South Africa yanks AI policy over 6 fabricated citations

A cabinet-approved policy document lasted 17 days before someone worked out its footnotes were invented. South Africa withdrew its Draft National Artificial Intelligence Policy in April after at least six of 67 sources in the bibliography turned out to be fictitious. Communications Minister Solly Malatsi called it an 'unacceptable lapse' and said 'the most plausible explanation is that AI-generated citations were included without proper verification,' per reporting from Rest of World.

What makes the story worth reading is that South Africa is not the outlier. Rest of World's cross-jurisdiction check catalogues a lengthening list of AI-fabricated citations turning up in official documents. Deloitte refunded $290,000 of the $440,000 it charged Australia's Department of Employment and Workplace Relations for an August 2025 report whose footnotes did not check out, and had a separate $1.2 million, 526-page healthcare report for Newfoundland and Labrador exposed in November for similar problems. The European cybersecurity agency ENISA had 26 of 492 footnotes come back incorrect in one report. The Trump administration's Make America Healthy Again report on children's health cited studies that did not exist, and the Washington Post reportedly found 'oaicite' fragments attached to URLs, which is often read as a marker of ChatGPT use.

The exposure is not really about the models. It is about the QA process at the consultancies and ministries that ship these documents. Chiara Gallese, an AI law and data ethics researcher quoted in the story, described ENISA as having let AI 'touch the one layer it should never touch unguarded: the truth layer.' If you are a leader signing off on an AI-assisted deliverable, the risk you are underwriting is now legal and reputational, not just editorial, and it is already translating into refunds and new contract language.

The honest caveat is that the reporting is a snapshot rather than a systematic audit. It does not tell you how many similarly affected reports are sitting unchecked, whether Deloitte's contracts previously disclosed AI use, or what actual consequences will follow for the South African drafters Malatsi flagged for review. What it does show is a template forming: Newfoundland and Labrador has already updated contracts to require disclosure of 'all intended uses of AI and/or machine learning,' with reserved rights to assess risks. Expect other procurement teams to copy that clause quickly, and expect a small market to open up for whoever can credibly certify that a report's citations are real.