Melissa Cogger, Partner at Bowmans

The Labour Relations Act, 1995 (LRA) provides every employee with the right not to be unfairly dismissed and not to be subjected to an unfair labour practice. Where artificial intelligence (AI) systems are used in HR-practices, for example, to generate performance ratings, identify employees for retrenchment, or make decisions about promotion, training, or dismissal, the interplay between algorithmic decision-making and the LRA’s protections becomes critical.

AI-generated performance ratings and dismissals for poor performance

The Code of Good Practice: Dismissal (Code), read with applicable case law, has established requirements for a fair dismissal on the ground of poor performance. These include, among other things, a clear and objective performance standard; communication of the standard to the employee; a fair opportunity to meet the standard; and a fair process before dismissal.

Where an AI system generates a performance rating that leads to a poor performance process and ultimately to dismissal, a number of LRA-related concerns come into play.

The first relates to explicability. The employer must be able to explain to the employee how the performance rating was generated. If the employer cannot explain the algorithm’s methodology (the so-called ‘black box’ problem) it will be difficult to demonstrate that the standard is clear, objective and fair.

Second, there is the question of objectivity. An AI-generated performance rating is only as objective as the data on which it is trained and the criteria it applies. If the algorithm incorporates irrelevant or biased variables, or focuses only on quantifiable metrics, the resultant rating may not constitute a fair and objective measure of performance. Soft skills such as interpersonal dynamics, leadership qualities, the ability to take on feedback, and collaboration with colleagues may be missed. Reliance on AI ratings accordingly risks providing a narrow and incomplete view of an employee’s overall contribution to the organisation.

Third, the employee’s opportunity to respond must be meaningful. An employee who is told that an algorithm has rated their performance as inadequate must be given an opportunity to challenge that rating. This requires, at minimum, that the employee understands what the algorithm measured and how it arrived at its conclusion.

Fourth, human oversight remains essential. A dismissal based solely on an AI-generated rating, without independent human assessment, is likely to be found procedurally and substantively unfair.

AI in retrenchment selection criteria

Section 189 of the LRA requires that, in retrenchment exercises, selection criteria must be agreed upon with consulting parties or, failing agreement, must be fair and objective. AI systems might be used to identify what potential selection criteria should be proposed by the employer, or could apply any agreed or determined selection criteria by the employer to suggest which employees are selected for retrenchment.

Where an AI system is used to identify employees for retrenchment (for instance, by scoring employees on skills assessments, qualifications, or experience) the following risks arise:

• the criteria applied by the algorithm may not be fair and objective (for example, where they incorporate biased variables that disadvantage employees on the basis of race, gender, age or disability);
• the employer may be unable to explain to employees or the consulting parties how the algorithm selected particular employees for retrenchment; and
• the algorithm may depart from agreed or conventional selection criteria (such as ‘Last in First Out’, known as LIFO) without a justifiable basis.

A retrenchment exercise that relies on opaque or unexplainable selection criteria driven by AI is vulnerable to challenge as procedurally and substantively unfair.

Promotion and training decisions driven by AI: unfair labour practices

Where AI systems are used to recommend employees for promotion or to identify employees for training and development opportunities, decisions that are arbitrary, irrational, or incapable of being explained and justified by the employer, may constitute unfair labour practices under section 186(2) of the LRA (as well as give rise to unfair discrimination claims). The following examples illustrate how AI-driven results could give rise to such disputes:

• An AI system recommends younger employees for upskilling programmes on the basis that they have a longer projected tenure with the organisation and therefore represent a better ‘return on investment’. Older employees who are excluded from these programmes may have a claim on the basis that the AI system’s recommendation constitutes unfair conduct relating to the provision of training or benefits, in addition to a separate claim of unfair discrimination on the basis of age.
• An employee passed over for promotion could argue that the AI-driven decision constitutes unfair conduct, particularly where the algorithm’s criteria indirectly discriminated against them or failed to account for legitimate absences such as maternity leave, family responsibility leave or periods of ill health, which can similarly give rise to unfair discrimination claims on the basis of sex, gender, family responsibility and disability status.

Over-reliance on algorithmic selection may inadvertently exclude employees with high potential who do not conform to established data patterns. As with performance ratings, the algorithm may focus on quantifiable metrics whilst overlooking qualitative factors, resulting in selections that do not account for the full value an employee contributes to the organisation.

AI-generated disciplinary outcomes

The emerging practice of disciplinary decision-makers using AI tools to assist in formulating findings, rulings, or sanction recommendations presents a distinct set of risks.

A particular danger is the phenomenon of AI ‘hallucinations’, where AI tools generate plausible-sounding but factually incorrect or fabricated information. In the legal context, this may manifest as fictitious case law, invented or incorrectly cited statutory provisions, or purported legal principles that are outdated or have no basis in authority. Relying blindly on such material in a disciplinary context, without independent verification, may compromise the fairness and integrity of the process and render the outcome vulnerable to challenge.

The Labour Court’s recent decision in Molawa and Others v Matjhabeng Local Municipality and Another provides instructive guidance on these risks. In that matter, the applicants sought to stay disciplinary proceedings pending the determination of a review application in which they challenged rulings made by the disciplinary chairperson. Central to their challenge was the allegation that the chairperson had relied on AI-generated authorities in support of his findings.

The Court found that several of the authorities cited by the chairperson either did not exist or did not support the propositions advanced. It had yet to be definitively established whether they were indeed generated by AI, but the Court held that whether the citations were generated by an ‘assistant’ or self-generated by the Chairperson himself, the result is the same: there was prima facie evidence that the chairperson failed to apply his mind to the law.  The Court accordingly granted the interdict, finding that it would not be in the interests of justice to require the employees to subject themselves to a disciplinary enquiry chaired by an individual whose decision-making had been called into serious question.

Employers should accordingly exercise caution if they permit decision-makers to use AI tools in disciplinary proceedings, and should implement guardrails on when and how (if any) AI can be used in disciplinary proceedings.

The requirement for fair process and the inability to explain AI decisions

A thread that runs through all of the above scenarios is the requirement for procedural fairness. The LRA and Code require that employees be given reasons for adverse decisions, an opportunity to respond, and a fair process before a decision is taken. Where the employer cannot explain the AI system’s reasoning, or where the decision is taken by the algorithm without meaningful human intervention, these procedural requirements cannot be met.

Employers must accordingly ensure that AI systems are used as decision-support tools, not as autonomous decision-makers, and that every material employment decision is subject to independent human review by a person who can articulate the reasons for the decision and respond to the employee’s representations.

To the extent that personal information is uploaded onto AI systems, employers must also have regard to their obligations under the Protection of Personal Information Act 4 of 2013.

The ultimate message for employers is clear: AI may inform, but humans must decide.

Ed’s note: This is Part 4 of a six-part series examining the responsible use of artificial intelligence (AI) in human resources from a South African legal perspective. You can find the previous three articles here:

AI in the workplace: The current regulatory landscape

AI in the workplace: POPIA considerations

AI in the workplace – Employment Equity Act considerations

ENDS