Authored By: Khanide Seti
University of South Africa
I. Introduction
The Fourth Industrial Revolution has transitioned from a futuristic concept to a present-day legal reality within the Republic of South Africa. As private corporations and state organs increasingly integrate Artificial Intelligence (AI) into decision-making processes — ranging from credit scoring and recruitment to predictive policing — the legal community faces a profound challenge. While these technologies promise efficiency and objectivity, they often act as “black boxes” harbouring opaque algorithms that can perpetuate or even amplify historical socio-economic biases.
In a nation defined by transformative constitutionalism, the intersection of emerging technology and human rights is not merely a technical concern but a constitutional imperative. The preamble of the Constitution of the Republic of South Africa, 1996 commits to “heal the divisions of the past and establish a society based on democratic values, social justice and fundamental human rights.” If AI systems are permitted to operate without stringent legal oversight, they risk becoming new instruments of systemic exclusion. This article provides a critical analysis of the current South African legislative landscape, examining whether Section 9 (the Equality Clause) and associated statutes like POPIA and PEPUDA are sufficient to mitigate algorithmic bias, particularly in light of the recently gazetted Draft National AI Policy of April 2026.
II. The Constitutional Anchor: Section 9 and Transformative Equality
To understand the legal challenge, one must first look to the supreme law of the Republic. Section 9 of the Constitution does not merely prohibit discrimination; it imposes a positive duty on the state and private parties to promote equality.1
2.1 Formal vs. Substantive Equality
South African jurisprudence, established in landmark cases such as Harksen v Lane NO,2 distinguishes between formal equality (treating everyone the same) and substantive equality (recognising historical disadvantage). AI algorithms are often designed with a “colour-blind” or “data-neutral” logic, which fits the definition of formal equality. However, because these algorithms are trained on historical datasets that reflect South Africa’s unequal past, they often produce results that are substantively unequal.
For instance, if an AI recruitment tool uses “historical success” in a corporate firm as a training metric, it will likely favour candidates from privileged backgrounds, thereby indirectly discriminating against those from marginalised communities. Under Section 9(3), the Constitution prohibits both direct and indirect unfair discrimination. The legal hurdle lies in the “unfairness” inquiry. In the context of AI, how does a litigant prove that an automated decision was “unfair” when the underlying logic of the machine is proprietary or computationally incomprehensible?
2.2 The Burden of Proof and the Presumption of Unfairness
Section 9(5) of the Constitution provides that once discrimination is established on a listed ground (such as race, gender, or disability), it is presumed unfair unless the contrary is established. In traditional litigation, this shift in the burden of proof is a powerful tool for the marginalised. However, in the “black box” era, even establishing that discrimination has occurred is a Herculean task. If a bank’s AI denies a loan application, the applicant rarely receives a breakdown of the variables used. Without transparency, the presumption of unfairness in Section 9(5) remains dormant protection, as the “trigger” — the evidence of discrimination — is hidden behind layers of code.
III. The Statutory Gap: POPIA, PEPUDA, and the Illusion of Control
While the Constitution provides the overarching values, the practical regulation of AI in South Africa currently rests on two primary legislative pillars: the Protection of Personal Information Act 4 of 2013 (POPIA) and the Promotion of Equality and Prevention of Unfair Discrimination Act 4 of 2000 (PEPUDA).3 However, a critical examination reveals significant gaps in how these Acts address algorithmic harm.
3.1 POPIA Section 71: The “Human-in-the-Loop” Fallacy
Section 71 of POPIA is the most direct legislative attempt to regulate Automated Decision-Making (ADM). It grants data subjects the right not to be subject to a decision which results in legal consequences or substantial effects, if that decision is based solely on the automated processing of personal information.
The critique here lies in the word “solely.” In many South African corporate environments, AI systems provide a “recommendation” which is then formally approved by a human consultant. This creates a “Human-in-the-Loop” scenario that technically bypasses Section 71. However, studies in automation bias suggest that humans rarely overrule algorithmic suggestions, especially when they do not understand the underlying logic. Consequently, Section 71 risks becoming a “paper tiger” — a protection that is easily circumvented by keeping a human figurehead in a process that is, for all intents and purposes, fully automated.
3.2 PEPUDA and the “Black Box” Evidence Problem
PEPUDA was enacted to give effect to Section 9’s right to equality. It prohibits unfair discrimination by the state and all persons. However, the Act was drafted with human agency in mind. Under Section 13, the complainant must prima facie establish the existence of discrimination.
In the context of AI, this creates an insurmountable evidentiary wall. If a recruitment algorithm consistently filters out candidates from specific South African townships, the victim may suspect bias but will lack the technical evidence to prove it. The “black box” nature of deep learning means that even the developers of the AI may not be able to explain exactly why a specific decision was reached. Without a statutory right to an “algorithmic explanation,” PEPUDA’s protections against indirect discrimination remain largely unenforceable in the digital economy.
IV. Analysis of the 2026 Draft National AI Policy: A New Dawn?
The gazetting of the Draft National AI Policy in April 2026 represents a pivotal shift in South African legal strategy. It acknowledges that existing laws are insufficient and proposes a proactive regulatory framework.4
4.1 The Risk-Based Categorisation
The 2026 Policy introduces a “Risk-Based Approach” similar to the EU AI Act, categorising AI systems into “Unacceptable,” “High,” and “Low” risk. High-risk systems — such as those used in healthcare, justice, and credit scoring — will be subject to mandatory Algorithmic Impact Assessments (AIAs).
From a critical legal perspective, this is a positive step. However, the effectiveness of AIAs depends entirely on the standard of transparency required. If these assessments are merely “checkbox” exercises performed by companies themselves, they will fail to protect the public. The policy must mandate that these assessments be audited by the proposed AI Regulatory Authority to ensure they align with the substantive equality requirements of Section 9 of the Constitution.
4.2 The AI Ombudsperson and Access to Justice
The proposal for an AI Ombudsperson is perhaps the most innovative aspect of the 2026 Policy. In a country with high levels of inequality and varying degrees of digital literacy, a dedicated office to handle algorithmic grievances is essential. The challenge will be jurisdiction. How will the AI Ombudsperson interact with the existing Information Regulator? There is a risk of “regulatory overlap,” where victims are shuffled between departments, leading to a dilution of accountability. For the 2026 Policy to succeed, the AI Ombudsperson must have the power to subpoena source code and mandate “human intervention” in cases where algorithmic bias is suspected.
V. Towards an Ubuntu-Based AI Jurisprudence
In seeking a regulatory framework that is uniquely South African, we must look beyond the Western-centric models of the European Union’s AI Act. A transformative legal approach requires the integration of Ubuntu — the constitutional value of communal interconnectedness and human dignity — into the heart of algorithmic governance.
Dignity as the Antidote to Dehumanisation
The “black box” nature of AI threatens the constitutional value of dignity because it treats citizens as data points rather than human beings with agency. By applying Ubuntu, we can argue for a mandatory “Right to Human Contact” in critical legal and administrative decisions. This ensures that no South African is ever denied a fundamental right (such as healthcare or housing) based solely on a machine’s calculation, without meaningful human review that considers the person’s unique socio-economic context.
VI. Recommendations for Legislative Reform
To bridge the gaps identified in this analysis, the following legislative and policy interventions are recommended:
Amendment of POPIA Section 71: The word “solely” must be removed or clarified. The law should protect citizens against any decision where an automated system plays a predominant or material role, preventing companies from using “human rubber-stamping” to bypass accountability.
Introduction of “Algorithmic Transparency” in PEPUDA: PEPUDA should be amended to include a statutory “Right to Explanation.” This would require entities using high-risk AI to provide a plain-language explanation of the logic involved in a decision, thereby enabling victims to establish the prima facie case of discrimination required for Section 9 litigation.
Technical Capacity for the Regulatory Authority: The 2026 Draft Policy’s proposed AI Regulatory Authority must be shielded from political interference and adequately funded to hire data scientists alongside legal experts. A regulator without technical auditing power will be unable to “pierce the veil” of complex proprietary code.
A “Rebuttable Presumption” of Bias: In cases involving historically marginalised groups, the law could introduce a rebuttable presumption that an algorithm is biased if its outcomes mirror historical patterns of exclusion. This would shift the burden to the corporation to prove that their AI is substantively fair.
VII. Conclusion
As South Africa navigates the complexities of the 2020s, the legal system stands at a crossroads. Technology, while a tool for progress, possesses no inherent moral compass; it reflects the values and the prejudices of its creators and the data upon which it is fed.
The Constitution of the Republic of South Africa remains our most powerful defence against new forms of digital oppression. However, the “lens” of Section 9 must be sharpened to detect the invisible threads of algorithmic bias. While the 2026 Draft National AI Policy is a commendable step forward, it must be anchored in the substantive equality and human-centric values of Ubuntu. By evolving our legislative frameworks — specifically POPIA and PEPUDA — we can ensure that the Fourth Industrial Revolution does not become a tool for “automated apartheid,” but rather a means of achieving the truly equitable society envisioned in 1996. The law must not merely react to technology; it must govern it with the same transformative spirit that dismantled the legal structures of the past.
VIII. Bibliography
Constitution of the Republic of South Africa, 1996.
Protection of Personal Information Act 4 of 2013.
Promotion of Equality and Prevention of Unfair Discrimination Act 4 of 2000.
Harksen v Lane NO and Others [1997] ZACC 12.
Department of Communications and Digital Technologies, Draft National Artificial Intelligence (AI) Policy (GN 123, GG 50456 of 1 April 2026).
Footnote(S):
1 Constitution of the Republic of South Africa, 1996.
2 Harksen v Lane NO and Others [1997] ZACC 12.
3 Protection of Personal Information Act 4 of 2013; Promotion of Equality and Prevention of Unfair Discrimination Act 4 of 2000.
4 Department of Communications and Digital Technologies, Draft National Artificial Intelligence (AI) Policy (GN 123, GG 50456 of 1 April 2026).
