The Role of Open-Access AI in Facilitating Global Inclusion in AI Safety Research

The concept of open-access AI has emerged as a practical solution to empower global majority contexts in the field of AI safety research. By enabling the sharing of AI models through various means, including staged releases, cloud-based access, API interfaces, and widely available weights, open-access AI promotes a more inclusive approach to defining acceptable model behavior.

Research indicates that open-access AI fosters greater inclusivity by allowing model scrutiny and modifications by external evaluators. This process aids in advancing safety research through safety fine-tuning, making AI safer and more reliable.

Open-Access AI Gains Momentum in Africa

In Sub-Saharan Africa, the open-source community is rapidly expanding. Countries like Rwanda and Nigeria have seen developer populations increase by more than 45% from 2022 to 2023. However, the growth of this community, including AI safety researchers, faces hurdles related to the limitations of open-access AI in Africa.

Dependency Dynamics

Open-access AI thrives when driven by external incentives rather than voluntary decisions. Yet, AI developers control which components to share and who can access them. African researcher Sienka Dounia has experienced skepticism when requesting model access, reflecting the asymmetry in this relationship.

Leading AI countries are increasingly adopting nationalist policies to secure AI supremacy, potentially alienating developing nations like those in Africa. Additionally, concerns about national security make governments wary of sharing proprietary AI models.

The U.S. Commerce Department’s Bureau of Industry and Security (BIS) introduced a global licensing requirement for model weights of models trained using more than 1,026 FLOPs, aiming to protect U.S. national security interests. This rule, however, excludes African countries from the “low-risk” list of key U.S. allies, potentially limiting future access to U.S. model weights.

China’s significant involvement in AI development in Africa adds another layer of complexity. Chinese companies like Alibaba and Huawei are expanding their tech presence on the continent, which may impact Africa’s ability to collaborate with Western partners.

African AI safety developers rely heavily on open-access AI, leading to power imbalances. This dependency leaves African developers at the mercy of model producers and the countries they operate in, undermining African autonomy and contribution to global AI safety.

The Boundaries of Open-Access AI in Africa

While open-access AI helps reduce development costs, accessing it remains challenging in African contexts. Critical AI infrastructure such as GPUs and cloud computing services are scarce. A study of Zindi Africa data scientists showed that only 1% had on-premises GPU access, while 4% could afford cloud access worth $1,000 per month, equivalent to around two hours of usage per day for an older Nvidia A100 GPU.

The cost of GPUs relative to GDP per capita in African countries is disproportionately high. For example, an Nvidia A100 GPU costs 22% of GDP per capita in South Africa, 75% in Kenya, and 69% in Senegal. This financial barrier is compounded by limited access to foundational utilities like energy and digital infrastructure. Africa accounts for only 6% of global energy use, with only 51.5% of the Sub-Saharan African population having electricity access in 2022.

Sub-Saharan African countries experience an average of 87 blackouts annually, compared to just one in North America. Internet access varies across the continent, with low-income countries having penetration below 13%. However, 5G adoption is expected to grow significantly, with Sub-Saharan Africa set to have 226 million 5G connections by 2030.

African developers face exclusion from global safety research networks due to financial constraints and visa requirements. Limited AI safety research globally exacerbates this issue, as AI safety accounted for only 2% of the broader AI research landscape between 2017 and 2022. In Africa, AI funding is often directed at developmental challenges like healthcare, agriculture, and education, rather than AI safety.

Despite AI investments in Africa, the prospects for AI safety developers remain limited due to disinterest in AI safety within the region.

Framing AI Safety Issues

Framing theory suggests that issues can be viewed from different perspectives, impacting multiple values. Linking AI safety to people’s values fosters a sense of ownership. In Africa, where AI for inclusive development is highly valued, developers can secure AI funding for developmental challenges by emphasizing the risks AI safety concerns pose to these solutions.

For example, African developers can make the case for collaboration with AI stakeholders offering educational solutions by highlighting the negative effects of misinformation and disinformation on system performance, student learning experiences, and organizational reputation.

African AI Safety Research Collaboration

Establishing safety research networks is essential. The European Network for AI Safety illustrates the power of collaboration. In Africa, consortium-based collaboration has been proposed to build general AI capacity. Initiatives like the Artificial Intelligence for Development (AI4D) program have supported African AI researchers, innovators, and policymakers.

Networks could coordinate safety efforts, such as distributed machine learning, allowing African researchers to scale algorithms, share computational resources, and minimize redundancy. The African Union (AU) could play a key role in fostering collaboration on AI safety by establishing open computing access and promoting shared research initiatives across member states.

Developing Context-Specific African AI Safety

African safety researchers can identify unmet needs in model evaluation and develop niche expertise to enhance their chances of gaining model access. Building on research networks, they could establish organizations with standardized policies for model evaluations to ensure consistency.

Africa’s rich cultural, linguistic, and demographic diversity positions it as ideal for conducting robustness testing. African researchers can design “stress tests” that simulate African scenarios for AI models, improving resilience. Another opportunity lies in testing multilingual models, which are predominantly tested in English.

Hamza Chaudhry warns that AI’s usage in non-English languages poses risks like misinformation and disinformation. Evaluating multilingual models in non-English languages addresses this concern. Establishing diverse teams of African evaluators could incentivize model producers to share access with African AI safety researchers.

Conclusion

While African AI safety researchers can employ small-scale interventions to overcome open-access AI obstacles, comprehensive systemic interventions are essential. These include global AI benefit-sharing commitments and negotiating model access in multilateral AI agreements with leading AI partners.

Given AI’s transformative potential, it is crucial that the power to govern it is distributed fairly. African countries must actively engage in these discussions to ensure their participation in AI safety governance.

Join the conversation and share your thoughts on how we can enhance AI safety globally. Comment below, subscribe to our newsletter for more insights, or share this article on your social media platforms. Together, we can drive progress in AI safety research worldwide.