Sixteen Days

Cabinet approved the draft National Artificial Intelligence Policy in late March and early April 2026. It went into the Government Gazette on 10 April, open for public comment until 10 June1 Source 1. Resolves. The document is served at this address. No scholarly index holds a record of it.. It was, by any ordinary measure, a serious document from a serious process.

Within days, journalists started pulling on the reference list. Several of the sixty-seven sources turned out not to exist. Researchers had been credited with foundational work on subjects they had never written a word about4 Source 4. Resolves. The document is served at this address. No scholarly index holds a record of it.. Nobody inside the process had caught it; the check that mattered happened in a newsroom, after publication5 Source 5. Resolves. The document is served at this address. No scholarly index holds a record of it..

On 26 April the Minister of Communications and Digital Technologies announced the withdrawal2 Source 2. Resolves. The document is served at this address. No scholarly index holds a record of it.. The formal retraction followed by gazette some weeks later3 Source 3. Resolves. The document is served at this address. No scholarly index holds a record of it.. Sixteen days from publication to withdrawal, and the drafting method became a political question rather than an administrative one7 Source 7. Resolves. The document is served at this address. No scholarly index holds a record of it..

The profession that has to live with documents read it precisely. One law firm titled its client note The Hallucinatory Irony of it All6 Source 6. Resolves. The document is served at this address. No scholarly index holds a record of it.. The irony is real and it is also the least interesting thing here.

A fabricated citation is not evidence that a model is untrustworthy. Everyone already knew that models produce fluent references to things that were never written. The fabrication is evidence that between the drafting of a claim and the publication of a national policy, there was no step at which anyone could establish where a source had come from.

That gap existed before this technology and will outlast it. What the technology did was make it cheap to fill sixty-seven slots in a reference list, and expensive to find out which ones were real. The tool changed the economics of a check that was always supposed to happen.

A document is defensible when a hostile reader, holding only what you published, can get back to where each claim came from. Everything else is a claim about your own good intentions.

Every source under this article was put through a real check. Where a DOI exists it was resolved against Crossref. Where one does not, the title was searched against Crossref and OpenAlex. Every address was requested to see whether anything is actually served there. The results are below, and the button above runs the whole thing again on demand.

The result was not the one being fished for. Only 2 of the 13 sources came back confirmable by machine. 11 of them are live, readable, and held by no index at all.

The gazette carrying the policy is a PDF on a state file server. The department’s own terms of reference for the evidence work behind the national GBVF plan is a procurement document10 Source 10. Resolves. The document is served at this address. No scholarly index holds a record of it.. The Africa Evidence Network’s account of a launch sits on the network’s own site9 Source 9. Resolves. The document is served at this address. No scholarly index holds a record of it.. Each of those is real, public, and carrying weight in actual policy. None of them has a DOI, and no index holds a record of any of them.

An automated checker reports the same thing about all three that it would report about a citation invented from nothing: no record found. The machine is not being stupid. It is being precise about a question that turns out to be the wrong one.

A machine cannot distinguish a source that was fabricated from a literature that nobody bothered to index. Those two look identical from the outside, and only one of them is a problem.

This matters more here than elsewhere. Gazettes, evaluation reports, terms of reference, departmental research, and a great deal of African scholarship live outside the indexes that automated verification depends on. Put a checker in front of a policy reference list and the sources it fails hardest are the local ones. A verification standard designed without that in mind will quietly push drafters toward whatever is easiest to confirm, which means Northern, indexed, and often less relevant.

Some African scholarship does clear the bar, and the way it clears it is instructive. The standard reference on relational ethics in African evaluation is confirmable in seconds11 Source 11. Indexed. Found in Crossref, 2019., because it was published in an American journal with an identifier attached. The scholarship was never the variable.

The first version of this page carried a DOI for that evaluation paper, written out from memory while the source list was being assembled. It looked entirely plausible. It was wrong, and the checker refused it on the first run, then found the paper under a different identifier and a different year.

Nothing was published, no journalist was involved, and the whole incident lasted about four minutes. That is the entire argument for building the check into the process rather than hoping someone runs it afterwards. The tool was built to catch other people’s errors and its first catch was its author’s.

The script now reports an identifier that does not resolve as its own kind of finding rather than quietly repairing it, because a fallback that fixes the problem invisibly recreates the original failure with better manners.

Sixty-seven references, checked the way this page checks thirteen, is a few minutes of compute and an afternoon of somebody deciding what to do about the ones no machine could confirm. Set against that: a withdrawn policy, a gazetted retraction, a parliamentary statement, and a national AI strategy that has to be rebuilt while every other country moves.

The revised policy is still coming, and so is every departmental AI guideline that will be written in its wake. The organisations drafting them are about to discover that the difficult part was never the position on artificial intelligence. It is being able to show where anything in the document came from.