REG-2026-04-26 · 26 APRIL 2026 Withdrawn

South Africa — national AI policy withdrawn over fabricated citations

A gazetted national policy was withdrawn sixteen days later. At least six of its 67 academic citations referred to research that does not exist.

Primary source
Government act · 26 April 2026
Withdrawal of the Draft National AI Policy by the Minister of Communications and Digital Technologies, Solly Malatsi

The policy was gazetted on 10 April 2026 and withdrawn on 26 April. The withdrawal is the official record this entry is anchored to.

Originating reporting · April 2026
News24 — analysis of the policy's academic citations

The fabricated citations were identified by News24's review, not by the department. Under the register standard, reporting that originates a finding is named, and the finding is recorded against the official act it produced. Corroborated against DLA Piper, Fasken, CNBC Africa, Semafor and the ANC Parliamentary Caucus statement.

What happened

Sixteen days from gazette to withdrawal.

South Africa's Draft National AI Policy was gazetted on 10 April 2026. On 26 April it was withdrawn by Communications Minister Solly Malatsi, following reporting by News24 which found that at least six of the document's 67 academic citations were fabricated.

The fabrications were of two kinds: references to journals that do not exist, and references crediting real, named authors with research they never wrote. AI-generated citations had passed unchecked through multiple layers of departmental review before publication.

Gazetted

10 April 2026, as the Draft National AI Policy.

Finding

At least six of 67 academic citations did not correspond to real research. Some named journals that do not exist; others attributed work to real authors who never produced it.

Withdrawn

26 April 2026, by the Minister of Communications and Digital Technologies.

Status of revision

A revised policy is reported to be targeted for January 2027. That is a target, not a publication date and not an effective date, and this entry does not treat it as either.

This entry records a verification failure in a publication process. It makes no claim about intent on the part of any official, department or contributor, and none should be inferred from it.

Which control failed

No step in the process checked a citation against the thing it cited. The document passed through several layers of review, and every layer read the references. None resolved one. Reading a citation and verifying a citation are different acts, and only one of them was ever in the process.

Review was assumed rather than assigned. Where several layers of approval exist and none owns verification specifically, each layer can reasonably believe the check happened elsewhere. Diffuse review is not redundancy. It is a single point of failure with more signatures on it.

The failure was invisible in the format. A fabricated citation is indistinguishable from a real one at a glance — correct shape, plausible journal, credible author. Where output is assessed on whether it looks right, generated text will pass, because looking right is the specific thing it is good at. Plausibility was doing the work that checking should have done.

Detection came from outside the institution. A newspaper checked what the department had not. By the time the finding existed, it was public, and the only remaining control was withdrawal.

Zevana exposure analogue

Not recorded for this entry.

The exposure analogue describes systems in operation — what a deployed system can reach, decide or affect. This entry concerns a published document, where the failure sits in a review process rather than in a running system. No analogue is assigned rather than inferred, and the register records the absence instead of estimating a value.

The applicable control question is not what the system could do, but what the organization published without checking.

Practice addressed
Atlas

Governance & policy

Ownership, decision rights, intake and approval, policy framework, accountability mapping. For AI-assisted work products specifically: a named owner for verification, a step that resolves sources rather than reading them, and an approval that cannot be given without it.

What rules and responsibilities govern our AI?
Ask your organization
One question, this week

What in our published work has never been checked against a primary source?

Take one document already out in the world and resolve three of its references end to end. The exercise takes an afternoon, and the answer is the control, not the reassurance.

← All register entries
Begin with visibility

Ask it of your own systems before someone else does.

A short conversation about what your organization is running, who owns it, and what evidence exists. Scope and fee are confirmed after that conversation — never before.