216. HUMAN OVERSIGHT REQUIREMENTS IN AI-OPERATED UTILITIES

1. Introduction

Artificial intelligence is increasingly capable of supporting electricity utilities through demand forecasting, predictive maintenance, grid balancing, outage detection, fraud monitoring, automated dispatch and customer-management systems. As utilities become more autonomous, the central legal question is not merely whether AI can make technically accurate decisions, but whether human officials retain meaningful control, accountability and review authority.

Human oversight therefore requires utilities to ensure that AI systems remain subject to supervision, intervention, verification, audit and correction. This is particularly important where an AI decision may affect electricity access, tariffs, disconnection, safety, privacy or essential public services.

2. Legal Basis for Human Oversight

In South Africa, human oversight can be derived from constitutional legality, administrative justice, accountability and privacy principles. Section 33 of the Constitution guarantees everyone the right to administrative action that is lawful, reasonable and procedurally fair, together with a right to written reasons where rights are adversely affected.

The Promotion of Administrative Justice Act 3 of 2000 (PAJA) gives practical effect to these requirements and permits judicial review of unlawful, unreasonable or procedurally unfair administrative decisions.

Accordingly, an electricity regulator or public utility should not simply argue that an adverse decision was produced by an algorithm. Delegating a function to an AI system does not necessarily eliminate the legal responsibility of the human institution exercising the statutory power.

3. Requirements of Effective Human Oversight

A legally meaningful oversight framework should include:

Human authorisation – clearly identify officials responsible for approving consequential AI-supported decisions.

Monitoring – continuously monitor system performance, anomalies and unexpected outputs.

Override powers – authorised personnel should be able to reject, modify or reverse AI recommendations.

Explainability and records – utilities should preserve relevant data, system outputs, human interventions and reasons.

Risk assessment – higher-risk applications require stronger supervision.

Independent auditing – algorithms should periodically be tested for reliability, discrimination, cybersecurity vulnerabilities and systemic error.

Emergency intervention – critical infrastructure must contain mechanisms allowing humans to safely suspend automated operations.

The EU AI Act provides a useful comparative model. Article 14 requires high-risk AI systems to be designed for effective human oversight and specifically contemplates the ability of human operators to understand limitations, detect anomalies, avoid automation bias, disregard outputs and interrupt the system safely.

4. Automated Decisions and Privacy

South Africa's Protection of Personal Information Act 4 of 2013 (POPIA) also has direct relevance. Section 71 generally restricts decisions producing legal consequences or substantially affecting a person where the decision is based solely on automated processing used for profiling, subject to statutory exceptions.

Therefore, an AI-operated utility using customer information for credit assessment, fraud detection, service eligibility or similar consequential decisions should incorporate appropriate human review and procedural safeguards.

5. Case Law

Case 1: Democratic Alliance v President of South Africa

Citation: Democratic Alliance v President of South Africa and Others [2012] ZACC 24.

Facts: The President appointed the National Director of Public Prosecutions after relying on information that did not adequately address significant adverse material concerning the candidate.

Legal Issue: Whether the decision-making process was rational.

Judgment: The Constitutional Court held the appointment irrational because relevant material had been ignored.

Legal Principle/Ratio Decidendi: Public power must be exercised through a rational decision-making process, and ignoring relevant information can undermine the legality of the ultimate decision.

Significance: For AI-operated utilities, human officials cannot blindly accept algorithmic outputs. They must consider relevant information and exercise independent judgment.

Case 2: State v Loomis

Citation: State v. Loomis, 881 N.W.2d 749 (Wis. 2016).

Facts: A court considered a proprietary COMPAS risk assessment during sentencing.

Judgment: The Wisconsin Supreme Court permitted consideration of the algorithm subject to safeguards and cautions concerning its limitations, accuracy and appropriate use.

Legal Principle/Ratio Decidendi: Algorithmic decision-support does not automatically replace human decision-making; users must understand the system's limitations and avoid treating its output as determinative.

Significance: The principle is highly relevant to utility regulation where AI recommendations may affect essential electricity services.

6. Conclusion

Human oversight is a fundamental governance requirement for AI-operated electricity utilities. AI may improve efficiency and reliability, but consequential decisions should remain embedded within legally accountable human institutions. South African constitutional principles, PAJA and POPIA provide important foundations for reviewability, reasons, fairness, rationality and accountability. The emerging international approach, particularly the EU AI Act, further demonstrates the importance of meaningful human intervention, override capability and protection against automation bias. Ultimately, the lawful utility of AI depends not only upon technological performance but upon maintaining human responsibility over the exercise of public and regulatory power.