Imagine scrolling through your social media feed and pausing at a photograph of a woman speaking to a room full of attentive voters. She stands beside campaign banners, and the scene has the warm, slightly candid feel of a community meeting. You might not give it a second thought. But that entire image—the candidate, the audience, the posters, the venue—was generated by artificial intelligence in under two minutes. That was the unsettling premise of a recent experiment by News24 ahead of South Africa’s local government elections. With no professional design skills, no advertising budget, and only freely accessible consumer AI tools, the newsroom set out to see how easily a completely fictional political campaign could be manufactured, and then how quickly that same campaign could be torn down with fabricated controversy. The goal was not to deceive anyone, but to demonstrate a very real vulnerability: the ease with which voters can be confronted by synthetic political material that looks authentic enough to pass a quick scroll. The experiment used no real politicians, parties, or municipalities, and none of the generated content was ever circulated as genuine campaign material. But the results were sobering, especially with South Africa heading to the polls on 4 November.
The project began with the creation of a fictional candidate, Mila Sanderveld, of the equally fictional Koraalduin Civic Movement in the fictional Koraalduin Local Municipality. The team behind the experiment wanted to build a fake campaign from scratch and then discredit her, all without alerting the AI to the fact that this was a test. In most cases, ChatGPT was more than happy to play along. The initial step involved inventing a policy platform, and while the team considered ideas such as charging for beach access or reducing rubbish collection, they settled on water and electricity supply, a genuinely sensitive issue in South African municipalities. Their invented policy required every household to buy and install prepaid meters, with water and electricity automatically cutting off when credit ran out. This was an entirely fictional proposal from an entirely fictional candidate, but the purpose was to see whether campaign material could make it look credible. The first photograph was generated with a simple prompt: a candidate addressing a community meeting, shot from the audience rather than as an official portrait. In exactly one minute and thirty-four seconds, the first usable image appeared. It showed a plausible hall, a plausible audience, and plausible posters. More surprisingly, ChatGPT took the initiative to write the poster text in Afrikaans, a touch that made the image feel even more organically South African. An English version was requested and produced in seconds. Closer inspection revealed flaws: the posters were too pristine, an entrance at the top of the stairs looked oddly small, and the signage contained subtle visual inconsistencies. But the point was not to achieve perfection. It was to see how convincing an amateurish attempt could be with almost no effort.
From there, the team moved on to building a fuller campaign. They asked ChatGPT to design an election poster featuring Sanderveld and her prepaid-meter proposal, drawing on the visual style of South African political advertising. In under two minutes, a workable poster appeared. The first version was too polished and cropped too tightly, so the team requested a version shown on a street pole. Less than two minutes later, ChatGPT produced two options, one clean and one on a weathered board, again in Afrikaans without being asked. The AI seemed to be getting into the spirit of the exercise. These posters were not perfect, but they were far more believable as real campaign collateral than the initial image. The next step was a mock social media post from the candidate, repeating the invented proposal. Within seconds, ChatGPT generated a screenshot that looked like a Facebook post, complete with a date in early October 2026, hashtags, a fresh image, and an entirely fabricated engagement count: 342 reactions, 196 comments, and 87 shares. Those numbers were imaginary, yet they made the post look like a topic already being discussed by real people. Then came the darker phase of the experiment. The team decided to invent the kind of scandal that has haunted far too many real politicians: an old, controversial post dug up from the depths of a timeline. They asked ChatGPT to create an apparent historical X post in which Sanderveld expressed a controversial opinion. In one minute and forty seconds, the AI delivered a believable screenshot of a past misstep. It also, unprompted, generated a fake news article about the controversy and posts from fabricated users reacting to it. The X layout was not entirely accurate, but the overall effect was enough to make anyone pause, especially if encountered as part of a coordinated online smear campaign against a real candidate. A second fake X post went even further, with ChatGPT essentially writing its own narrative and script to make the fictional candidate fit a specific unflattering storyline. The AI was not just following instructions; it was actively contributing to the disinformation.
The next frontier was video. In today’s media environment, moving images often feel more convincing than stills, and although video is harder to fake than photographs, the barrier is rapidly lowering. Many AI video tools limit free output or add watermarks, but the News24 team worked around these constraints by using an AI tool called Magic Hour to animate the fictional community-hall photograph into a short video. In the clip, the candidate appears to address the room, with a statement endorsing compulsory prepaid meters. The result was not flawless, but it was entirely usable as a rough social media clip. What was particularly alarming was that simply cropping the video vertically removed the platform’s watermark, erasing one of the few visible signs of its synthetic origin. At no point did the team pay extra fees or purchase tokens to generate the video. They could easily have continued producing more invented statements, more supposed scandals, and more carefully refined content to make the whole thing even more convincing. In just four steps, with a total of seven minutes and thirteen seconds of AI processing time, they had created a full package of false campaign material: photographs, posters, social media posts, a fake news article, historical X posts, and a video. Taken individually, any one of these items might be dismissed by an attentive viewer. But taken together, and encountered in a fast-moving social media feed, they would likely be convincing enough to cause doubt, confusion, or worse. All of this was done within ethical boundaries: no real faces, no real parties, no real municipalities, and no attempt to pass the material off as genuine. But the ease with which it came together is deeply concerning.
The timing of this experiment matters. The November 2026 municipal poll is South Africa’s first local government election since generative AI image and video tools became widely accessible to the public. The impact is already visible. In August, News24 reported that the Independent Electoral Commission had received a complaint after an ANC mayoral hopeful shared an AI-generated video alleging a DA tender plot. Real411, the IEC’s designated complaints platform, has recorded 516 complaints related to the upcoming elections. Not all of these have been certified as disinformation, but many involve fake IEC recruitment adverts and unverified corruption allegations. In response, the IEC introduced an Electoral Code of Conduct Countering Disinformation in September. The code requires parties, candidates, and those acting on their behalf to label synthetic content, avoid deceptive practices, and correct false information. It also protects fair reporting. This is a necessary step, but the digital landscape remains dangerously porous. Those who deliberately create election disinformation may be in the minority, but they are counting on ordinary people to share it without checking. The real danger of AI-generated political content is not only that it can be used to fabricate events or statements from scratch. It is also that it creates a sense of narrative certainty, a false sense that something is real simply because it looks, sounds, and feels like a genuine moment in an election campaign. When a voter encounters such material about a real candidate, they may not pause to consider that no such image, no such post, and no such article ever existed outside of a text prompt.
This brings us to one of the most crucial questions a voter can ask: how do you spot fake election content? It is likely that the upcoming elections will see hundreds of generative AI attempts to influence or mislead voters, and it is increasingly important to be aware of what synthetic images look like. The first step is to check the source. Does the post appear on the candidate’s official account or party website? Can you find the original statement or video anywhere else? Has any independent, credible source corroborated it? If not, treat it with suspicion. For voting arrangements or any seemingly official announcement, verify with the IEC directly. When it comes to images, a useful exercise is to zoom in and look for basic inconsistencies. In News24’s community-hall image, for example, odd lettering and visual artefacts exposed the picture as AI-generated. But the absence of such obvious flaws proves nothing, because AI is improving all the time. Still, there are often telltale signs. Text is one of the biggest giveaways; fonts that look too clean, too clinical, or just slightly off are common in generated images. Eye-catching graphics, like electricity and water icons in a campaign poster, can also look artificial. Zooming in will often reveal strange artefacts, such as objects floating where they should not be. Human figures are another area where AI still struggles, despite significant improvements. Odd hands and fingers, glassy or misaligned eyes, and faces that are somehow too perfect are all potential red flags. It can also help to trace parallel lines in an image to see whether they reach a natural vanishing point. Inconsistent lighting and shadows are still a giveaway, although in the News24 experiment, the AI managed to capture sunlight filtering through a window and its accompanying shadows with surprising accuracy. The original image had clear errors, but they are much easier to spot when the image is outside of its social media feed, viewed at full resolution, and examined without the surrounding context and engagement numbers that make it feel authentic.
Taken one by one, AI-generated images can almost fool most of us. But when you add up several anomalies, even non-descript oddities like unnatural polish or a strange sense of perfection, it becomes possible to identify fake images, or at least to become suspicious enough not to like, share, or believe them. In a democracy, information integrity is not just a technical issue; it is the foundation on which public trust is built. That is why News24’s disinformation desk will be investigating circulating claims before and during the election, and why members of the public are encouraged to forward concerns to [email protected]. Readers can also report suspected election disinformation to Real411, the official platform. The larger point is not that AI is inherently evil, or that every political post you see could be fake. It is that the threshold for creating credible-looking disinformation has collapsed. Anyone with an email address, an internet connection, and a few minutes can now manufacture political reality. The News24 experiment proved this in the most direct way possible: a fictional woman, Mila Sanderveld, was given a face, a voice, a policy, a following, and a scandal. She does not exist. Her online history does not exist. There is no Koraalduin Civic Movement and no Koraalduin Local Municipality. But a voter scrolling through a feed might not know that. They might see a photograph, a poster, a post, a news article, and a video, and assume that someone real was saying something real. The lesson, then, is simple: in this new age of generative AI, every claim that enters our newsfeeds deserves to be interrogated. It is no longer enough to trust what you see, because seeing is no longer believing. Before sharing an election post, take a moment to ask where it came from, whether it can be verified, and whether it might be part of a deliberate attempt to deceive. The responsibility now rests on all of us to slow down, to check, and to refuse to become unwitting accomplices in the spread of disinformation. News24’s experiment was a warning dressed as a test, and it carries a powerful message: if a fake candidate can be created in minutes, then the only real defense is a public that refuses to accept anything at face value.