South Africa's draft national AI policy was ambitious on paper: a new National AI Commission, an ethics board, a regulatory authority, tax incentives for private-sector AI investment, and plans to reduce the country's dependence on foreign-owned hardware infrastructure. It was unveiled for public comment, and within three weeks it was withdrawn. The reason was not political opposition, funding disputes, or constitutional challenge. The reason was that the document cited journal articles that do not exist.
Investigative outlet News24 found that at least six of the policy's 67 academic citations were fabricated: the journal names were real, but the articles themselves had never been written. Editors of the South African Journal of Philosophy, AI and Society, and the Journal of Ethics and Social Philosophy all confirmed independently that the cited works were nowhere in their archives. Communications Minister Solly Malatsi did not contest the finding. "The most plausible explanation is that AI-generated citations were included without proper verification," he wrote on X. "This should not have happened."
The hallucinations revealed something specific about how LLMs fail. These models are trained to predict what a plausible next token looks like, not to verify whether any given claim is true. When asked to produce an academic citation, a model will generate something that sounds authoritative: a plausible author name, a real-sounding title, a journal that exists, a volume and page number in the right format. The cited article itself is pure confection. The model has no way to check whether the text it is predicting corresponds to anything real. It is, in the most literal sense, making things up that sound right.
The scale of this problem is growing. A study published in Nature found that over 2.5 percent of academic papers published in 2025 contained at least one potentially hallucinated citation, up from 0.3 percent in 2024. That translates to more than 110,000 papers in a single year containing references to works that do not exist. The South Africa case is not an isolated embarrassment; it is the governance version of a pattern spreading through research, law, and policy writing globally.
The political fallout was immediate. The African National Congress, whose own members hold seats in the Government of National Unity alongside Malatsi's Democratic Alliance, called for the minister to appear before Parliament's Portfolio Committee on Communications and Digital Technologies. The ANC's study group demanded he explain who was responsible for drafting the policy and whether those people would face consequences. "South Africans expect their government to do the work," the ANC statement read, "not outsource the nation's most consequential policy documents to tools known to hallucinate fabricated facts and non-existent sources."
The deeper tension in the episode is almost theatrical. South Africa was attempting to write a policy governing a technology prone to confident fabrication, and used that technology to do it, without the verification steps that would have caught the problem. The minister's response -- "this unacceptable lapse proves why vigilant human oversight over the use of artificial intelligence is critical" -- is correct as far as it goes. But the commentary from civil society went further. Writing in GroundUp, analyst Tyronne McCrindle argued that the debacle illustrated not just an internal process failure but a broader pattern: AI systems being layered into government work faster than the institutions using them can critically evaluate their outputs.
The policy will be redrafted. The ANC is demanding the process be "driven by human expertise and rigorous evidence, not AI-generated shortcuts." Whether that demand produces a meaningfully different document is an open question. The tools that produced the hallucinated citations are the same tools that drafters will reach for again when they need to locate supporting literature quickly. The difference, presumably, will be a human checking each reference before it is published. That is not a sophisticated institutional response to AI's epistemological failures. It is what careful researchers have always done, before AI made it easy to skip.