South Africa's power system has made a substantial operational recovery, with Eskom reporting a year-to-date Energy Availability Factor of 67.87% in August 2026, its highest level since 2020. The country's experience provides an important lesson in resilience that can be applied to artificial intelligence systems. Rather than focusing on making AI systems perfect, it's crucial to develop mechanisms for recovering from failures and degrading authority when confidence deteriorates.
The global AI debate often assumes that better models should receive more responsibility, with higher confidence enabling greater automation. However, resilient infrastructure is designed with the understanding that components can fail. The next generation of AI may need to focus on how safely authority can be taken away when confidence begins to deteriorate. This approach prioritizes the development of mechanisms for reducing confidence, re-examining assumptions, and restoring normal authority.
South Africa's draft national AI policy was withdrawn after fictitious sources were found in its reference list, highlighting the need for vigilant human oversight. The incident demonstrates that resilience is shown not by hiding errors but by having mechanisms in place to address them. AI systems operating in critical organizations will require similar capabilities to degrade authority and restore human control when confidence deteriorates.
AI systems can make mistakes, and the critical question is what happens to their authority after an organization has reason to doubt them. Many governance frameworks emphasize human oversight, but humans may not be able to effectively override AI if they have become structurally dependent on it. This is why Human Override and Operational Fallback are not the same thing, and organizations need to prioritize the development of Operational Fallback capabilities.
The UAE provides a useful counterpart to South Africa, with its rapid adoption of AI-enabled government and digital infrastructure. The two countries' different experiences point to the same principle: AI authority should be designed to degrade before AI capability disappears. This can be achieved through Authority Degradation, where operational permissions granted to an AI system are reduced as uncertainty about its reliability increases.
Organizations often define what an AI system can do when functioning normally but fail to specify what it can do when reliability becomes uncertain. To address this weakness, organizations should 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.
A control room that relies heavily on AI should occasionally operate with part of that capability unavailable to test its resilience. Such exercises demonstrate not that AI is unreliable but that the organization using AI is resilient. South Africa and the UAE could provide a useful environment for testing AI resilience, with a joint exercise placing an AI-supported infrastructure operation under progressively deteriorating conditions.
Key points
- AI authority should be designed to degrade before AI capability disappears.
- Organizations should identify a Minimum Decision Capability to ensure safe decision-making when advanced digital assistance is unavailable.
- Resilience in AI systems is demonstrated not by pretending errors did not occur but by possessing mechanisms to address them.
