On 21 May 2026, tabling the Government Communication and Information System (GCIS) budget vote in Parliament, Minister in the Presidency Khumbudzo Ntshavheni announced that South Africa's National AI Policy 'must compel digital media platforms to disclose the AI-generated content carried on their platforms (AI content branding)' and 'prohibit the broadcasting of Fake and Misleading News about South Africa.' Alongside the rules, GCIS will build a state-run fact-checking website to 'proactively debunk deepfakes and tactical misinformation' and 'position GCIS as the anchor of the truth' (SAnews). The timing is not incidental: South Africa votes in local-government elections on 4 November 2026.
The case for doing something
The threat is real, and the government is not inventing it. Synthetic media is cheap, convincing, and arrives faster than newsroom corrections. Voters in an election year deserve some defence against fabricated clips of candidates 'announcing' withdrawals or 'confessing' to crimes. Disclosure-of-provenance rules — labelling content as machine-generated — are among the least intrusive interventions available: they add information rather than removing speech, and they mirror the EU's transparency-first approach under Article 50 of the EU AI Act, which requires that AI-generated or manipulated content be marked as such. A narrow, technology-standards-based labelling mandate is defensible.
The problem is that the South African proposal is not that narrow — and the institution proposing to adjudicate 'truth' has just demonstrated, in public, why governments make poor arbiters of it.
The hallucination problem the state can't outrun
Less than a month before Ntshavheni's announcement, the very policy she invoked collapsed under its own credibility failure. In April 2026, researcher Tyronne McCrindle of Article One identified at least six fabricated academic citations — including a fictitious journal — in the draft National AI Policy Framework, sources experts attributed to AI 'hallucinations' generated by a large language model and inserted without verification. On 27 April 2026, Communications Minister Solly Malatsi withdrew the draft, conceding: 'The most plausible explanation is that AI-generated citations were included without proper verification. This should not have happened' (GroundUp).
That episode is not a footnote; it is the heart of the matter. A department that could not keep fabricated sources out of its own foundational document now proposes to operate the national authority on what is fabricated. The asymmetry is glaring. When a platform mislabels content, users can appeal, fork to a competitor, or be corrected by rival outlets. When the governing party's communications arm declares itself 'the anchor of the truth,' there is no competing anchor — and the people deciding what counts as 'Fake and Misleading News about South Africa' are the same people whose performance is being judged at the ballot box five months later.
'Fake news about South Africa' is the wrong legal object
Prohibiting 'fake and misleading news about South Africa' is a category, not a crime. South African law already addresses the genuine harms of synthetic media with precision. The Cybercrimes Act 19 of 2020 criminalises non-consensual intimate imagery, explicitly including simulated images. The Electoral Act 73 of 1998 already restricts the publication of false information intended to influence an election. POPIA provides damages for reputational harm from unlawful data processing, and the Films and Publications Act and Protection from Harassment Act cover further ground (Cliffe Dekker Hofmeyr). These statutes target conduct and harm — defamation, fraud, intimate-image abuse, election interference — not the open-ended offence of being wrong about the country.
Section 16 of the Constitution protects freedom of expression and carves out only narrow exceptions (incitement, hate speech, propaganda for war). False or misleading speech is not, by itself, outside its protection. A prohibition on 'misleading news about South Africa,' enforced by a government website, sits uncomfortably close to the kind of content-based, viewpoint-sensitive restriction the Constitutional Court has repeatedly treated with suspicion.
A co-regulatory model already exists
South Africa does not need to invent a Ministry of Truth, because it already built something better. For the 2026 elections, the Electoral Commission (IEC) is relying on Real411 — a co-regulatory complaints platform run with civil society — and has committed to issuing a draft Code of Conduct on Misinformation for public comment to 'embed safeguards against fraudulent and manipulated content' (SAnews/IEC). This is the proportionate model: an arms-length body, civil-society participation, a published code, and complaints adjudicated against transparent criteria rather than a single ministry's say-so. Stacking a GCIS 'anchor of truth' on top of it is redundant at best and, at worst, politicises a function the IEC has deliberately kept independent.
What proportionate looks like
The pro-innovation path is not to do nothing. It is to legislate the standard, not the verdict: require provenance metadata and clear AI-content labelling through a technology-neutral rule with safe harbours, leave adjudication of specific election-period falsehoods to the independent IEC and the courts under existing statutes, and fund media literacy and the co-regulatory Real411 channel rather than a partisan rebuttal site. Disclosure expands what voters know; a state truth-portal narrows who gets to decide. After an AI policy that hallucinated its own evidence, South Africa should be especially wary of handing the last word on reality to the government that wrote it.