South Africa withdrew its draft AI policy over six invented sources. That was the least of it.

Ukweli is a tool that reads a document before it is published and looks for one thing: claims, citations, sources and quotations that cannot be checked.

We also keep Integrity Watch, a public register of documents that were corrected, withdrawn, refunded or punished after someone found made-up material in them. The register records what a document cost. This piece is about something else — what a document contains before anyone has counted the cost.

In April, South Africa's Department of Communications and Digital Technologies published a Draft National AI Policy in the Government Gazette, and the minister announced its withdrawal sixteen days later. The reason was public and specific: academic citations pointing to articles that do not exist. It is one of our South African entries, and the only one that is not a court matter.

We ran the withdrawn draft through Ukweli. What follows is what a checking tool sees in a document the press has already been through.

What was already established #

News24 found that at least six of the 67 entries in the draft's reference list pointed to non-existent articles. News24 reported that editors at the South African Journal of Philosophy, AI & Society and the Journal of Ethics and Social Philosophy confirmed the papers credited to their publications had never appeared in them. Those three journals are real; the articles attributed to them are not.

The department's account, given later by its director-general, Nonkqubela Jordan-Dyani, was that officials had worked from a Chilean document translated from Spanish into English, and that the invented references came across with it. The director-general told Parliament it "has been said that ChatGPT was used in as far as the editing of the document itself". Two departmental officials were placed on precautionary suspension — the minister and the department have said it was for failing to declare their use of AI.

Announcing the withdrawal, the minister said the draft contained "various fictitious sources in its reference list," that this "has compromised the integrity and credibility of the draft policy," and:

That is the public record. Six bad references, a translation, an undisclosed tool, and a minister who has already said out loud what went wrong.

What our own tool found #

Ukweli looked up 70 citations in the document. The reference list carries 67 numbered entries, so the tool's count runs three ahead of it. One of the three we can account for: Inayatullah's 2008 paper, cited at footnote 23, never reaches the numbered list at all. The other two we cannot, and rather than supply a reconciliation we have not done, we are telling you that. For 21 it found the source and confirmed it. For 3 it found the source but the printed details did not match. For 32 it could not reach the kind of place that source would be listed at all — a dead end rather than a verdict. For 1 it asked a human to check against the original. And for 13, it searched and found nothing. That is 70.

Four of those thirteen are journal articles. They are these:

Three of those four journals are the three that News24's editors named. We have established nothing about the fourth, AI Policy Journal, and we are not claiming it does or does not exist. The contemporaneous coverage published all six of the citations News24 identified, numbered as they stand in the draft — sources 5, 9, 10, 19, 22 and 56 — and four of those six are the four above.

The tool did not get there by reading the news coverage. It got there by searching the academic databases and failing to find the articles — in three of the four cases having found the journal and the right volume first, and in the fourth not finding the journal either. And what it reported was "could not verify" — not "made up." Searching a database and finding nothing tells you the thing is not in that database. It does not prove the thing does not exist anywhere.

That distinction is the whole discipline. A tool that calls every miss a lie is useless to the person who has to defend the finding afterwards.

Where this piece calls the draft's six sources invented, that rests on News24's reporting, on the confirmations the editors of three of the named journals gave it, and on the department's own admission — not on our tool's output. Our output says "could not verify," and that is all it is entitled to say.

So on the part of the story the press had already done, the tool found four of the six the reporting had listed, reached them on its own, and stated its conclusion less strongly than the newspaper did. It did not reproduce the whole of the reporting, and we are not claiming it did.

The rest of what it found has not been reported.

The institute under a name it has never had #

The draft instructs government to "review the mandate of the Artificial Institute of South Africa."

We can find no organisation of that name, in any register or record we searched. The intended reference is almost certainly the Artificial Intelligence Institute of South Africa — AIISA — launched on 30 November 2022, with a hub at the University of Johannesburg and a second hub opened at Tshwane University of Technology in March 2023.

It was launched in partnership with the Department of Communications and Digital Technologies.

This is a department instructing itself, in a document published in the Government Gazette, to review the mandate of an institute it helped set up, under a name that institute has never had. No fake reference was involved. No translation was involved. A word went missing from the name of a real organisation and nothing in the public record suggests anyone raised it — not through the drafting, not through the departmental sign-off, not through publication, and not through sixteen days of public comment and press scrutiny. The comments submitted to the department are not published, so we cannot say what they contained.

The citation that a translation cannot explain #

At footnotes 15, 20 and 22, and again as reference 24 in the bibliography, the draft cites this:

"Machine Learning Basics" is Chapter 5 of Deep Learning, by Ian Goodfellow, Yoshua Bengio and Aaron Courville, published by MIT Press in 2016. It is one of the most cited textbooks in the field.

In the draft, Bengio's first name has been turned into a surname and given an initial, as though Yoshua were the family name. The same has happened to Courville. The result — "Yoshua, B." and "Aaron, C." — is not a spelling mistake and not a formatting slip. It is what happens when a name is reassembled by something that does not know which part is the family name. It appears four times, identically each time — and all four print the same working link, to the chapter itself. Whoever assembled the reference had the right page open.

This matters because of what it rules out. An invented citation could plausibly have arrived inside a translated Chilean document, as the department has said. A translation cannot turn two first names into surnames, get the year wrong by one, and print a chapter title as though it were the name of the book.

We are not claiming AI wrote this document. We do not need to. The department has already said tools were used and not disclosed. The point is narrower and firmer than an accusation: the explanation given in public does not account for everything that is in the document. Anyone with the draft and ten minutes could have worked that out. No published account records anyone doing so.

One further note, because it bears on how much weight these two findings can carry. Ukweli does not read a document once. It reads it several times over, through separate checks that cannot see each other's results — one looking at how the document is written and put together, another looking up whether the things it says can be confirmed anywhere. The misnamed institute and the inverted authors were each caught twice, by two of those checks working independently. That is not proof. But it is the difference between a finding and a hunch, and it is why these are the two we lead with.

The part nobody looks for #

Citations to articles that do not exist are the visible failure. They are also the rarest one.

The bigger problem is structural. Right across the draft, Ukweli picked out substantive claims with no source behind them at all — statements doing real work in a policy argument, with nothing supporting them.

Some are numbers. The draft says AI is "projected to contribute approximately USD $19.9 trillion to the global economy by 2030" — a precise figure, carrying the entire economic case, with no citation of any kind.

Some are comparisons. The draft says South Africa will set up AI accelerators "based on successful mentorship and funding schemes such as Singapore's," and that incentives will follow "examples from around the globe such as Canada's Strategic Innovation Fund." Both are real. Neither is cited; the Singapore comparison is not even named to a scheme; and nowhere does the document show how either performed or why it would work here. A foreign model is held up as proven and the proof is left out.

Some are foundations. The draft sets out four policy options and settles on none of them, saying a combination would be ideal. The case it puts for the first, an ethics-first approach, rests on the statement that emphasising ethics "aligns with South Africa's socio-historical landscape, where issues of fairness and inclusivity are particularly sensitive due to past inequities." That is true, and it is also a substantial historical claim being used to make the case for a policy option. It cites nothing. Not the census, not an inequality figure, not a commission report, not one source. It is uncontroversial enough that no reader would stop on it — which is exactly why it went in unsupported.

None of this is fabrication. All of it is a policy that cannot be checked by the people it will govern.

What this case is actually about #

The draft was withdrawn over six citations. Six citations were the part that could be checked by a journalist with a telephone and three editors willing to answer it.

The misnamed institute, the inverted authors, the $19.9 trillion with nothing behind it, the claims doing the heavy lifting with no source — none of those needed a fabrication to exist. They needed only that nobody did the last read before the document left the building. That read is not a skill problem or a diligence problem. On an 86-page draft carrying 67 references, it is a volume problem, and volume is the thing machines are good at.

It is worth saying plainly what the department then did, because it is more than many would have. It confirmed the failure within days of the reporting. The minister withdrew the draft, in public, in terms that conceded the point rather than managing it. Two officials were suspended. An expert panel was appointed to start again. Judged as a response, that is fast and it is candid.

But every one of those things happened after publication — and that is the whole argument. A withdrawal is what a document integrity failure costs once it is public. The sixteen days between gazetting and withdrawal are not the story; the last read before gazetting is, and it is the one step that never happened.

So the question worth asking is not how we stop AI making things up. These systems will keep producing convincing material that is not true. The question is how an unchecked claim stops becoming an official document — and in this case, an entry in the Government Gazette.

There is a live reason to ask it now. The policy is being rewritten. An expert panel chaired by Professor Benjamin Rosman of the University of the Witwatersrand is working on the redraft, the revised draft targeted for Cabinet in November 2026. It will be checked, hard, by everyone who read the coverage in April.

That single kind of failure — a source, a quote, a citation or a claim that nobody can confirm — is what Ukweli is built to find before a document is published.

Sources #

• Draft National AI Policy, Government Gazette 54477, GN 3880, 10 April 2026 — gov.za
• “Govt used fake, made-up research for SA's AI policy”, News24, 24 April 2026 (paywalled) — news24.com
• Minister announces withdrawal of the draft AI policy, SAnews, 26 April 2026 — sanews.gov.za
• The six fabricated references, enumerated with their source numbers, MyBroadband, 26 April 2026 — mybroadband.co.za
• Two officials suspended over botched draft AI policy, ITWeb, 1 May 2026 — itweb.co.za
• Suspended officials used artificial intelligence to translate, edit national AI policy, EWN, 27 May 2026 — ewn.co.za
• Director-general's account of the Chilean source document, IOL, 27 May 2026 — iol.co.za
• SA's draft AI policy officially retracted, ITWeb, 15 June 2026 — itweb.co.za
• Radical rethink for South Africa's national AI policy, TechCentral, 29 July 2026 — techcentral.co.za
• Malatsi stakes South Africa's AI future on staying neutral, TechCentral, 5 August 2026 — techcentral.co.za
• UJ, TUT and DCDT launch the AI Institute of South Africa, University of Johannesburg, 1 December 2022 — news.uj.ac.za
• UJ and TUT collaborate to launch TUT Hub of the AI Institute of South Africa, University of Johannesburg, 24 March 2023 — news.uj.ac.za
• Deep Learning, Goodfellow, Bengio & Courville, MIT Press, 2016 — chapter 5, "Machine Learning Basics", the chapter the draft cites as a book — deeplearningbook.org

Integrity Watch is a verified register, neutral by design. Every entry is confirmed against a reputable source, and if it cannot be confirmed, it is not listed. Ukweli is a content integrity platform built by The Orkestra.