A narrow rule in a region with few
On 4 September 2026 the Electoral Commission of South Africa (IEC) promulgated the Electoral Code of Conduct Countering Disinformation, in terms of section 88(a) of the Local Government: Municipal Electoral Act 27 of 2000. It applies from publication until results are declared in the 4 November local government elections. The code asks parties and candidates to label AI-generated ("synthetic") content clearly before they distribute it. It also sets a 36-hour clock for reporting online disinformation and for publicly correcting false claims.
I could not find a broad deepfake statute elsewhere on the continent. This election-period code is therefore unusual: it is a binding synthetic-media obligation, and it comes from an electoral regulator rather than from a new law passed in a hurry.
The case for the rule
The strongest argument for it is simple. Synthetic audio and video are cheap to produce, and a convincing fake released days before a vote can circulate widely before anyone corrects it. A candidate who is wrongly shown saying something damaging cannot always repair that in time, and voters cannot easily un-hear it. A short election window with an identifiable set of regulated actors is the setting where a mandatory provenance label is easiest to justify.
What the text actually says
The June draft, Government Notice 3995 of 30 June 2026, shows how the design works. Clause 11(c) required parties and candidates to ensure that "any synthetic content (including AI-generated content) is appropriately white-labelled and marked as such prior to distribution". The test is contextual: it takes into account the content's nature, purpose and likely consequence, and the public interest in avoiding confusion about authenticity. The draft separately told parties not to use "deceptive content using audio, images or video ... to falsely or deceptively alter or fake another person" (clause 9(c)).
According to the IEC's own statement, parties must report instances of online disinformation to Real411 within 36 hours of becoming aware of them, and must publicly retract and correct false or misleading information within the same period. Business Day reported that the code defines synthetic content as "audio, images, video or text generated using AI".
Enforcement runs through existing machinery. The IEC statement says the code carries "the same sanctions" that apply to breaches of the Electoral Code of Conduct under the Act, and the draft likewise says a breach is "subject to sanction as contemplated under the Act". The sources I could verify do not list specific penalties, so readers should check the Act for the sanction schedule rather than rely on press summaries.
What the design gets right
Three features keep this proportionate, and each is worth copying.
- It regulates the speakers who are accountable. The duties fall on registered parties and candidates, who have reputations and legal exposure. They do not fall on anonymous users or on platforms.
- It is a labelling rule, not a ban. Satire, parody and legitimate AI-assisted campaigning remain lawful if marked. Labelling is a much lighter intervention than takedown.
- It is anchored in existing law. The code builds on section 69(2) of the Act, which already prohibits knowingly publishing false information with intent to disrupt an election or influence its outcome. The code does not create a new speech offence.
The draft also contained a free-speech safeguard (clause 14): nothing in the code limits "lawful criticism, opinion, commentary or fair reporting" about the Commission. It further barred parties from improperly labelling legitimate criticism as false (clause 6(d)). A rule that names its own abuse case is more credible than one that does not.
Where the risks sit
Proportionality will be tested in practice, and four risks stand out.
- The 36-hour clock is short. It runs from when a party "becomes aware", which is a subjective trigger. Large parties with social teams can comply, but small parties and independent candidates, who far outnumber them, may struggle. The deadline raises the risk of defensive over-reporting, with parties filing everything to stay safe.
- Labelling is only as good as the adjudication behind it. Real411 is a complaints platform with real volume: Business Day cites Media Monitoring Africa as reporting more than 285 complaints in the 2024 election. If decisions are slow or inconsistent, the label becomes a trap for honest actors and a non-event for bad ones.
- Bad actors will not label. A foreign influence operation or an anonymous account is not bound by a party code. The rule therefore mainly constrains parties that were already trying to comply, and for those actors the question is whether the label works as a norm.
- Vague tests invite selective enforcement. Terms like "appropriately" labelled and "likely consequence" need published guidance and reasoned decisions. Otherwise they can be applied unevenly against opposition voices.
What to watch after 4 November
The code lapses when results are declared, which makes it a useful experiment. A short, time-limited rule with a defined set of regulated actors gives regulators a chance to measure outcomes before anyone proposes a permanent statute. The IEC should publish how many Real411 complaints involved synthetic media, how many were upheld, how long decisions took, and whether any sanction was applied to a labelled satirical or clearly parodic item.
That data would answer the question that matters for the rest of the continent. Governments elsewhere may be tempted to answer deepfake anxiety with broad, permanent offences that sweep in satire and journalism. South Africa's approach suggests a better order: start with transparency duties on accountable political actors, build in explicit speech protections, and let the evidence on enforcement accumulate before legislating more. If the post-election review shows the labelling rule was administrable and rarely misused, it will be a stronger model than any blanket deepfake ban.