AI need not inevitably be the domain of Big Tech. Smaller-scale, community-led work, dedicated to solving local problems and empowering marginalized groups, can have real impact. By embedding cultural and linguistic diversity into AI, community-based approaches could enable globally equitable outcomes and help counter the technological monoculture of big business.
If we have learned only one thing from a decade and a half of social media, it might be that technology designed, built and operated in just one place but deployed worldwide should give us pause for thought. Values, assumptions and rules written into the technology we use matter enormously. Global technology frequently just means Western technology.
The development of technology enabled by artificial intelligence (AI) risks following a similar path. Of the many challenges this creates around ensuring AI is developed to the benefit of all, two in particular stand out. The first is the possibility that governance of so-called ‘global’ technology will map poorly to reality in geographies or cultures unfamiliar to its creators and operators. The second is that dominant, Western-developed technologies may squeeze out other technologies built by those very same under-represented geographies and cultures.
In both cases, this would not only hurt local communities in numerous ways (from entrenching cultural biases to limiting the creation of AI solutions relevant to local needs). It would also be a loss for AI development globally, limiting the pool of technical talent for AI work and inhibiting the diversity of perspectives and technical approaches needed to drive innovation.
‘Global technology’ and the linguistic dominance of English
ChatGPT, the OpenAI product that thrust generative language models into headlines and fuelled the widespread use of chatbots, is a good example of what could be considered a ‘global technology’. ChatGPT is a large language model (LLM), a type of AI that is created by trawling huge datasets of text with the aim of generating human-like responses (i.e. resembling examples in the text datasets) when prompted with questions or comments in a conversational manner.
The magic tricks that LLMs can perform in English are frequently astonishing. LLMs can generate sophisticated and (at least superficially) plausible text that is often indistinguishable from that written by a human. But LLMs are far more useful to English speakers than to anyone else. This is due to, on a surface level, issues in the web-crawled data for low-resource languages; and fundamentally, systemic issues in society that are then reflected in the lack of availability of data for these languages, and in the poor quality of the data on the occasions when it is available. Low-resource languages are languages for which insufficient data is available to enable development of robust natural language processing (NLP) capabilities in AI systems. One study found that for at least 15 such poorly represented languages, the data used to train LLMs was totally deficient. In other words, for non-English communities, there is a higher likelihood that AI tools nominally developed for a given language might actually spit out gibberish that is ‘like’ the language in question; this is in addition to the AI having minimal factual knowledge of the local contexts in which the language is spoken.
Other researchers evaluating the performance of LLMs developed for a global audience in relation to that of LLMs developed to serve subsets of African languages found that global AI tools are ‘still not achieving the accuracy of low-resource and Africa-centric language models, [even] on simple tasks…’. For example, an evaluation by Lelapa AI – a South Africa-based AI lab – of ChatGPT’s performance in Zulu found that LLMs built by native-language teams and focused on a subset of languages performed significantly better at named entity recognition than LLMs developed with a global scope did. (Named entity recognition is a crucial step in information extraction, and is used to classify proper nouns in formerly unstructured text.) For machine translation, the gulf was even wider, with ChatGPT 3.5 scoring a round zero as its BLEU score. The team concluded its study by stressing ‘the huge value of context-specific AI work’.
Colonial AI, or local AI?
In principle, universal, global AI, should such a thing be possible, would have none of these problems. Yet the above-mentioned issues with output quality, combined with prevailing economic systems that leave dominant populations as the owners of such tools, suggest a risk that the globalization of AI could facilitate what might be termed a form of ‘colonial AI’, with its insistence on English and little regard paid to local cultures and languages. The near-invisibility, at least until recently (see below), of non-Western and non-English-speaking AI stakeholders in much of the AI debate is also evident in the fact that international AI conferences are typically held in the Global North; attendance of African delegates is usually low, due to distance and high travel costs and registration fees.
There are also concerns that globalized AI could lead to or entrench exploitative economic dynamics, as suggested by reports on the human cost of preparing high-quality training data for building AI models. Data annotation (the practice of human coding and labelling of text, images or videos) is usually outsourced, and often poorly paid. This labour is essential for limiting bias, hate speech, violence and sexual abuse content generated by AI models. However, it has been reported that workers in this field often endure difficult working conditions and are exposed to toxic textual and visual content, which affects their mental health. Even though outsourced data annotation is, in a sense, the backbone of a highly lucrative industry, this is seldom reflected in the status, compensation and protections provided to those performing such roles.
If AI content continues to be developed in such ways, it will have significant negative ramifications for both its producers and its consumers. Failings in the governance of social media have led to violence in parts of the world where companies have not invested in appropriate oversight or care. AI companies need to avoid repeating this mistake: a risk assessment for a model in the UK or US, for example, should not mirror assessments for Bangladesh or Kenya. With AI tools potentially acting as news sources, personal assistants, recruiters or political advisers, it should be a critical priority to ensure each tool is safe and fit for use in a given culture or country.
There is encouraging evidence of a growing African AI community that is taking ownership of AI development and the issues around it, and building and shaping AI technologies that respond to local needs.
Fortunately, the story of AI development in the future is unlikely to be confined to Western companies and cultures. Quite the opposite. For example, there is encouraging evidence of a growing African AI community that is taking ownership of AI development and the issues around it, and building and shaping AI technologies that respond to local needs. Women in Machine Learning and Data Science, an organization that champions opportunities for women and gender minorities in these technical fields, has locally organized chapters in 13 African countries. Data Science Africa, Data Scientists Network (formerly Data Science Nigeria) and the Deep Learning Indaba are grassroots organizations championing capacity development in the African AI community. These groups organize events, summer schools, boot camps and conferences where members of the African AI community nurture the interest of budding young developers and support academic careers.
An increasing number of higher education opportunities in Africa are emerging in AI and machine learning: the African Institute for Mathematical Sciences (AIMS) runs a master’s degree course in machine intelligence; Google DeepMind provides scholarships for students at Stellenbosch University in South Africa and Makerere University in Uganda; and the African Centre for Technology Studies offers AI4D doctoral scholarships to candidates from 21 African countries. Organizations like the Masakhane Research Foundation, GhanaNLP, EthioNLP and HausaNLP are also providing capacity-building, all working to increase NLP research on African languages and with various regional or linguistic focuses.
This flourishing AI ecosystem has had profound effects on the nature of technology in African countries. Language models that reflect local communities are being built. Problems that have previously received little attention from the teams and communities that traditionally develop AI in the West are now being placed front and centre. One example is treatment of Leishmaniasis, a neglected disease most common in Brazil, East Africa and India. Closely associated with poverty, the disease has historically received limited funding for discovery, development and delivery of new treatments. Moreover, the pre-existing treatment was costly, lengthy, painful and sometimes toxic.
In 2021, the ‘Deep Learning Indaba Grand Challenge’ focused on this disease, bringing in local AI practitioners to lead model-building and assist with drug discovery and data analysis. Together, the participants started to tackle a disease that might have been ignored by the mainstream. Over 350 community volunteers participated in the challenge. This led to the selection of several promising drugs that went on to be evaluated by the Drugs for Neglected Diseases initiative – DNDi. The winning solution from the Grand Challenge was published as a conference paper at the International Conference on Learning Representations (ICLR) in Vienna in 2021.
Further evidence that the African AI community’s international profile is increasing – and that practitioners may gain a greater say in the global development of AI – can be seen in the fact that in 2023 the ICLR was held in Africa for the first time. African attendance at the conference in Kigali, Rwanda grew by over 1,000 per cent. While these numbers are a testament to the increase in AI activity on the African continent, the geographic accessibility of the conference certainly played a role.
Open but vigilant
Building local capacity is the foundation of an anti-colonial technology movement, and it is delivering results. But protecting the interests of historically marginalized communities and their technology requires more than simply moving with the times. Steps can be taken to further strengthen the power and autonomy of small technology communities. Licensing the data created and curated by these communities is one example. It would be sadly ironic if such data were simply sucked up by a technology giant and then sold back to the very people who generated it.
The case of the Kaitiakitanga licence offers a positive example of what can be achieved. Te Hiku Media, a collectively owned charitable media organization, started gathering data for the Te Reo language, primarily spoken by the Māori of New Zealand but at risk of extinction following British colonial policies. Te Hiku Media noted the risk that, if such data was left unprotected, foreign technology companies might simply be able to develop products and sell these back to the Māori people. To address this risk, Te Hiku Media developed the Kaitiakitanga licence, designed to ensure that access to the language dataset and any related resources aligns with the customs, protocols and values of the Māori people.
Risks and opportunities
Without addressing colonial power structures recreated in the design and deployment of technology, there is a risk that old inequities will find new life in the technological tools used by local communities, with consequences for such communities’ power and identity. Displacement of local language and culture from our technology threatens communities’ futures, cultural heritage and indigenous knowledge. The threat is a future in which local identities are erased by technological development. In short, there is a risk of dependency on tools and processes built by someone else, for someone else, for purposes that fail to deliver good outcomes for those excluded from their design and value chain.
In contrast, community AI can provide a route to digital self-determination. On local questions, community-based AI is likely to outperform global solutions and to be preferred by local populations; it simply needs to overcome the hurdle of getting the right tools into the hands of the people it is built for. In turn, community AI promises to bring value back to communities as the builders and the owners of these technologies.
