Synthetic Hatred: AI, Disinformation and Xenophobia in South Africa
AI makes fake evidence against foreign nationals cheap and fast, and every safeguard South Africa has is still slower.
Sign up for my Substack daily AI newsletter here.
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
South Africa knows where a rumour can lead. A story passed from phone to phone about who is to blame for crime or joblessness has, more than once, ended with shops looted, families displaced and people killed. What has changed is that the cost of a rumour has collapsed. Generative AI can now fabricate the “evidence” in minutes: an edited interview, a forged government notice, a voice note from someone who does not exist. With local elections three weeks away and anti-migrant mobilisation already intense this year, the question is not whether synthetic content will be used to inflame hatred. It already is. The question is whether platforms, police and the law can move faster than the lie.
CONTEXT AND BACKGROUND
The human cost is documented. Xenowatch data show 406 verified xenophobic incidents between 2022 and 2025, with 75 deaths and 151 incidents in 2025 alone. Between 2022 and 2024, more than 6,000 people were displaced and more than 800 shops looted.
This year the online campaign intensified. Africa Check identified four recurring patterns: recycled or AI-generated visuals, falsely attributed quotes, fake official documents and anonymous WhatsApp voice notes, including one falsely claiming that foreign-owned spaza shops were poisoning children’s food. Xenowatch verified 23 incidents in the first half of 2026, 14 of them violent. During protests in late June, fact-checkers found a forged notice offering a R250 bounty for reporting “illegal foreign nationals”, and a genuine broadcast interview edited with AI tools to put hostile words in a businesswoman’s mouth. Of the four claims checked, only that one involved AI; the others were recycled videos and the forgery
INSIGHT AND ANALYSIS
The striking finding is how ordinary most of this content is. As those checks show, recycled footage from years ago and crude forgeries still outnumber deepfakes. AI matters because it removes the last barriers of cost and skill, so that a convincing fake can be produced by anyone and timed to any tension. Speed is the weapon, not sophistication. Inside encrypted WhatsApp groups, platforms cannot see or remove a voice note, so any correction has to travel back through the same trusted chats, usually long after the rumour has done its work.
The amplification is organised. ISS consultant Karen Allen reports about 3.6 million mentions of the anti-immigration campaign since January, AI-generated videos circulating on TikTok, and about 15 Facebook and X accounts doing much of the boosting. False claims even dragged in a national retailer, which had to distance itself from a fabricated jobs story, showing that businesses can be pulled into the campaign without warning. The response has been thin: civil society groups say requests to platforms for takedowns produced only a handful of actions, and one group “heard nothing” back.
The law is slower still. The Prevention and Combating of Hate Crimes and Hate Speech Act was signed in 2024 but remains inoperative while its regulations are finalised, and Human Sciences Research Council surveys cited by Lawyers for Human Rights find hostility towards immigrants at its highest since 2003.
I have previously written about this in the context of fraud, arguing that the best defence against a convincing fake is verification through a channel the forger cannot reach (AI for Business: https://www.aiforbusiness.net/post/the-cheapest-defence-against-a-deepfake-is-a-question-it-cannot-answer). Communities need the same habit, because in social cohesion the cost of being fooled is measured in lives.
IMPLICATIONS
The strongest counter-argument is that blaming AI or bots distracts from real grievances about unemployment, crime and housing, and that aggressive takedowns or hate speech prosecutions could chill legitimate debate about immigration policy, especially before an election. Both concerns are valid. The answer is to target fabrication and incitement, not opinion.
That requires speed. Platforms should treat South Africa as a high-risk market, with faster escalation for verified civil society flaggers and labels on synthetic content. The Electoral Commission’s digital campaign rules, as drafted in June, require parties to label AI-generated material and report suspected disinformation within 72 hours, and they should be enforced visibly when xenophobic fakes surface. Government should bring the hate speech law into force. Businesses should monitor for false claims made in their name and correct them quickly. Employers, community leaders and faith groups can spread a simple rule: if you cannot verify a voice note, do not forward it.
CLOSING TAKEAWAY
AI did not invent xenophobia, but it has made the fuel cheaper and the fire faster. South Africa cannot detect its way out of this problem, because fakes will always outpace forensic tools. What it can do is shorten the time between a fabrication and a credible correction, and close the gap between incitement and consequence. The next flashpoint will arrive without warning. Whether it is prevented will depend on preparations made before the voice note arrives, and on a society that refuses to let a stranger’s fabrication decide who its neighbours are.
Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net
Comments