Introduction
The increasing use of automated systems and artificial intelligence (AI) in recruitment is transforming how employers find and hire staff. These technologies promise efficiency and objectivity by analysing vast numbers of applications against set criteria. However, this automation raises significant legal and ethical concerns, particularly regarding the potential for discrimination. In South Africa, the legal framework for preventing workplace discrimination is principally the Employment Equity Act 55 of 1998 (EEA). This essay will argue that while the EEA's prohibition of indirect discrimination provides a theoretical basis for challenging biased algorithmic hiring, a significant regulatory gap exists in practice. The unique characteristics of automated systems, such as their complexity and opacity (the "black box" problem), create substantial evidentiary hurdles for applicants and challenges for legal oversight, rendering the EEA's protections difficult to enforce in this new technological context.
The Framework for Indirect Discrimination under the EEA
The primary objective of the Employment Equity Act is to achieve equity in the workplace by promoting equal opportunity and fair treatment through the elimination of unfair discrimination (EEA, s 2). Section 6(1) of the EEA prohibits an employer from unfairly discriminating, directly or indirectly, against an employee or job applicant on one or more of the listed grounds, including race, gender, sex, pregnancy, marital status, ethnic or social origin, colour, sexual orientation, age, disability, religion, and others.
While direct discrimination is often easy to identify, indirect discrimination is more subtle. It occurs when a policy, practice, or criterion that appears neutral on its face has a disproportionately negative effect on a particular group protected under the EEA, and which cannot be justified. The Constitutional Court in Harksen v Lane NO and Others [1997] ZACC 12 established a foundational test for unfair discrimination under the Constitution, which informs the interpretation of the EEA. The test involves determining whether a practice differentiates between people, whether this differentiation is on a specified ground, and whether the discrimination is unfair.
In the employment context, an applicant claiming indirect discrimination must first establish a prima facie case. This involves showing that a particular "employment policy or practice" has had an adverse impact on them and others from a protected group (EEA, s 11). Once this is shown, the burden shifts to the employer to prove that the discrimination was not unfair or that the policy or practice is justifiable in the circumstances, for example by showing it is based on the inherent requirements of the job (EEA, s 6(2)(b)). Cases like Independent Municipal & Allied Trade Union v City of Cape Town [2007] 3 BLLR 195 (LC) illustrate how a seemingly neutral policy, such as a particular recruitment test, can be found to be indirectly discriminatory if it disproportionately excludes a protected group without adequate justification. This established legal framework, therefore, clearly contemplates challenges to facially neutral employment practices that have discriminatory outcomes.
The Challenge of Algorithmic Decision-Making
Automated hiring tools range from simple keyword-based CV-screening software to complex machine-learning algorithms that analyse video interviews and predict job performance. These systems are intended to be neutral, applying the same rules to every candidate. However, they can embed and amplify human biases, leading to discriminatory outcomes. This algorithmic discrimination often manifests indirectly. For example, an algorithm trained on historical hiring data from a company that has traditionally employed more men in senior roles may learn to associate male characteristics with success, thereby down-selecting female applicants (Le Roux, 2018). Similarly, a tool that uses postcodes to filter candidates might inadvertently discriminate on the grounds of race due to the enduring spatial inequalities left by apartheid-era policies like the Group Areas Act.
The central problem that algorithms pose for the EEA's framework is one of transparency. Many advanced algorithms are considered a "black box," meaning that even their creators may not be able to explain exactly why a particular decision was made (Herselman, 2021). The decision-making process is contained within complex mathematical models and vast datasets. For a rejected job applicant, this creates an almost insurmountable challenge. Unlike a traditional recruitment process where a human decision-maker can be questioned, the "policy or practice" in an algorithmic context is the algorithm itself—its code, its training data, and its weighting of variables. An unsuccessful applicant is unlikely to know that an algorithm was used, let alone have access to the information needed to demonstrate that it had a disproportionate impact on their protected group (Singh, 2020).
Analysing the Regulatory Gap in Practice
When the legal test for indirect discrimination is applied to the context of algorithmic hiring, the practical regulatory gap becomes clear. The EEA was drafted long before the widespread adoption of AI in the workplace and is not equipped to deal with the specific challenges it presents.
First, the evidentiary burden on the applicant is a major obstacle. Section 11 of the EEA places the initial onus on the complainant to establish a prima facie case. To do this in an algorithmic context, an applicant would need statistical evidence showing that the hiring tool disproportionately disadvantages their group. Gathering this evidence is nearly impossible for an individual who has no access to the employer's overall application data or the algorithm's decision-making logic (Singh, 2020). Without this initial evidence, a claim cannot get off the ground, and the burden of proof never shifts to the employer. This effectively shields algorithmic discrimination from legal scrutiny.
Second, the concept of justification becomes problematic. If an applicant were to succeed in establishing a prima facie case, the employer would then have to justify the practice. An employer might argue that the algorithm is simply selecting for the "inherent requirements of the job." However, if the employer does not fully understand how the algorithm reaches its conclusions, they cannot meaningfully explain or defend its rationality. Relying on a third-party vendor's assurance that their tool is "bias-free" may not be a sufficient defence, as the employer remains liable for any discrimination that occurs in its name (EEA, s 60). The lack of transparency makes it difficult for both the employer to meet its legal obligations and for a court to assess the fairness of the justification.
Finally, there is an issue of accountability. The EEA holds the employer responsible for discriminatory acts. Yet, when the discriminatory tool is designed and maintained by an external tech company, it creates a complex chain of responsibility. While the employer is the legally liable entity, the source of the bias lies with the technology provider. The current legal framework provides little guidance on how to apportion liability or compel transparency from third-party vendors who may consider their algorithms to be protected trade secrets. This situation leaves applicants without a clear path to a remedy and employers exposed to legal risk for technology they may not control or understand.
Conclusion
In conclusion, while South Africa's Employment Equity Act provides a strong principled stand against indirect discrimination, it is ill-suited to the practical realities of algorithmic hiring. The legal framework, which relies on the identification of a "policy or practice" and the ability of a complainant to provide evidence of its adverse impact, breaks down when faced with opaque and complex automated systems. The evidentiary burden on applicants is too high, the process for justifying algorithmic decisions is unclear, and the lines of accountability are blurred. This creates a significant regulatory gap where discrimination can occur with little chance of detection or legal recourse. For the EEA to remain effective in protecting against unfair discrimination in the 21st-century workplace, legislative updates or the creation of a specific Code of Good Practice on the Use of AI in Employment are urgently needed. Such reforms would need to address issues of algorithmic transparency, auditing, and the allocation of responsibilities between employers and technology developers to ensure that the promise of fairness and equity is not lost in the age of automation.
References
- Herselman, K. (2021) 'Regulating artificial intelligence in South Africa: A case for a Sector-specific approach?', Potchefstroom Electronic Law Journal, 24, pp. 1-36.
- Le Roux, R. (2018) 'Is the age of the algorithm the age of discrimination? Some preliminary thoughts on the legal implications of using artificial intelligence in the workplace', South African Mercantile Law Journal, 30(1), pp. 108–131.
- Singh, A. (2020) 'Artificial intelligence and employment law: A South African perspective', SA Journal of Labour Law, 1(1), pp. 1-13.
Case Law
- Harksen v Lane NO and Others [1997] ZACC 12; 1998 (1) SA 300 (CC).
- Independent Municipal & Allied Trade Union v City of Cape Town [2007] 3 BLLR 195 (LC).
Legislation
- Employment Equity Act 55 of 1998 (South Africa).
- The Constitution of the Republic of South Africa, 1996.
