South Africa's Draft National Artificial Intelligence Policy was withdrawn on 26 April 2026. It had been live for sixteen days.
The reason: the reference list contained citations to academic journals that had never published the articles attributed to them. News24 journalists contacted the journal editors, who confirmed the articles did not exist. The document had passed through Cabinet twice and through multiple layers of departmental review before that check was made.
Two officials were placed on precautionary suspension on 30 April. The minister called it "not a mere technical issue" that had "compromised the integrity and credibility of the draft policy." The parliamentary committee chair called for a redraft without, in her words, using ChatGPT this time.
The irony is structural. A governance framework for AI was drafted using AI, submitted without a citation verification step, approved at the highest level of government, and published for public comment. The fabrications passed every approval gate because no approval gate checked them. Cabinet ratification is a political decision, not a quality gate. The two are not the same thing.
If six citations were fabricated, what else in the 86 pages was not verified?
This is not a unique failure. In October 2025, a AU$440,000 government report produced by Deloitte Australia for the Department of Employment and Workplace Relations contained fabricated academic citations and a misattributed federal court judgment. Chris Rudge of the University of Sydney identified the errors. Deloitte acknowledged the use of Azure OpenAI and refunded approximately AU$97,000. The same failure mode: AI output accepted as authoritative without a verification step.
The pattern is simple. AI systems generate plausible-sounding citations. Plausibility is not accuracy. The check that catches fabrications is not a political approval process. It is a verification step applied before the output is accepted, at the point where it still can be corrected.
South Africa's policy had substantive ambitions: a National AI Commission, an AI Ethics Board, an AI Regulatory Authority, an AI Ombudsperson, a National AI Safety Institute, and an AI Insurance Superfund modelled on the Road Accident Fund for liability cases where responsibility is uncertain. The governing bodies designed to catch AI failures in practice were built into the policy. The verification step that would have caught the AI failure in the policy itself was not.
The draft will be revised. The lesson is available now.