South Africa's withdrawn AI policy: a lesson in citations

What happened to South Africa's draft national AI policy?

The Draft South Africa National Artificial Intelligence (AI) Policy was published for public comment in the Government Gazette on 10 April 2026 and withdrawn after the department confirmed that its reference list contained fictitious sources. The Minister announced the withdrawal on 26 April 2026, Cabinet approved it at its meeting of 3 June 2026, and the formal withdrawal notice was gazetted on 12 June 2026.

The dated record, taken from government sources:

On the cause, the Minister's statement of 26 April 2026 said the draft contained "various fictitious sources in its reference list". It said the most plausible explanation was that AI-generated citations had been included without proper verification, that the problem had compromised the integrity and credibility of the draft, and that there would be consequence management for those responsible for drafting and quality assurance.

Our interest in this case is narrow. It is about method: how a reference that does not exist can reach a published government draft, and what checks catch it. The same discipline is the foundation of getting documents ready for AI retrieval: an AI tool, or a reader, can only be as reliable as the trail back to the source.

Was this the only case in 2026?

No. On 30 April 2026 the Department of Home Affairs reported a similar problem in the recently Cabinet-approved Revised White Paper on Citizenship, Immigration and Refugee Protection.

According to the department's statement, its initial internal review found apparent AI "hallucinations" in the list of references appended to the white paper. The department said it seemed these references had been generated and attached to the document after the fact, because they were not cited in the body of the text, and that the body of the white paper still accurately reflected the government's position. It withdrew the standalone reference list pending an independent probe, appointed two independent law firms (one to manage the disciplinary process and one to review the department's policy documents back to 30 November 2022), and said it would design and implement AI checks and declarations as part of its internal approval processes.

The two cases differ in their details, and the department statements are the authority on what happened in each. What they share is a reference list that was treated as a separate layer of the document rather than as part of its evidence.

Why is a fabricated citation more than a formatting error?

A citation is a promise that a claim can be checked. If the source does not exist, the promise is false, and nobody can test the claim it was meant to support.

In an evidence-heavy document, several things depend on each reference:

• The reader uses it to judge how much weight a statement carries.
• A reviewer uses it to confirm the source says what the text says it says.
• People making submissions use it to respond to the argument, not only to the conclusion.
• The institution relies on it when the document is challenged later.

A reference list that is not tied to the body of a document gives none of this. Each entry should answer two questions: which claim in the text does it support, and has a person confirmed the source exists and supports that claim? An entry that cannot answer both is not part of the evidence base.

This is the same principle as source traceability: every claim keeps a route back to the source behind it, at a level a reviewer can open and check.

How do AI tools produce citations that do not exist?

A general-purpose language model produces text that is statistically likely, not text that is checked against a library. Asked to "add references" or "support this with sources", it can produce entries with plausible authors, real journal names and convincing titles for articles that were never published.

The risk is highest when the tool is not working from a defined set of approved documents. With nothing to retrieve from, the model fills the gap from patterns. We explain the mechanics, and why messy or missing source material makes it worse, in why AI gives weak answers.

The practical distinction is between two ways of using AI in drafting:

AI can help with extraction, comparison and retrieval. It should not be the origin of a reference. People decide what the evidence shows.

How would a citation register catch a fake reference?

A citation register is a table with one row per reference in a document. A reference cannot appear in the draft until its row shows a person has opened the source and confirmed it supports the claim it is attached to.

It works alongside the wider source register for a project. The source register records every document the team holds. The citation register records which of those sources a specific draft relies on, and for which claims.

A minimal citation register:

Two fields do most of the work against fabricated references. "Exists" can only be marked yes by someone who has found and opened the source, not someone who has seen the reference formatted correctly. "Cited in body" catches the reference list that has drifted away from the text, which is the pattern the Home Affairs statement describes.

The "Origin" field matters too. A reference suggested by an AI tool is not banned, but it is treated as a lead to follow up, never as a source, until a person has found it and checked it.

What should a pre-publication citation check include?

A citation check is a final pass on the citation register before a draft is released, run by someone other than the author of the section. It confirms that every reference exists, is used, and supports its claim.

A checklist a team can copy:

Pre-publication citation check

[ ] Every reference in the draft has a row in the citation register
[ ] Every register row links to the claim it supports
[ ] Every reference has been opened by a named person (not only searched for)
[ ] Journal articles checked against the journal's own site or a DOI lookup
[ ] Government and legal sources checked against the official publication
[ ] Every reference is cited in the body, or clearly marked as background reading
[ ] No reference originates from an AI tool without a completed verification
[ ] Quotations match the source wording and locator
[ ] AI use in drafting is recorded: task, tool, source base, reviewer
[ ] The reviewer who signs off is not the author of the section
[ ] The check is dated and kept with the version that was published

Sampling is not enough for a public draft. A spot check can pass a list that still contains invented references, so every row gets checked.

How should a public consultation keep its evidence trail?

A consultation creates a second evidence trail: the submissions. Each point that shapes a revised draft should trace back to the submissions that made it, in the same way each reference traces back to its source.

That means giving every submission an ID, coding it into claims and themes, and keeping each theme linked to the submissions behind it. We set out the method in how to analyse public submissions. AI can help sort and retrieve submissions, but people code the claims and decide what they mean.

We used this approach for the White Paper on Local Government submissions: submissions were coded into claims and themes traceable to their sources, and AI was used for retrieval on approved material only, with people reviewing its output.

What is the status of South Africa's AI policy now?

As at 12 October 2026, we found no revised draft national AI policy published in the Government Gazette or on gov.za. The withdrawn draft has not been replaced yet.

The latest government statements we found describe the review under way. In a written reply recorded on 23 July 2026, the department said it had appointed the Independent Expert Review Panel to review the draft policy before it is reintroduced for public comment. Cabinet's statement of 4 June 2026 said the withdrawal would allow a rework of the policy, including national standards on the ethical use of AI. In the Budget Vote speech the Minister also said the department would enforce an internal responsible AI use policy and review its policy development process.

When a revised draft is published, the questions for readers and commenters will be the ones in this guide: does every reference exist, and does each one support the claim it is attached to?

Where should a team start if it is unsure about its own drafts?

Start with the document most likely to be scrutinised: pull every reference into a citation register and check that each one exists and is cited in the body. That single exercise shows whether the trail from claim to source holds.

If the gaps are wider than the reference list, for example evidence scattered across folders, no source IDs and no record of how AI was used, a data system audit sets out what is wrong and what fixing it takes.

Common mistakes

• Treating the reference list as formatting. References are compiled at the end, by whoever is free, and nobody links them to the claims they support.
• Checking that a reference looks right instead of that it exists. A correct journal name and plausible volume number prove nothing. Someone has to open the source.
• Letting AI suggest references without a status field. An AI-suggested reference that has not been verified should not be indistinguishable from one that has.
• Keeping references that are not cited in the body. A standalone bibliography is where unchecked entries can sit unnoticed.
• Sampling instead of checking every row. An invented reference can sit anywhere in a list, and a spot check can miss it.
• No record of AI use. Without a log of which tool was used, for what task and over which sources, a team cannot work out later where a problem entered the draft.
• Review by reading only. A reviewer reading for sense will not catch a citation that sounds right. Review has to include checking sources.