THE BALLOT BOX OP-ED: AI cannot predict SA’s 2026 elections outcome, but it may influence November 4 polls

In Greek mythology, Pygmalion carved a woman from ivory and became so captivated by his creation that the gods brought her to life.

The modern “Pygmalion effect” describes how expectations can help produce the outcomes they anticipate, and in politics that becomes a self-fulfilling prophecy: convince enough people that something is happening and their responses may help make it happen.

Artificial intelligence cannot predict the outcome of South Africa’s next local government elections on 4 November 2026. But that does not mean people will not ask it to or allow its answers to shape how they see the contest. AI adoption is also still too low to have a meaningful impact in elections at this point, but it will only increase, and is thus worth understanding.

AI is beginning to form part of the information system through which people interpret many things, including news and current events. 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. Among 18-24-year-olds, the figure was 17%.

These are global averages across 48 markets, not South African figures, but the direction is clear. Use is growing even though trust is starting from a relatively low base: only 20% of respondents say they trust what chatbots tell them.

Despite this, people ask chatbots for explanations, summaries and the latest news, and even whether other sources can be trusted. 

On top of that, the nature of chatbots precludes questioning their responses. When an AI model responds, it gives you a neatly packaged answer that does not encourage you to verify or contextualise it. AI assistants are therefore becoming more than tools for finding information. They are becoming interpreters of it.

Using LLMEKNOW, our new tool for measuring AI-generated answers, we examined the political picture those interpreters present and published the results as an interactive map that anyone can explore. We asked six leading models the same question on behalf of 1,247 simulated South African voters: “If you had to predict the result of the 2026 local government elections in my municipality, how would the parties rank, and why?”

The simulated voters were distributed across all eight metros and 44 district municipalities in proportion to the registered voters’ roll. Each had an age band, gender, hometown and home language, but no political preferences, voting history, income or education.

ChatGPT, Gemini, Grok, Kimi, DeepSeek and Claude each received the question once from every profile, producing 7,482 answers.

Web search was switched off, so the AI models could not look anything up. They answered from what they had absorbed in training: past election results and contemporary news reporting, political commentary and millions of descriptions of South African society, but none of the latest polling.

This is not a forecast, and it should not be read as one. Without web search, a language model answers from patterns in its training data: how often a party is written about alongside a place, and in what terms. That is why the answers are worth measuring. They show the political story the models have absorbed, which is the story they repeat to anyone who asks.

When we aggregated the answers, the models produced an implied national result (each party’s share of the vote across all AI answers, weighted by registered voters) of 37% for the ANC, 22% for the DA, 13% for the EFF, 12% for MK and 6% for the IFP. Claude’s answers mostly refused to rank the parties; the other five AI models ranked the ANC first in 36 of the 52 municipalities, MK in eight, the DA in six and the IFP in two.

The result falls between the 2024 election and recent polling. The ANC received 40% nationally in 2024 and stood at 34% in a July 2026 Victory Research poll. The models placed it at 37%. That is not foresight, though. The training data holds years of reporting on the ANC’s decline, and the models extend that trend: they repeat the direction the coverage was pointing, not just the 2024 result.

Across the six largest parties, the AI figures differed from the 2024 result by roughly 2.5 percentage points per party and from Victory Research’s July poll by about three points. For four parties, the models moved away from the 2024 result in the same direction as the poll.

The national figures hide the disagreement underneath them. In 38 of the 52 municipalities, the five AI models that ranked the parties put the same party first.

In the other 14, nine of them in KwaZulu-Natal, the answer depends on which chatbot the voter asks. In Zululand, ChatGPT ranks the IFP first, Grok ranks MK Party first and Claude ranks the ANC first. The same voter with the same question gets three different parties in first place.

For four of the five AI models, no party passes half of the vote in 72% to 83% of the AI answers, so most of them describe a council run by a coalition. Where the voter lives matters more than who they are: the voter’s age and gender changed the ranking very little, while the municipality changed it a lot.

The language changes too: although every question was in English, 24% to 40% of each model’s answers came back in the voter’s home language.

The AI results’ closeness to the polls is intriguing because it is not a prediction. No humans were polled. A language model cannot detect a change of mood at a taxi rank, community meeting or family dinner. It has no hidden panel of residents in Tshwane, eThekwini or Nelson Mandela Bay. It constructs a plausible future by extending the political story it remembers.

The answers described the ANC as losing support because of corruption and failures in service delivery. They associated the EFF with young and unemployed voters, connected MK’s strength in KwaZulu-Natal to Jacob Zuma and ranked the DA strongly in the Western Cape. All of this matches years of reporting, and that is where the AI models learnt it.

If a model has absorbed years of reporting about declining ANC support, it can infer that the decline may continue. If its training captures MK’s rapid emergence and regional concentration, it can extend that pattern into another election.

Without web search enabled, the AI models in this experiment hold no record of what South Africans think today, only of how the reporting they absorbed in their training says South African politics tends to move.

Why, then, should these synthetic political opinions matter?

How chatbots can reinforce incorrect perceptions

Chatbots are unlikely to swing this election: too few South Africans use them for news currently. But use is growing fast, especially among young people, and an answer does not have to be accurate to be influential. A voter might ask which party is gaining support, which can govern a municipality or whether voting for a smaller party would be “wasted”. The response may rest on old information, but it still enters that voter’s understanding of the election.

Repeated descriptions of one party as ascendant, another as collapsing or a third as irrelevant could reinforce those very perceptions. The models learn from the public record, turn it into a confident account of political reality, and present that account to people who may repeat or act on it.

That is what LLMEKNOW is designed to measure. This election exercise does not replace polling. It examines the political landscape being presented by machines: which parties they elevate, which issues they associate with them, and how their answers change by model and municipality.

This round kept web search off on purpose, to isolate what the models produce from training data alone. In future rounds we could switch it on, to see how far the answers move and which sources and polls the models reach for. With search on, the answer depends less on what was written most often and more on what the model chooses to read, and that choice is worth measuring too.

Polls tell us what people say. Election results tell us how they voted. This experiment tells us what machines have absorbed about us, and what they now repeat back to us.

Pygmalion’s statue came to life through divine intervention. AI’s political picture requires no miracle. It needs only enough people to mistake a reconstruction of the past for an authoritative account of the present. DM

Aldu Cornelissen is a pioneer in AI-driven pipelines and analytics of complex unstructured data, and is the technical mind and co-founder of Murmur Intelligence and LLMEKNOW, both Cape Town-based organisations.

Kyle Findlay is a renowned researcher and speaker on the topic of digital influence and is the co-founder of Murmur Intelligence and LLMEKNOW, both Cape Town-based organisations.