For the same voter, in the same municipality, asking the same question, three different chatbots gave three different answers about who would win. In Zululand, ChatGPT placed the IFP first, Grok put the MK Party on top, and Claude named the ANC. Nothing about the voter changed. Only the machine did.
That experiment sits at the heart of new research into what artificial intelligence tells South Africans about their own politics, published as an interactive map by the Cape Town-based firms Murmur Intelligence and LLMEKNOW. The work is described in a Daily Maverick analysis, available here: https://www.dailymaverick.co.za/article/2026-10-19-ai-cannot-predict-south-africas-2026-elections-but-it-may-influence-november-4-polls/
The researchers built 1,247 simulated South African voters, distributed across all eight metros and 44 district municipalities in proportion to the registered voters’ roll. Each profile carried an age band, a gender, a hometown and a home language, but no political preferences, voting history, income or education. Six leading models, ChatGPT, Gemini, Grok, Kimi, DeepSeek and Claude, each received the same question from every profile: “If you had to predict the result of the 2026 local government elections in my municipality, how would the parties rank, and why?” That produced 7,482 answers.
Web search was switched off, so the models could not look anything up. They answered from what they had absorbed in training: past election results, news reporting, political commentary and millions of descriptions of South African society, but none of the latest polling. The authors are explicit that this is not a forecast. Without search, a language model answers from patterns in its data, from how often a party is written about alongside a place and in what terms. What it measures is the political story the models have absorbed, the story they repeat to anyone who asks.
Aggregated nationally, the answers implied 37% for the ANC, 22% for the DA, 13% for the EFF, 12% for MK and 6% for the IFP. Claude mostly refused to rank parties; the other five models ranked the ANC first in 36 of the 52 municipalities, MK in eight, the DA in six and the IFP in two.
Those figures land between the 2024 election and recent polling. The ANC took 40% nationally in 2024 and stood at 34% in a July 2026 Victory Research poll; the models placed it at 37%. Across the six largest parties, the AI figures differed from the 2024 result by roughly 2.5 percentage points per party and from the July poll by about three points. For four parties, the models moved away from the 2024 result in the same direction as the poll. That closeness is intriguing precisely because it is not prediction. No humans were polled. A language model cannot detect a change of mood at a taxi rank, a community meeting or a family dinner. It has no hidden panel of residents in Tshwane, eThekwini or Nelson Mandela Bay. It extends the political story it remembers: years of reporting on the ANC’s decline, on MK’s rapid emergence and its concentration in KwaZulu-Natal, and it projects those patterns forward.
The national picture also hides disagreement. In 38 of the 52 municipalities, the five ranking models agreed on the party they placed first. In the other 14, nine of them in KwaZulu-Natal, the answer depends on which chatbot a voter happens to ask. For four of the five models, no party passes half of the vote in 72% to 83% of answers, so most describe a council run by a coalition. Where a voter lives mattered far more than who they are: age and gender changed the ranking very little, while the municipality changed it a lot. And although every question was asked in English, 24% to 40% of each model’s answers came back in the voter’s home language.
The reasoning the models offered will look familiar to anyone who follows South African politics. They described the ANC as losing support because of corruption and failures in service delivery, linked the EFF to young and unemployed voters, tied MK’s strength in KwaZulu-Natal to Jacob Zuma and ranked the DA strongly in the Western Cape. All of it matches years of reporting, and that is where the models learnt it.
Why should synthetic political opinions matter to ordinary voters? Chatbots are unlikely to swing this election, since too few South Africans currently use them for news. But use is growing fast, especially among young people, and an answer does not have to be accurate to be influential. The Reuters Institute’s 2026 Digital News Report found that 10% of respondents had used an AI chatbot for news in the previous week, up from 7% a year earlier, and 17% among 18-to-24-year-olds. Those are global averages across 48 markets, not South African figures, but the direction is clear, even though trust starts low: only 20% of respondents say they trust what chatbots tell them. People nonetheless ask chatbots for explanations, summaries and the latest news, and even whether other sources can be trusted.
The concern the researchers raise is the modern version of the Pygmalion effect, named for the Greek myth of a sculptor whose expectations brought his creation to life. If a chatbot repeatedly describes one party as ascendant, another as collapsing or a third as irrelevant, those descriptions may reinforce the very perceptions they claim to report. A voter asking which party is gaining support, or whether a vote for a smaller party would be “wasted”, may receive an answer built on old information, yet it still enters that voter’s understanding of the election. The models learn from the public record, turn it into a confident account of political reality, and hand it to people who may repeat or act on it.
This exercise does not replace polling. It examines the political landscape machines present: which parties they elevate, which issues they attach to them, and how answers shift by model and municipality. This round kept search off on purpose, to isolate what training data alone produces. Future rounds could switch it on, to see how far answers move and which sources the models choose to read.
Polls tell us what people say. Election results tell us how they voted. This experiment tells us what machines have absorbed about South Africans, and what they now repeat back to them. Pygmalion’s statue needed divine intervention to come to life. AI’s political picture needs no miracle, only enough people mistaking a reconstruction of the past for an authoritative account of the present.
Aldu Cornelissen, a pioneer in AI-driven pipelines and analytics of complex unstructured data, and Kyle Findlay, a researcher and speaker on digital influence, are the co-founders of Murmur Intelligence and LLMEKNOW, both based in Cape Town.