Regulating Algorithmic Decision-Making in South Africa

Overview

South Africa increasingly uses algorithmic and automated systems to process information, assess risk, verify eligibility and support decisions across the public and private sectors. These systems can improve efficiency, but when they influence access to social assistance, credit, employment, insurance, healthcare, education or public services, they can also affect rights, livelihoods and socioeconomic opportunity.

South Africa has an important legal foundation in the Constitution, the Promotion of Administrative Justice Act, the Promotion of Access to Information Act and the Protection of Personal Information Act. However, these protections do not yet form a comprehensive operational framework for all high-impact algorithmic decision-making—particularly where a system materially influences an outcome without technically making the final decision itself.

SADGF proposes a risk-based framework built around algorithmic due process. People affected by consequential automated decisions should know when automation influenced an outcome, receive a meaningful explanation, correct materially inaccurate data, contest the decision, obtain genuine human reconsideration and access appropriate remedies.

The objective is not to restrict beneficial automation. It is to ensure that no consequential decision becomes less accountable merely because technology is involved.

Where automated decisions affect people

Algorithmic systems increasingly influence decisions involving creditworthiness, social grants, employment, insurance, public services and the allocation of public resources.

While automation can improve efficiency, decisions that materially affect people’s lives cannot be treated as purely technical. Important concerns include data accuracy, decision logic, discrimination, accountability and the ability to challenge an outcome.

Section 33 of the Constitution guarantees lawful, reasonable and procedurally fair administrative action, while the Promotion of Administrative Justice Act gives effect to those protections. The Protection of Personal Information Act also provides safeguards relating to decisions based solely on automated processing.

A governance gap nevertheless remains. Modern systems may rank, recommend, flag, score or verify while a human formally makes the final decision. A nominal human-in-the-loop should not remove a high-impact system from meaningful oversight when the technology materially determines the outcome.

Automated decision-making creates an accountability gap when an affected person cannot understand:

• Whether an algorithm materially influenced the decision
• What information was used
• Whether that information was accurate and relevant
• How the information was interpreted
• Why the system produced or recommended a particular outcome
• Who is legally and institutionally responsible
• How the outcome can be corrected, reviewed or challenged

The consequences of an incorrect automated decision are not merely technical. They can be social, economic, legal and constitutional.

Social grants and automated verification

The Social Relief of Distress grant litigation involving the Institute for Economic Justice and #PayTheGrants illustrates the risks of automated decision-making in the public sector.

The eligibility system used automated bank verification to assess whether applicants had insufficient means. Money entering an applicant’s bank account during a particular month could be treated as income without determining why the money was deposited or whether it constituted actual income.

Once-off family payments, emergency assistance, money received for specific expenses or funds held on behalf of another person could therefore be treated in the same manner as employment income.

The High Court found aspects of the system unfair and unreasonable. The case demonstrates a central governance problem: an automated system can apply a rule consistently while still producing substantively unfair outcomes when the data does not capture important human context.

The state’s appeal was heard by the Supreme Court of Appeal on 25 August 2026, and judgment was reserved. SADGF does not prejudge the outcome. The case nevertheless remains an important illustration of the governance questions raised when automated verification materially influences access to essential public assistance.

Credit and financial services

Automated systems can assess creditworthiness, detect fraud and influence access to loans and other financial products. Where these decisions rely on incomplete, inaccurate, proxy or discriminatory data, individuals may face financial exclusion without adequately understanding why they were rejected or how to challenge the outcome.

Employment

Algorithmic systems can support recruitment, candidate screening, productivity assessment and other workplace decisions. Their use raises questions about bias, discrimination, transparency and whether applicants or employees can meaningfully challenge an adverse outcome.

Insurance and essential services

Automated risk scoring can influence insurance pricing, coverage and access to other essential services. Models relying on opaque correlations or inaccurate information can produce significant consequences even when no individual decision-maker intends an unfair result.

Public services

As government digitises service delivery, automated systems may increasingly determine eligibility, prioritisation, verification and access. Efficiency should not come at the expense of procedural fairness, equality or accountable public administration.

A risk-based accountability framework

SADGF supports a risk-based and technology-neutral approach in which governance requirements are determined by the context and consequences of a system’s use.

Human accountability must remain. Institutions cannot avoid responsibility by attributing decisions to an algorithm or private technology provider.

Decisions must be explainable. Affected individuals should understand the material basis of significant decisions and know which aspects can be challenged.

Data must be accurate and relevant. Systems should not rely on materially inaccurate, outdated or irrelevant information, and practical mechanisms for correction must be available.

Human review must be genuine. Reviewers must possess the authority, information and discretion necessary to reconsider and overturn an automated outcome.

Higher-risk decisions require stronger safeguards. Greater impacts on rights, opportunities or essential services should trigger stronger assessment, transparency, audit, review and monitoring requirements.

Public accountability cannot be outsourced. When government procures or outsources an algorithmic system, constitutional and statutory accountability remains with the public institution.

Algorithmic due process

SADGF proposes algorithmic due process as an organising principle for consequential automated decision-making.

When an automated or algorithmic system materially affects a person’s rights, benefits, opportunities or access to essential services, that person should, according to the level of risk, be able to:

• Know that automation materially influenced the decision
• Know the significant information or categories of information relied upon
• Receive an understandable explanation of the material basis for the outcome
• Correct materially inaccurate or outdated data
• Contest the decision and make relevant representations
• Obtain meaningful human reconsideration by a person with authority and discretion
• Access an effective remedy where an unlawful or materially erroneous automated decision causes harm

What counts as high-impact algorithmic decision-making?

SADGF proposes stronger governance requirements for systems that materially influence decisions concerning:

• Social grants and social security
• Access to public services or benefits
• Allocation of public resources
• Creditworthiness and financial services
• Employment and recruitment
• Healthcare and education
• Law enforcement, security, immigration or identity
• Insurance and other essential services
• Administrative penalties or enforcement
• Other legal, constitutional or socioeconomic interests

The focus should be on consequences rather than technological sophistication. A rules-based automated system, statistical scoring model or machine-learning system can each create significant governance risks when used in a consequential decision context.

From automation to accountable governance

South Africa should not approach algorithmic decision-making as a choice between innovation and regulation. Well-governed automation can improve efficiency, expand access and support better public administration. Governance allows those benefits to be pursued without weakening constitutional rights or public trust.

When an automated system materially influences whether someone can access essential social assistance, financial opportunity, employment or a public service, the decision is no longer merely technical. It becomes a question of rights, fairness and accountability.

SADGF calls for a regulatory approach in which:

• No consequential automated decision is without identifiable accountability
• No person affected by a high-impact outcome is denied meaningful review
• No system is beyond appropriate scrutiny because it is technically complex or commercially proprietary
• No nominal human sign-off is accepted where the reviewer lacks real authority, information or discretion
• No institution is permitted to outsource accountability to an algorithm or technology provider

South Africa’s digital future should not only be automated. It should be accountable.