SELF-LEARNING REGULATORY RULE SYSTEMS

1. Concept and Regulatory Significance

Self-learning regulatory rule systems are regulatory frameworks in which artificial intelligence, machine learning, automated analytics, or adaptive algorithms continuously analyse market data and modify regulatory responses. In electricity markets, such systems could detect manipulation, predict congestion, adjust compliance thresholds, identify abnormal bidding, prioritise inspections, or recommend changes to grid and tariff rules.

Unlike conventional regulation, where rules are formally amended by human institutions, a self-learning system may adapt after observing new data. This creates significant legal questions concerning delegation of regulatory power, transparency, explainability, procedural fairness, accountability, bias, judicial review, and the rule of law.

South Africa currently does not have a specialised statutory regime authorising autonomous AI systems to create binding electricity regulations. Indeed, Cabinet announced in June 2026 that the draft national AI policy published earlier that year would be withdrawn and reworked to establish appropriate national standards for ethical AI use. Consequently, any such electricity-regulation system would remain subject to existing constitutional, administrative-law, and sector-specific requirements.

2. Legality and Human Regulatory Authority

A fundamental limitation is the principle of legality. An electricity regulator such as NERSA may exercise only powers conferred upon it by legislation. A machine-learning system cannot independently acquire regulatory jurisdiction merely because its predictions are technically accurate.

If an algorithm changes licence conditions, tariff methodology, compliance thresholds, or enforcement priorities, the regulator must be able to identify the statutory authority permitting that result. Self-learning systems should therefore ordinarily operate as decision-support mechanisms, unless legislation expressly authorises automated decision-making.

Section 33 of the Constitution and the Promotion of Administrative Justice Act 3 of 2000 require administrative action to be lawful, reasonable and procedurally fair and recognise rights to reasons and judicial review. These requirements remain applicable even where computational systems contribute to regulatory decisions.

3. Explainability, Accountability and Dynamic Rules

Adaptive systems create a particular problem because their decision logic can evolve after deployment. A regulator must therefore preserve sufficient records to explain:

what data informed the decision;

which algorithmic model was used;

what legal rule authorised the outcome;

whether humans reviewed the recommendation; and

why affected parties were treated in a particular manner.

This is crucial in electricity regulation because automated rules could materially affect generators, suppliers, network companies and consumers.

Where adaptive systems effectively generate generally applicable regulatory rules, notice-and-comment procedures may also become necessary. PAJA expressly recognises procedural requirements where administrative action materially and adversely affects members of the public.

4. Case Law

Case Name/Citation: Pharmaceutical Manufacturers Association of South Africa: In re Ex Parte President of the Republic of South Africa 2000 (2) SA 674 (CC)

Facts: The President brought legislation into force before the administrative systems necessary for its implementation were ready.

Legal Issue: Whether exercises of public power must satisfy rationality requirements.

Judgment: The Constitutional Court held that rationality is a minimum constitutional requirement applicable to all exercises of public power.

Legal Principle/Ratio: Public power must be objectively rationally connected to the purpose for which it was granted.

Significance: A self-learning regulatory system cannot produce legally valid outcomes merely because an algorithm generated them. Its decisions must remain rationally connected to statutory objectives.

Case Name/Citation: Minister of Health v New Clicks South Africa (Pty) Ltd [2005] ZACC 14; 2006 (2) SA 311 (CC)

Facts: Pharmaceutical businesses challenged regulations establishing a pricing system adopted following recommendations by a statutory pricing committee.

Legal Issue: Whether administrative regulation-making could be reviewed under constitutional administrative-law standards.

Judgment: The Constitutional Court held, in the relevant reasoning, that regulation-making under empowering legislation could constitute administrative action and emphasised procedural fairness and public participation in rule-making.

Legal Principle/Ratio: Delegated regulatory rules remain subject to statutory authority, administrative justice and appropriate participatory procedures.

Significance: Algorithms cannot lawfully bypass consultation merely by continuously updating regulatory standards.

Case Name/Citation: Bato Star Fishing (Pty) Ltd v Minister of Environmental Affairs and Tourism 2004 (4) SA 490 (CC)

Facts: Fishing-right allocations made by an administrative authority were challenged as unreasonable.

Legal Issue: How courts should review technically complex administrative decisions.

Judgment: The Constitutional Court confirmed that judicial review derives principally from PAJA and the Constitution, while recognising appropriate respect for specialised administrative expertise.

Legal Principle/Ratio: Expertise does not remove administrative decisions from legality, reasonableness and judicial review.

Significance: The same principle applies to sophisticated AI-assisted electricity regulation.

5. Conclusion

Self-learning regulatory rule systems could make electricity governance faster and more responsive, but technological adaptability cannot displace legality, human accountability, transparency, procedural fairness, explainability and judicial review. Their legitimacy ultimately depends not on algorithmic intelligence but on continued compliance with constitutional and statutory regulatory authority.