By Chisom Obiudo
We keep using the phrase ‘AI governance’ as though the words themselves confer meaning.
Frameworks multiply. Principles get adopted. Advisory panels convene. Yet beneath this performance of activity lies a question we rarely ask aloud: do we truly understand what we are trying to govern?
The gap between policy ambition and technical comprehension widens with each breakthrough. We draft regulations for systems whose inner workings remain opaque to their own creators. We mandate “transparency” without defining what evidence would demonstrate it.
We demand “fairness” whilst lacking any shared method to measure it. Words like “alignment” and “inclusivity” sound virtuous, but they often conceal power rather than distributing it. This is governance theatre, not governance practice, and the disconnect has real consequences for humans making decisions about health, employment, credit, and safety.
Governance without literacy is meaningless. You can legislate, regulate, audit, and enforce with great vigour. Still, if the people drafting these rules cannot explain how these systems work, if executives approving deployments cannot describe what occurs when models drift from their training data, and if front-line workers using these tools daily cannot recognise when outputs are erroneous, then the entire system becomes mere formality.
Guardrails written without an adequate technical foundation either miss critical risks entirely or create compliance burdens that obstruct beneficial innovation without improving safety.
Effective AI governance must start with understanding, not just rules. It requires technical and social comprehension of how these systems operate, the data that powers them, where their decisions end up in practice, and who bears the consequences when they fail. Only then can we turn principles into practice through clear roles, verifiable evidence, and ongoing learning.
The What: Mapping Your AI Landscape
AI governance starts by knowing what systems exist. Most organisations maintain clearer inventories of their office furniture than of their AI tools. This must change. Develop a record of every model, API, and vendor tool that affects organisational data or influences decisions. For each system, document its purpose, its owner, its data sources, its evaluation methods, its identified risks, its active controls, and its next review date. Treat this inventory as a living document and with the same rigour you apply to financial ledgers. Without knowing what you have, you cannot possibly govern what matters.
The When: Setting Clear Decision Points
AI governance is most effective when it takes place at specific, defined moments rather than being viewed as a vague, ongoing background process. Think of these as checkpoints where someone must actively decide whether a system is ready to proceed, requires improvement, or should be stopped.
Before launching any system, establish the requirements that must be met. Test systematically for fairness across different groups, privacy protection, security vulnerabilities, and potential misuse. A designated AI governance lead must review this evidence and approve the launch.
Whilst systems run in production, monitor their actual performance continuously. Watch for signs that the system’s behaviour is changing from its original design (this is called “drift” and happens when real-world data differs from training data), that accuracy is degrading, or that users are complaining or overriding its recommendations.
When monitoring reveals problems, send automatic alerts to the responsible person for that system, who must investigate and respond within specified timeframes, such as within 48 hours for moderate issues and within 4 hours for severe problems.
After incidents occur, pause to learn instead of simply moving on. Conduct reviews to understand what went wrong and why, then update your controls and share lessons across teams. This transforms failures into opportunities for improvement rather than allowing mistakes to recur.
The How: Translating Values into Evidence
Every organisation claims to value transparency, fairness, and safety. These remain abstractions until they are tested. The “how” dimension involves defining what transparency looks like for a specific system, identifying tests that demonstrate fairness, and establishing thresholds that indicate acceptable safety margins. Make verification part of the standard development process through pre-launch checklists that cannot be bypassed, automated dashboards that display key metrics on every team’s monitor, and simple templates that guide people through the right questions. When doing the right thing becomes the path of least resistance, governance becomes actionable rather than aspirational.
The Who: Giving Voice and Authority
Inclusion involves more than just consultations in the conference room. It means creating regular opportunities for affected communities to influence the AI design itself, not just respond to finished products. This could include monthly review sessions where community members evaluate sample outputs, quarterly workshops where they test prototypes, or ongoing advisory panels that give input on design decisions before they are finalised. Explain systems in clear language. Show how input affects outcomes. At the same time, make accountability tangible by assigning clear authority for approving, pausing, or retiring systems. When everyone knows who can make which decisions, responsibility moves from being theoretical to operational.
Moving Beyond Tokenism
Real inclusion extends beyond expert panels into the communities where AI decisions actually have an impact. The people designing credit algorithms should engage directly with loan applicants. The teams building hiring tools should speak with job seekers. The developers creating healthcare systems should listen carefully to patients. These conversations require structure, clear explanations, and demonstrated impact.
When community members suggest changes and subsequently see those changes implemented, with public acknowledgement of what shifted and why, participation becomes meaningful rather than performative. This is not charity. It is recognising that the people who live with AI decisions possess crucial knowledge that technical experts and data scientists lack. Their insights deserve equal weight in design decisions.
Accountability Requires Literacy
Accountability cannot exist without literacy. People cannot meaningfully question what they cannot comprehend. Build AI fluency at three levels.
Universal literacy enables everyone to understand what AI systems can and cannot do, their fundamental limitations, and how to use them safely.
Operational literacy equips developers and operators with testing, monitoring, and incident response capabilities.
Oversight literacy equips auditors, regulators, and executives to interpret evidence, ask the right questions, and tell substantive transparency from cosmetic disclosure. Spread across all levels, it makes accountability real.
The Mindset That Matters
Effective governance comes from ongoing learning rather than static compliance. AI technology is constantly evolving, with new capabilities and risks emerging faster than policy updates can keep up. Organisations must recognise that today’s AI systems have fundamental limitations that cannot be eliminated through engineering. These systems do not truly understand what they process – a medical AI analysing scans cannot grasp illness the way a doctor does. They absorb and amplify existing patterns in their training data, meaning historical biases in hiring, lending, or policing can be embedded into automated decisions.
These AI systems also produce confident-sounding outputs even when completely wrong, leading to errors that seem authoritative rather than tentative. Recognising these limitations is not pessimism; it is essential for developing systems that are reliably effective in practice. Creating shared practices across institutions, industries, and communities becomes crucial when challenges surpass individual organisational boundaries.
Control and coexistence work together rather than in opposition. Testing, monitoring, and approval processes create confidence that enables broader beneficial deployment. Human oversight, appeal mechanisms, and transparency allow AI systems to augment rather than replace human judgment.
Begin With Understanding
Before setting limits, take the time to understand how systems truly function. Before establishing ethical principles, involve those affected in shaping them from the outset. Before developing compliance processes, base them on evidence of what produces better results.
So, how do we govern what we barely understand? We stop performing governance and start practising it. That means learning before legislating, listening before designing, and measuring before claiming success. The curtain must come down on governance theatre.
*Chisom Obiudo is an admitted legal practitioner of the High Court of Namibia and a Chief Legal Officer at the Namibian Law Reform and Development Commission. She serves as a member of the National Artificial Intelligence Technical Advisory Committee on law and governance. Chisom holds a master’s degree in Corporate Governance and professional certificates in Non-Executive Directorship, AI Governance, and Legislative Drafting.
