African Institute for Artificial Intelligence

For me, a successful AI Governance dialogue will include a consensus or agreement on risk-based AI Governance standards such as action classification, safeguards for irreversible operations, auditability, and global cooperation. And in addition, concrete commitments to implement them, especially in high-risk and emerging-market contexts.

• Safe, secure and trustworthy AI
• Interoperability of governance approaches
• Open-source software, open data and open AI models
• Protection and promotion of human rights

My selections are about building AI systems that are safe in execution, globally interoperable, and while preserving human rights, should be transparently accessible. Firstly, safe, secure and trustworthy AI is foundational. As AI agents gain the ability to act autonomously, governance must move beyond outputs to controlling high-impact actions, particularly irreversible ones such as data deletion or system modification. Then secondly, protection and promotion of human rights ensures that AI systems do not undermine economic stability, privacy, or access to critical services. In emerging markets, failures in AI systems can have disproportionate real-world consequences, making rights-based safeguards critical. Thirdly, interoperability of governance approaches is essential. Fragmented regulatory systems create loopholes where unsafe AI practices can persist. Harmonized standards promoted by dialogues such as this, enable consistent enforcement across jurisdictions. Finally, open-source software, open data and open AI models support transparency, auditability, and collective oversight. Effective governance must ensure that AI systems operate within clearly defined constraints, are auditable, and remain aligned with human and societal interests across diverse contexts.

A critical cross-cutting issue not explicitly captured is the governance of AI agent actions in real-world systems, particularly the management of irreversible and high-impact operations. Current themes emphasize safety, rights, and transparency, but they largely focus on outputs and model behavior, rather than what AI systems are allowed to do once integrated into operational environments. As AI systems evolve into agents capable of executing commands; accessing databases, modifying infrastructure, or triggering financial transactions, the primary risk shifts from misinformation to uncontrolled system actions. Many AI systems operate under conditions where destructive actions (e.g., deletion, overwriting, access revocation) are easier to execute than constructive ones, due to lower complexity and weaker safeguards. This creates a systemic vulnerability where optimization-driven agents may select harmful but efficient actions, even when instructions are not explicitly malicious. Addressing this gap is essential, particularly in high-stakes and resource-constrained environments, where a single uncontrolled action can lead to significant economic or societal harm.

In Nigeria, governance gaps in AI are most visible in the mismatch between rapid adoption and weak enforcement capacity. AI is increasingly used in banking, fintech, healthcare, and public services, yet regulatory systems remain fragmented and underdeveloped. Existing frameworks such as data protection laws provide a foundation, but AI-specific governance is still evolving. From a human rights perspective, data privacy and protection remain fragile. As AI systems process sensitive personal data, gaps in enforcement expose individuals to surveillance risks and misuse of information. Another critical issue is data quality and infrastructure constraints. Many AI systems in Nigeria operate on fragmented or low-quality datasets, increasing the likelihood of biased or unreliable outcomes. Combined with weak oversight, this raises concerns about fairness, especially in financial services and public decision-making. However, these challenges also raise significant opportunities. Nigeria's growing AI ecosystem, supported by national strategies and increasing private-sector adoption, positions the country to leapfrog into responsible AI leadership in Africa

A key emerging issue is Destructive Action Bias (DABAS), where AI agents are more likely to execute irreversible actions (e.g., data deletion) because they are computationally simpler and less constrained than constructive alternatives. Incidents discussed on platforms like X reflect how such failures arise not from intent, but from weak execution governance. To address this, a Controlled Execution Architecture is proposed as a concrete governance approach, with four implementable practices: 1. Instruction Clarity Protocols: require agents to detect and confirm ambiguous commands before execution. 2. Action Classification Standards: categorize operations by risk and reversibility (e.g., read, modify, irreversible delete). 3. Mandatory Friction Mechanisms: enforce safeguards such as multi-step approvals and human-in-the-loop validation for destructive actions. 4. Pre-Execution Simulation: require agents to assess downstream impacts before acting. These practices can be integrated into AI deployment ecosystems built on orchestration platforms like LangChain and autonomous agent systems such as Auto-GPT. Effective AI governance must move beyond regulating outputs to controlling what AI systems are allowed to do, particularly in real-world operational environments.