Expert: Solly Malatsi, Minister of Communications and Digital Technologies, South Africa When a government publishes a national AI policy, the bibliography is not decoration - it is the evidence base. In April 2026 South Africa's Draft National Artificial Intelligence Policy was published in the Government Gazette, and 17 days later the Department of Communications and Digital Technologies withdrew it. The reason: some of the research it cited does not exist. Civil rights group Article One wrote to the department after checking the policy's sources and found that at least six of the 67 entries in its bibliography were AI hallucinations. News24 reported that several of the academic journals cited in the document were "completely fictitious". Minister Solly Malatsi confirmed the finding in a statement: "The most plausible explanation is that AI-generated citations were included without proper verification. There will be consequence management for those responsible for drafting and quality assurance." The cost was not only reputational. South Africa was set to become the first African nation to adopt a policy establishing a formal ethics board to oversee AI outside the West, and it lost that milestone to a bibliography that no one checked. As Rest of World noted, this is the first time a government has withdrawn a document because of AI hallucinations. It is not the first time AI-generated citations have reached official material. Deloitte confirmed in a September 2025 letter to Australia's Department of Finance that "the use of the generative AI tools had resulted in inaccurate outputs whereby certain citations, in the form of footnotes and sources in the accompanying reference list, contained errors" in a report it had prepared for Australia's Department of Employment and Workplace Relations. Deloitte republished the corrected study and, in November, refunded the Australian government $290,000 of the $440,000 it had charged for it. That refund belongs to the Australian engagement, not to the South African policy. Europe's cybersecurity agency ENISA admitted that two of its 2025 threat reports contained AI-hallucinated sources: researchers from Westfalische Hochschule found that 26 of the 492 footnotes in one report were incorrect, according to Der Spiegel. AI law and data ethics researcher Chiara Gallese framed the pattern as a process failure rather than a skill failure: "No mandatory verification step. No provenance checks. No clear rule for AI use. Just speed, convenience, and trust-by-default." What makes the South African case the sharpest of the group is where the fabricated citation pointed. A policy designed to govern AI was itself built partly on sources that an AI invented, and the error survived drafting and quality assurance all the way to gazette publication. The check that catches this is mundane: open the reference, confirm it exists, confirm it says what it is claimed to say. The minister's answer for accountability was consequence management. The technical answer is that any workflow in which AI-generated citations reach publication without reference-level checking has no quality assurance at all. Source: https://restofworld.org/2026/government-ai-hallucinations-south-africa-deloitte/