South Africa offers an important lesson in resilience that has little to do with artificial intelligence at first glance. After years in which electricity shortages became part of national life, the country’s power system has recorded a substantial operational recovery. In August 2026, Eskom reported that its year-to-date Energy Availability Factor had reached 67.87%, its highest level since 2020, while unplanned outages had fallen substantially. Earlier in the year, Eskom completed a full year without load shedding for the first time since 2018. The lesson is not that infrastructure can become perfect, but that complex systems can recover when reliability, operational discipline and intervention become part of how the system is managed. Artificial intelligence now needs the same philosophy.
Much of the global AI debate is built around an assumption that better models should progressively receive more responsibility. Higher confidence enables greater automation, while successful performance encourages organisations to place AI deeper inside operational processes. Yet resilient infrastructure is never designed on the assumption that its components will never fail. The more important question for the next generation of AI may therefore not be how much authority a system can acquire, but how safely that authority can be taken away when confidence begins to deteriorate.
South Africa recently provided an unusually relevant example of institutional correction. Its draft national AI policy, which emphasised responsible governance, ethical and inclusive AI and human-centred deployment, was withdrawn after fictitious sources were confirmed in its reference list. Communications and Digital Technologies Minister Solly Malatsi said the failure had compromised the integrity and credibility of the document and highlighted the need for vigilant human oversight. There is a principle here that extends beyond the policy itself: resilience is demonstrated not by pretending an error did not occur, but by possessing a mechanism through which confidence can be reduced, assumptions re-examined and normal authority eventually restored.
AI systems operating inside critical organisations will require exactly this capability. Consider an AI platform assisting operators in an electricity network. Under normal conditions, it may forecast demand, identify anomalies, prioritise maintenance and recommend operational adjustments. Then something changes: a trusted data source deteriorates, a cyber incident affects an input, sensors begin to conflict, or unusual conditions fall outside the model’s previous experience. The dangerous question is not whether AI can make a mistake. Every complex system can. The dangerous question is what happens to its authority after the organisation has reason to doubt it.
Many governance frameworks emphasise human oversight, but a human may technically remain “in the loop” while the surrounding organisation has become structurally dependent on AI. If operators no longer practise the task manually, alternative information channels have disappeared, or procedures assume automated analysis is always available, switching off AI may not restore human control. It may simply remove a capability on which humans have learned to depend. This is why Human Override and Operational Fallback are not the same thing. Human Override asks whether a person can stop or overrule the machine; Operational Fallback asks whether the organisation can still perform its essential mission afterwards.
The UAE provides a useful counterpart. Its rapid adoption of AI-enabled government and digital infrastructure makes continuity under declining technological confidence increasingly important. South Africa brings hard-earned experience of restoring critical systems under operational strain, while the UAE brings experience in rapidly integrating advanced digital capabilities into institutions. Their different experiences point toward the same principle: AI authority should be designed to degrade before AI capability disappears.
Imagine an AI system with several levels of operational authority. Under normal conditions it may analyse information, prioritise risks, recommend actions and automatically execute certain low-risk decisions. If data quality deteriorates, its execution authority could disappear while its analytical role remains. If uncertainty rises further, its recommendations could become advisory only. At a still lower confidence level, operators could receive underlying information without AI interpretation, before eventually moving to predefined manual or alternative procedures. The AI has not necessarily failed; its authority has degraded proportionately to the trust that can still reasonably be placed in it.
I would describe this as Authority Degradation: the operational permissions granted to an AI system are reduced as uncertainty about its reliability increases. The concept addresses a weakness in conventional AI governance. Organisations often define what an AI system may do when it is functioning normally, but far fewer define with comparable precision what it may continue doing when its reliability becomes uncertain. That uncertainty may arise not from the model itself but from the environment around it: unstable power, intermittent communications, unavailable data feeds, conflicting sensors, network isolation or reduced staffing. A functioning AI operating on a degraded representation of reality is not necessarily a trustworthy AI.
Organisations should therefore identify a Minimum Decision Capability: the smallest combination of people, information, communications and procedures required to continue making safe essential decisions when advanced digital assistance is unavailable or untrusted. That capability should be designed before a crisis and exercised regularly. A control room that normally relies heavily on AI should occasionally operate with part of that capability unavailable so operators understand which decisions become slower, which actions require additional approval and which operations must be suspended. Such an exercise is not evidence that AI is unreliable; it is evidence that the organisation using AI is resilient.
South Africa and the UAE could provide a useful environment for such testing because their strengths are different. A joint exercise could place an AI-supported infrastructure operation under progressively deteriorating conditions while initially leaving the AI itself available. One trusted data source could become unreliable, communications latency could increase, sensors could conflict, and a cyber incident could force isolation of part of the network. Finally, operators could be told that the AI remains technically operational but its recommendations can no longer be assumed reliable. The test would then examine how quickly the organisation reduces AI authority, who can do so, whether essential operations continue and how long degraded operation can be sustained.
The exercise should also test what evidence is required before authority is returned to AI. Organisations often plan how to stop automated systems but spend much less time defining how trust should be restored. Restoring a failed data feed should not automatically restore the AI’s previous permissions. Authority restoration should be deliberate, producing a complete resilience cycle: Normal Authority, Reduced Confidence, Authority Degradation, Human or Alternative Control, Minimum Decision Capability, Validation and Controlled Authority Restoration. Trust is not a permanent property awarded to a machine after an initial test; it is an operational condition that can rise, fall and require reassessment.
South Africa’s policy episode provides a broader institutional analogy. The significant point is not only that a serious problem occurred, but that the problem triggered withdrawal and reconsideration. Advanced organisations will need the same willingness with operational AI. There should be no embarrassment in reducing the authority of a system when evidence changes. The more dangerous culture is one in which an organisation has invested so much confidence, money and prestige in automation that operators hesitate to question it.
AI resilience therefore depends partly on technology and partly on organisational permission to doubt. The strongest operator is neither the person who blindly trusts automation nor the person who instinctively rejects it. It is the person who understands when the system deserves authority, when that authority should be restricted, what evidence justifies the restriction and how the mission continues afterwards. This is especially important in control rooms, where operations rarely offer a clean choice between perfect automation and perfect human control.
South Africa’s infrastructure recovery and the UAE’s rapid digital integration illuminate different sides of the same challenge. We should continue building AI systems that are more accurate, explainable and trustworthy, but we should stop designing organisations as though trust, once earned, remains constant forever. Conditions, data, threats and models change. Sometimes the responsible decision is not to switch AI off completely, but to make it surrender authority gracefully while the organisation continues to function. The ultimate test of human control over AI is not whether we can give a machine authority. It is whether, at the moment trust begins to disappear, we can take that authority back without losing the mission with it.
Abdulla Saeed Alhebsi (A.S. Alhebsi) is an Emirati author and researcher specialising in security, risk management and control room operations.