Case study

South Africa’s Draft AI Policy: Fake Sources in the Plan for AI

In April 2026, South Africa published its draft national policy on artificial intelligence for comment. Some of the research it cited had never been published. The minister withdrew it 16 days later.

What are the labels? The Evidentiality Framework asks an AI to mark each claim it writes: (g) generated, its own work; (u) given, passed to it by someone else; or (m) checked against a named source. The marks stay on a claim while people work with it. How the labels work.

What Happened

The Cabinet, the president’s team of ministers, approved the draft for publication in March 2026. It was published for public comment on 10 April.
The news site News24 reported that at least six of its references appeared to be made up. Editors of three real journals, including the South African Journal of Philosophy, confirmed that articles credited to them had never appeared.
On 26 April the communications minister withdrew it. He said: “The most plausible explanation is that AI-generated citations were included without proper verification.” He called it a failure that “has compromised the integrity and credibility of the draft policy”, and promised “consequence management”, meaning action against those responsible.
Two officials were suspended while this is looked into; that is not a finding against them. The department’s head later said that “ChatGPT was used in as far as the editing of the document itself.”
The minister later said much of the policy’s content “had not faced significant challenge”. A panel of seven experts is reviewing it, and a revised draft is planned for public comment in January 2027.

Follow the Claim

This is an illustration of how the labels would have worked, not a test. It only holds if the conditions under “What Would Have Had to Be True” held.

Follow one reference: from a draft to a national policy

Key: as it happened, the type gets bigger as the claim sounds more certain. With labels: red (g) generated: written by the AI; green (u) given: passed on, with who said it; blue (m) checked against a named source.

As it happenedWith labels

Early 2026 · AI tool (suspected) → policy drafters

As it happenedA reference to a journal article that was never publishedLooks like research. (The exact wording wasn’t reprinted in sources we could read.)

With labels(g)An article in the South African Journal of Philosophy(/g)If an AI tool wrote it and used the labels: marked as the AI’s.

Early 2026 · Drafters → the draft policy

As it happenedListed in the policy’s referencesNow it looks like the research behind a national policy.

With labels(u)An article in the South African Journal of Philosophy(/u: AI tool, unconfirmed)Passed on with where it came from, and that nobody has opened it.

Before March 2026 · The check

As it happenedNobody looked it upThe references were never opened.

With labels(m)The cited articles never appeared(/m: journal editors, as reported by TechCentral, checked 3 Oct 2026)This check could be made at the time. We name the source that confirmed it later.

March–April 2026 · Cabinet → the public

As it happenedApproved and published as the national draftNow it’s the country’s official draft.

With labelsLeft out. The check found nothing, so the claim never reaches the finished document.Finished documents carry no labels. They only carry claims that passed the check.

The articles never appeared. The reference is described, not quoted, because the sources we could read didn’t reprint it.

How the Labels Could Have Helped

Each source shows where it came from. A suggested reference can’t pass as research someone read.
Approval waits for a check. Unchecked sources are visible before the Cabinet signs off.
The rest can stand. The minister said most of the content held up. Labels would show which parts were checked.

What Would Have Had to Be True

The four conditions every case shares: the AI tool used the labels; it labelled its own work correctly (the weakest link: in our tests, AI sometimes mislabels its own work); the label stayed on when the text was copied; and someone owned a rule that unchecked claims don’t go further. In this case:

The rule is owned by the officials who prepare policy papers for the Cabinet. And an AI tool has to have produced the references; the minister called that the most plausible explanation, not a finding.

What Already Existed

Review stages existed. The minister promised action against those responsible for “drafting and quality assurance”, so a quality step was there. The fake sources got through it.
Public comment worked. Outside readers found the problem, but only after publication.
A simpler check would have caught it: looking the articles up. Labels add one thing: they show reviewers which references nobody had checked.

What the Labels Wouldn’t Have Caught

Openness about tools. The officials didn’t say they had used AI. Labels assume people are open about their tools.
Is the policy right? Labels say where claims came from, not whether the policy is good.
The system around it. MPs asked how the department leading national AI policy had such gaps. That is a management question.