Two Home Affairs Officials Suspended After AI 'Hallucinations' Found in Policy Paper

South Africa’s Department of Home Affairs (DHA) suspended two senior officials this week after discovering fictitious AI-generated references in the revised white paper on citizenship, immigration and refugee protection. The department acknowledged the error caused embarrassment and framed the incident as a catalyst to modernize its internal processes.

The suspended officials include the Chief Director of the citizenship and immigration unit, whose suspension took effect Thursday, and a director involved in drafting the document, set to be suspended at the start of the next work week. DHA stated the standalone reference list attached to the white paper contained hallucinations, a term for erroneous or fictitious large language model outputs that stem from a lack of credible training data (OpenAI’s guide to AI hallucinations). These references were not cited in the body of the policy document and appear to have been generated and attached after the fact.

To address the issue, DHA appointed two independent law firms, one to manage the disciplinary process for the suspended officials and another to review all policy documents produced by the department since November 30, 2022. That date marks the public release of ChatGPT, the first widely adopted consumer large language model, which DHA identified as the likely starting point for potential AI use in its policy work.

This incident follows a similar case one week earlier at the Department of Communications and Digital Technologies (DCDT), which withdrew its draft National AI Policy after finding fictitious sources and citations in the document. Minister Solly Malatsi stated the most plausible explanation was AI-generated citations included without proper verification, adding that the draft policy would be reworked. DHA has also withdrawn the reference list attached to its white paper but maintains the core policy content remains valid, noting the document was developed through cross-departmental collaboration and public consultation, with no material impact from the flawed references.

The DHA acknowledged AI as a transformative, disruptive technology changing how organizations operate across sectors, noting that public institutions must adapt to harness its benefits responsibly. The department will design and implement AI checks and declarations as part of its internal approval processes, part of a broader push to update workflows to account for generative AI use.

From a startup ecosystem perspective, this pair of incidents highlights a persistent gap between AI adoption hype and operational reality. Government agencies, like many enterprises, have rushed to integrate generative AI tools to speed up administrative work, skipping the verification and governance steps that early-stage startups treat as table stakes. For startups building AI governance, hallucination detection, and citation verification tools, the fallout creates a clear market opportunity. Public sector agencies, which handle high-stakes documents with legal and policy implications, are now actively seeking ways to vet LLM output, a need that has driven growing venture interest in AI safety startups.

The DHA’s plan to audit all policy documents back to late 2022 also signals a broader trend of organizations retroactively checking work for unvetted AI use, another area where specialized startups can offer services. While the DHA did not disclose plans to partner with private vendors for its new AI checks, similar government adoptions in other markets have already driven contracts for AI governance firms. Skeptics of AI hype will note that the DHA’s claim that core policy content remains unaffected relies on the assumption that only the reference list was AI-generated, a claim that the ongoing document review may test.

This is not an isolated issue. Enterprises across industries have reported similar problems with LLM-generated content, from fake legal citations in court filings to incorrect medical advice in patient materials. For the startup ecosystem, the pattern reinforces that the most sustainable opportunities in AI are not in the hype-driven development of new models, but in the less glamorous work of making existing models reliable, verifiable, and safe for official use.