure diversity in AI models, tech giants have in recent cases found themselves favouring underrepresented communities, at the expense of ones that are ‘superior’. 

Recently, we witnessed a case whereby Google’s Gemini AI displayed diversity errors, leading to a pause in Gemini’s ability to generate images of people. 

Google’s CEO Sundar Pichai admitted to the errors that “offended our users and shown bias.” This was after the model, in an racial diversity attempt, inaccurately portrayed the races of some popular and historical figures. 

“It is not easy to ensure balanced representation of all races, genders, and cultures with AI models. There is always a group that will be unfortunately still underrepresented despite all the best efforts put in place,” Michie says on such incidences, adding that even though a balance is needed in representation, there is more harm in not being diverse. 

“It is (currently) not possible to build culturally perfect and diverse AI models when we as a species are still struggling with this same issue, but we can build tools that are sensitive to the user's race, gender and more ensuring a form of personalized exchange with the technology that does not undermine culture or beliefs.”

SETBACKS IN AI GLOBALIZATION

The unequal access to technology and infrastructure between nations creates an uneven playing field for harnessing AI opportunities. This discussion is set to take a centre stage during a public debate that will be held at Strathmore University on April 5, 2024. 

Some countries are far ahead of others in the development and use of AI, with other only being consumers of these technologies, due to the lack of infrastructure such as power, internet, compute, and technical resources – creating an AI-divide. 

While issues such as technological access in the Global South are glaring, interconnectedness and limitations to exchange of cultures worldwide also hinder globalization. 

“The belief that one's own culture is superior to others creates a significant barrier to appreciating and respecting diverse practices even with the exposure to cultures that technology has provided to us, there are those who still view their cultures superior and AI through sycophancy could likely make this worse. Stereotypes can also hinder cultural globalization and make it difficult for harmonious exchange and appreciation of cultures and their diversity,” the Nia za Akili Zetu founder says. 

Another barrier to globalisation is scenarios where languages carry cultural nuances that can be easily misinterpreted. 

These nuances if not considered can create a barrier to cultural globalization and acceptance of cultural diversity through languages.

Also, difference governance and data protection regulations can also hinder globalization, as it can make it difficult to share data across borders, hindering the development of truly global AI solutions. 

“While there is need for data protection especially when it comes to sensitive data, limiting collaboration due to laws and regulations could hamper the development of AI into a global presence. Nations can come up with agreements on how to share infrastructure and data while still having their own sovereign AI,” Michie shares. 

AFRICA’S ROLE IN ENSURING REPRESENTATION 

To enable an AI for Africa, organisations such as Nia za Akili Zetu envision a future where every person, irrespective of their technological sophistication, understands and controls the technology directly influencing their lives.

The potential of AI can spur a lot of economic development in Africa, increasing internet and electrical adoption throughout the continent, making data more accessible and usable for training models. A combined effort by Africa could be focused on transforming AI into the next AI frontier this could be achieved through collaboration of African nations to harness Africa’s data potential, workforce potential, infrastructure potential, investment potential. 

This would allow for Africa to quickly move from a net user of AI applications to a prime area for AI research and development globally. 

“Offering cleaner power for compute, more human generated content to train models on, adequate legislation and cyber security for data exchange for training models, researchers and entrepreneurs to use the models to solve for Africa’s challenges and for the world as well,” Michie says. 

Themed AI and Cultural Globalization; the debate that will be held at Strathmore University seeks to delve into fundamental questions surrounding the influence of dominant cultures on local traditions, the ethical dimensions of cultural exchange, and the role of technology in shaping global cultural interactions.