South Africa's local government elections are fewer than two months away and at the beginning of this month the Electoral Commission (IEC) announced a new Disinformation Code aimed at protecting the elections from false and misleading information.

Political parties and candidates will have to clearly label artificial intelligence (AI)-generated content before distributing it. They must avoid fake accounts and bot networks and, if they become aware that false or misleading information has been circulated by them or on their behalf, publicly correct it within 36 hours. The code was promulgated on 4 September and forms part of the existing Electoral Code of Conduct.

If you want to understand what the IEC is preparing for, take a look at Australia.

Research published this week uncovered more than 200 Facebook pages impersonating Australian politicians. Prime Minister Anthony Albanese, opposition leader Angus Taylor, and One Nation leader Pauline Hanson are among those who have been targeted. Senator Fatima Payman was shown in an AI-generated clip pleading not to be deported to Afghanistan. Independent senator David Pocock appeared in another, saying he had joined forces with One Nation. In reality, neither person said these things.

Reset Tech, a non-profit organisation that researches technology and its impact on democracy, linked the network to an influencer business in Sri Lanka. There is no evidence that the Sri Lankan government was behind it. In fact, the motivation appears to have been at least partly money.

Facebook pays some content creators for material that attracts views and engagement, and Reset Tech found that roughly one in ten of the pages had been monetised. The same network was targeting political audiences in several other countries. In total, researchers identified 1 864 coordinated Facebook pages with 5.9 million followers. You don't necessarily have to care who wins an election to mess with the political conversation around it. Sometimes there is money to be made simply by getting people angry enough to click.

Not New

This isn't entirely new to South Africa. During the 2024 election, the independent fact-checking organisation Africa Check found AI-generated impersonations of politicians and celebrities. One showed President. Cyril Ramaphosa apparently addressing the nation and announcing a plan to deal with the energy crisis. He hadn’t. Interestingly, Africa Check found far less AI-generated election misinformation than it had expected. Most of the false information it encountered during the election had nothing to do with AI. But that may not be the case forever.

Earlier this year, IEC chairperson Mosotho Moepya warned that the 2026 local elections could see misinformation move from broad national claims to much more targeted, ward-specific deception using generative AI. He referred to expectations of a “flurry of deepfakes”. For me, the danger is not only watching a fake video and believing it is real. It’s what happens when we watch a real video and decide it must be fake.

There is a name for this. Researchers call it the “liar’s dividend”. The more we know that convincing video and audio can be artificially created, the easier it becomes for someone caught saying or doing something real to claim that the evidence was generated by AI.

Doubt

Think about that in an election. A genuine recording emerges of a politician saying something damaging. The politician says it isn't real. A few years ago, we might have asked whether it had been edited or taken out of context. Now there is another possibility: perhaps the person never said it at all. You don’t need everyone to believe that. You just need enough people to wonder.

One of the things I’ve learnt from covering conflicts is that you don’t always have to convince people that your version of events is true. Sometimes creating enough doubt about everybody else’s version does the job. That’s what worries me about where AI could take election misinformation. The IEC can require political parties to label synthetic material. It can require them to correct false information. Those are sensible rules. But how do you regulate doubt?

Democracy doesn't require voters to agree with politicians, the media, or each other. South Africans certainly don’t. But we do need some way of establishing whether something actually happened. For most of our lives, seeing something with our own eyes has counted for something. Photographs could be manipulated and videos could be edited, but there was still an assumption that if we could see and hear someone saying something, they probably said it. AI is making that assumption much harder to hold on to.

By the time South Africans vote in November, a voter could watch a video of a politician saying something and have good reason to ask whether the politician ever said it at all. We know what happens when people believe something that isn’t real. I'm not sure we have thought enough about what happens when we stop believing something that is.