The concept of Authority Degradation highlights how an AI system’s permissions should be reduced when its reliability is uncertain. This addresses a gap in traditional AI governance, where organisations usually define operational parameters for AI under normal conditions but fail to specify what actions it can take when reliability is questioned. This doubt may stem from various external issues, such as unstable power, poor communications, unreliable data, or staffing shortages. An AI operating in such degraded circumstances is not necessarily trustworthy.
Organisations should establish a Minimum Decision Capability, which is the essential mix of personnel, information, communication, and procedures needed to continue making safe decisions when advanced digital assistance is either unavailable or deemed unreliable. This capability should be devised prior to a crisis and tested regularly. For instance, a control room that typically relies on AI should occasionally function without it to help operators understand which decisions may be delayed, which actions need extra approval, and which operations must be paused. Such exercises highlight an organisation’s resilience rather than the device’s unreliability.
South Africa and the UAE could offer valuable environments for testing, given their differing strengths. A collaborative exercise could involve an AI-supported operational framework facing gradually worsening conditions while the AI remains available at first. Scenarios could include a trusted data source becoming unreliable, increased communications delays, conflicting sensors, or a cyber incident causing network isolation. Eventually, operators might learn that even though the AI is technically operational, its recommendations can’t be considered reliable. The test would assess how swiftly the organisation reduces the AI’s authority, who has the power to do so, whether critical operations persist, and how long a compromised operation can be maintained.
Testing should also include what evidence is necessary to reinstate the AI’s authority once it has been reduced. While organisations typically plan for stopping automated processes, they spend less time on how to rebuild trust. For example, merely restoring a failed data feed shouldn’t automatically regain the AI’s previous permissions. There should be a structured process for authority restoration, encapsulating a full resilience cycle: Normal Authority, Reduced Confidence, Authority Degradation, Human or Alternative Control, Minimum Decision Capability, Validation, and Controlled Authority Restoration. Trust is not an enduring trait assigned to machines; it is a dynamic operational condition requiring continuous evaluation.
The current situation in South Africa offers an institutional analogy, emphasising the need for organisations to be ready to reassess their AI’s authority when new evidence arises. There is no shame in scaling back a system’s power in response to changing circumstances. A more detrimental culture emerges when an organisation has invested heavily in automation, leading operators to hesitate in questioning it.
AI resilience hinges on technology and the organisation’s ability to foster scepticism. The most effective operator is not one who blindly trusts automation nor one who outright dismisses it, but someone who knows when to grant authority, when to restrict it, what evidence warrants such action, and how to ensure operations continue effectively thereafter. This understanding is critical in control rooms, where choices between complete automation and full human control are seldom straightforward.
The infrastructure recovery in South Africa and the UAE’s swift digital integration illustrate two facets of the same challenge. We must strive to develop AI systems that are accurate, explainable, and trustworthy while acknowledging that trust is not a permanent state. As conditions, data, threats, and models evolve, organisations should sometimes opt for a gradual transfer of authority rather than completely shutting down the AI, allowing the broader mission to persist. Ultimately, the test of human control over AI lies not in granting machines authority, but in reclaiming that authority gracefully when trust begins to wane.
Abdulla Saeed Alhebsi is an Emirati author and researcher specialising in security, risk management, and control room operations.
