When AI Fails: Building Accountability, Safety, and Trust in African Machine-Learning Systems

Project overview

AI Safety Nigeria (AISN), with the support of SPRITE+, convened a two-day Innovation Forum in Abuja titled “When AI Fails: Building Accountability, Safety, and Trust in African Machine-Learning Systems” on 18–19 August 2026. The hybrid, closed-door workshop brought together researchers, regulators, technologists, and civil-society practitioners to move beyond discussion and directly co-author practical accountability tools for AI governance in Nigeria and across Africa.

Day 1 grounded participants in documented evidence of AI harms, including biased credit-scoring and gender-based bias in fintech systems. It then moved into a facilitated session mapping the data, accountability, and governance opportunities across African AI ecosystems. Day 2 built on this foundation with a technical session on AI control and the case for an African-specific AI harm taxonomy, a presentation on designing proportionate accountability tools, and further case simulation work on credit scoring and debt recovery. It also included case studies on facial recognition deployment and data sovereignty.

Key findings included the implications of the continued absence of a dedicated African AI Safety Institute, and clear opportunities to strengthen implementation of existing data protection provisions and coordination across institutions with complementary mandates. Participants also highlighted the value of citizen-led interventions alongside institutional efforts, and the need for greater representation of African contexts, languages, and populations in AI training data. Participants agreed that mitigation must be prioritised over redress, while recognising that the two strategies must complement each other. They also agreed that gender and data sovereignty must be built into AI systems from the outset, and that effective governance depends on government, industry, researchers, and civil society working together.

The forum’s two lasting outputs are the African AI Accountability and Trust Playbook, now completed and published on the AI Safety Nigeria website as a free, open resource for policymakers and developers, and the Trusted AI for Africa Network, formally launched on Day 2 to sustain collaboration between researchers, regulators, industry, and civil society across the continent.

Objectives

As set out in the original Expression of Interest, the forum’s six objectives were to:

1. Examine documented cases of AI harm, bias, and erosion of trust in African settings across fintech credit scoring, healthcare, public services, and emerging surveillance applications, building a shared empirical foundation for the discussions. Achieved: covered in the Day 1 session on documented harms (credit-scoring bias, the Kano textile manufacturers’ lawsuit, the Lagos Buy-Now-Pay-Later scandal, and the collapse of Babylon Health’s diagnostic platform in Rwanda) and revisited through a facial recognition case study on Day 2.

2. Identify and map structural gaps in Nigeria’s and Africa’s current accountability mechanisms. Achieved: delivered through the Day 1 “Mapping the Gaps in Africa’s Governance” workshop, facilitated by Hanniel Jafaru. The workshop structured the discussion around three connected national themes:

   • The AI Data Gap: does the country have the right data to build, train, and evaluate AI?

   • The AI Accountability Gap: when AI causes harm, who answers for it?

   • The AI Governance Gap: institutions, laws, capacity, and coordination across the lifecycle, alongside a broader infrastructure and skills challenge (compute, connectivity, energy).

   The session drew on current data: Nigeria ranks 70th of 195 countries on Oxford Insights’ 2025 Government AI Readiness Index, and the African Union adopted its Continental AI Strategy in June 2024. Ten accountability areas (AI audit, algorithmic transparency, explainability, human oversight, AI liability, bias testing, impact assessment, redress, incident reporting, and procurement/vendor accountability) and six areas where implementation and coordination can be strengthened (policy, regulatory, institutional, capacity, coordination, and enforcement) were mapped in detail.

3. Build a shared cross-sectoral understanding of ethical, technical, and policy responses, drawing on comparative lessons from the EU AI Act, the NIST AI Risk Management Framework, and the AU Continental AI Strategy. Achieved: delivered through the Day 2 technical session led by Dr Ayuba Tuodolo, which grounded participants in comparative governance models before extending into loss of AI control and the case for an African-specific AI harm taxonomy.

4. Co-design practical accountability tools and governance recommendations proportionate to the African context. Achieved: delivered through Prof. James Agajo’s presentation and workshop on “Designing Proportionate Accountability Tools” (detecting, documenting, explaining, and remedying harm). The session introduced the ALARP principle (“As Low As Reasonably Practicable”, borrowed from aviation and nuclear safety) and a proportionality formula linking accountability intensity to risk, impact, scale, and vulnerability across four risk tiers. Eight tools were presented: AI Accountability Impact Assessment, Data Quality & Bias Audit, Model Documentation & System Cards, Explainability & Decision Traceability, Human Oversight, Audit Trails & Evidence Logs, Incident Reporting & Redress, and Continuous Monitoring. These were worked through a live credit-scoring case study mapping ten risk areas to specific tools, practical solutions, and evidence/KPIs.

5. Forge lasting professional connections across research, government, and industry. Achieved: the room brought together representatives from civil society, industry, academia, and government, and this cross-sector base was formalised through the Trusted AI for Africa Network launched on Day 2.

6. Produce at least two concrete proposals for future applied research or policy initiatives, ready for submission to funders within three months. Achieved: this work was carried out during the Day 2 “From Forum to Funders” session, with proposals identified for progression to international research and policy funders over the coming three months.

Outcomes & outputs

Stakeholders: the forum targeted impact across four stakeholder groups (academia, government/regulation, technology industry, and civil society), on the premise that convening them jointly, rather than consulting them separately, would produce more durable governance outputs. All four groups were represented across both days, matching the sectors identified at proposal stage.

Indicators: (i) approximately 25 in-person participants across sectors, achieved; (ii) two lasting outputs, the Playbook and the Network, both delivered, with the Playbook published on the AISN website and the Network launched at the forum; (iii) at least two fundable proposals within three months, identified during the forum and on track for submission.

Social: discussions on gender bias in fintech credit scoring, and on embedding gender considerations into AI systems and organisational culture from the outset, informed the Playbook. They have the potential to improve financial-inclusion outcomes for women-led SMEs and rural borrowers.

Economic: the forum’s focus on proportionate, resource-appropriate accountability tools, rather than direct transplantation of EU or US frameworks, aims to support innovation and investment in Nigeria’s AI sector while reducing the economic harm documented in cases such as the Lagos Buy-Now-Pay-Later scandal and biased fintech underwriting.

Organisational: the Trusted AI for Africa Network creates a standing structure linking AI Safety Nigeria with NITDA, NDPC, NCAIR, the Nigeria AI Collective, and academic partners. It strengthens implementation and coordination across institutions with complementary AI governance mandates.

Unintended and indirect effects: two themes emerged that were not scoped in the original proposal: the case for an African-specific AI harm taxonomy (raised in the Day 2 technical session), and an emerging joint initiative on African-representative AI training datasets (raised by Hanniel Jafaru and Prof. James Agajo). Both are being carried forward as additional outputs of the forum. The keynote by Andrew Ayobami also surfaced a multilingual safeguards mismatch and untested assumptions in some African governance mechanisms (for example, open-weight models versus API-permission-based approaches) as under-researched areas. It named three sovereignty considerations affecting Africa’s ability to address AI harm: infrastructure dependence, market leverage, and knowledge/information asymmetry. This converges with a design principle echoed by both technical facilitators: accountability tools must reflect local languages, informal economies, and connectivity constraints rather than importing controls that are difficult to implement locally.

Future work

• Dissemination and uptake of the published Playbook among policymakers, regulators, developers, and civil society, with feedback gathered to inform future updates.

• Formalisation of the Trusted AI for Africa Network’s founding charter with its partner institutions, building on its launch at the forum.

• Development of the African-representative AI dataset initiative raised by Hanniel Jafaru and Prof. James Agajo during the forum.

• Submission of the two proposals identified during the forum to international research and policy funders within three months.

• Monitoring uptake of the Playbook and Network activity to inform future impact reporting.