The most revealing part of South Africa’s AI-policy embarrassment was not that a computer apparently invented references. It was that fictitious sources survived long enough to appear in a draft approved for public consultation.
In April, communications & digital technologies minister Solly Malatsi withdrew the policy after an internal inquiry confirmed the problem. The question now is who should have checked those sources, with what evidence and what authority, to stop publication. That question matters far beyond one department.
A revised policy is taking shape. At an industry event on October 21, a senior official said it was likely to reach the cabinet in March 2027. The extra time creates an opportunity to set a more demanding standard for government and business: demonstrate that workers can find and correct AI errors before those errors become decisions affecting citizens, customers or employees.
Business Day has already called for human oversight. Quite right. But the phrase can conceal a weak process. An official may be instructed to check an AI-generated briefing without being given access to the underlying evidence, enough time to investigate, or permission to challenge a senior colleague’s preferred conclusion. That is oversight on paper. Organisations need a way to establish whether a real person could catch a plausible mistake under ordinary working conditions.
Consider a junior analyst asked to summarise evidence for a municipal procurement decision. An AI tool produces a polished paragraph with three citations. One paper does not exist, another discusses a different country, and the third supports a narrower claim than the summary suggests. A supervisor who simply reads for fluency might approve it. The US National Institute of Standards & Technology describes precisely this risk: generative systems can confidently produce false information and misleading citations. The problem is amplified when an answer looks ready for an official signature.
Here is a more useful test. Give an employee the draft, the relevant source materials and a deadline resembling the actual job. Ask the employee to locate the original publication for every consequential claim; check that its authors, date and findings match the summary; flag uncertainty; and record what was corrected or escalated. Seed some practice documents with known mistakes. An exercise where every citation is genuine teaches less than one where the employee must decide which claims can survive scrutiny.
Supervisors should explain why a proposed correction was accepted or rejected. Junior staff then learn to separate a broken link from an invented study, a genuine source from an inaccurate paraphrase, and an acceptable estimate from an unsupported assertion. They also learn to say, “I cannot verify this”, even when a fluent answer appears convincing. Certification for completing an introductory AI course gives an employer little evidence of such judgment.
This is particularly important in a country struggling to create pathways into skilled work. Stats SA reported 47.4% unemployment among people aged 15-34 in the second quarter of 2026. The International Labour Organisation’s 2026 review of workplace evidence also warned about the erosion of opportunities for younger workers as tasks change. If entry-level research, checking and drafting disappear from job descriptions, employers may inadvertently remove the very tasks through which inexperienced workers acquire professional judgment.
A practical response is a supervised error detection apprenticeship built into real workflows. Public departments, banks, insurers, professional firms and customer-service operations could allocate protected time for newcomers to review AI-assisted work, compare it with trusted records and discuss difficult cases with experienced colleagues.
The trainee would start with low-stakes material, then progress to more consequential documents as performance improves. Unesco’s AI competency framework places critical judgment alongside technical understanding. That is a sensible starting point for workplace training as well.
Performance should be assessed through tasks rather than attendance. Can a trainee find a fabricated source? Detect that a real regulation was quoted out of context? Identify a figure whose date makes it unsuitable for a current decision? Explain the correction in language a nonspecialist manager can understand? A worker who merely clicks “reviewed” has demonstrated none of these abilities. The assessment should record mistakes found and genuine material wrongly rejected, so employees are rewarded for accuracy rather than indiscriminate suspicion.
Then publish meaningful correction metrics, at least in aggregate if confidentiality matters. A department could disclose how many consequential AI-assisted documents were sampled, the proportion with material errors, how many errors reached decision-makers, and how long corrections took. A company could track the same measures internally and publish appropriate summaries for customers or regulators. Assign a named accountable role to each workflow and a route for stopping a document when serious doubts remain. An unmeasured instruction to “keep a human in the loop” cannot tell the public whether safeguards are improving.
Useful precedents
There are useful precedents. The UK’s Algorithmic Transparency Recording Standard requires specified public bodies to publish records for certain algorithmic tools affecting public decisions. South Africa need not copy that system wholesale. It could borrow the discipline of naming the tool, the responsible owner, its role in decisions and the evidence used to evaluate it. The local policy should go further by asking whether human reviewers actually detect failures when tested.
Existing South African law gives this argument practical weight. Section 71 of the Protection of Personal Information Act places limits, subject to exceptions and safeguards, on certain decisions based solely on automated profiling that substantially affect people. Human involvement should therefore be substantive where it is required, rather than a ceremonial click after the system has effectively decided. In a disputed credit, employment or service decision, an accountable employee needs a defensible record of the evidence and reasoning.
The revised policy should require pilots of these checks in selected public services, with published methods and results, before declaring the approach successful. Procurement contracts should provide access to source records and audit trails. Managers should be evaluated on correction quality as well as turnaround time. Businesses can begin testing their own workflows now, without waiting for cabinet. For AI adoption at work, the most valuable investment may be giving employees the time, tools and authority to disagree with a machine.
South Africa has already paid a reputational price for a document that nobody checked adequately. Another policy filled with worthy promises about skills would repeat the mistake in a different form. The credible test is straightforward: put a convincing but flawed AI answer in front of a trained worker, and see whether the error is stopped before the answer becomes somebody else’s problem.
• Dr Tsipursky, a behavioural scientist, is CEO of Disaster Avoidance Experts and author of ‘The Psychology of AI Adoption at Work: From Resistance to Results’.
Business Day
Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.
Please read our Comment Policy before commenting.