Malawi is entering an important phase of its AI policy debate. COMESA recently brought stakeholders together in Lilongwe for a four-day consultation designed to shape a Regional Artificial Intelligence and Digital Inclusion Strategy, an action plan, and model policies and regulations for member states. That is the right level of ambition. But a strategy for adopting AI also needs an operating standard for deciding when people should trust its output.
The core problem is easy to miss because modern AI is designed to sound useful. A system can summarize a rule, recommend an action, classify a case, or draft a response in seconds. Yet fluency is not the same thing as authority. In public services, the question is not simply whether an answer sounds plausible. It is whether the answer rests on the right source, reflects the latest rule or record, and can be challenged before it creates a meaningful consequence for a citizen.
That makes verification a public-administration capability, not just a technical safeguard. Malawi can build it into AI adoption from the beginning with five practical requirements.
First, every high-impact AI workflow should define an authoritative-source hierarchy. If a system is helping with a public benefit, permit, health-related service, regulatory question, or other consequential process, staff need to know which official records and rules outrank model-generated language. AI can help people find and interpret information, but the underlying authority should remain identifiable.
Second, agencies should set consequence-based approval thresholds. Low-risk uses such as drafting an internal summary can tolerate more automation than a decision that changes eligibility, imposes a requirement, or communicates a binding instruction. The greater the consequence, the clearer the requirement for a person to verify the relevant source and approve the action.
Third, human escalation must be visible to the person receiving the service. A citizen should not have to guess whether an automated response can be reviewed. When the system encounters missing information, conflicting records, unusual circumstances, or a request that falls outside its defined scope, the path to a responsible human decision-maker should be explicit.
Fourth, important AI-assisted actions should leave an audit trail that a person can understand. That does not require logging every internal model calculation. It means preserving the information that matters for accountability: what the system was asked to do, what sources or records supported the result, what recommendation it produced, who approved a consequential action, and what changed afterward.
Fifth, agencies should practice correction before problems occur. If an AI-supported process gives a wrong answer, relies on stale information, or triggers an inappropriate action, staff should know how to reverse the effect, notify the affected person, correct the source of the error, and prevent the same failure from recurring. A recovery process that exists only on paper will often fail when it is first needed.
Malawi already has adjacent building blocks for this approach. UNDP has highlighted the country’s blockchain-enabled Digital ID Wallet initiative, which is aimed at making identity use across public services more secure and interoperable. The Ministry of Education has also validated an AI readiness assessment that emphasizes policy and institutional capacity alongside privacy, human rights, cybersecurity, and infrastructure. Those efforts point toward the same broader lesson: digital capability becomes useful when institutions can govern how it is applied.
A verification standard would also help leaders distinguish between two very different questions. The first is whether an AI tool works in a demonstration. The second is whether an institution can rely on it repeatedly, explain its use, catch mistakes, and recover when something goes wrong. Public trust depends far more on the second question.
This matters as regional AI policy moves from consultation to implementation. The most visible measures of progress will naturally be new systems, new pilots, and faster services. Those indicators are useful, but they can hide the operational burden of correcting poor recommendations or tracing decisions after the fact. Malawi should therefore measure verification performance as well: how often high-impact AI recommendations are checked, how quickly questionable cases reach a human, how long correction takes, and whether repeated errors decline.
Done well, this would not slow adoption. It would make adoption safer to expand. Staff could experiment with useful tools inside clear boundaries. Citizens would have a visible route when automation fails them. Leaders would gain evidence about which uses deserve broader authority and which need redesign.
The COMESA consultation gives Malawi an opportunity to shape regional AI rules while implementation choices are still flexible. A simple verification standard would turn broad commitments to responsible AI into an everyday administrative discipline. Before public-service AI scales, people should be able to answer five questions: What is the authoritative source? How consequential is the action? When must a human step in? What record will be preserved? How will an error be corrected?
If those answers are built into the workflow, Malawi can gain more from AI without asking citizens or public servants to trust a machine merely because it sounds confident.
—
Gleb Tsipursky, PhD, a behavioral scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026). https://disasteravoidanceexperts.com/aibook
