South Africa Regulated What Parties Say With AI. Nobody Regulated What AI Says to Voters.
The Electoral Commission’s new code binds 142,072 candidates, but not the chatbots many voters now consult before they mark a ballot.
Sign up for my Substack daily AI newsletter here.
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
South Africans vote in local government elections on 4 November, and for the first time the rules speak directly to artificial intelligence. Parties and candidates are bound by a code on disinformation that reaches into the digital world, AI included. What the rules do not reach is the voter who simply opens a chatbot and asks who to vote for, what a ward candidate stands for, or how coalitions work in their municipality. The machine answers fluently and with confidence. Nobody tested that answer, and nobody in South Africa is accountable for it.
CONTEXT AND BACKGROUND
The scale of this election makes the gap matter. A record 142,072 candidates are contesting 10,526 council seats, 43% more than in 2021, including 975 independent ward candidates. The Electoral Commission's counter-disinformation code is immediately binding and applies to the election-related communications, conduct and campaign activities of political parties and candidates, online and offline. It is a meaningful step, but it is addressed entirely to the people on the ballot, not to the systems voters increasingly consult about them.
Elsewhere, that consultation is already routine. About half of American adults use AI chatbots, and 42% of them use chatbots to search for information. Reporters found voters asking ChatGPT how to vote in the midterms, with the system declining to pick a candidate but offering to compare them. In Quebec, researchers estimate that about 40% of voters asked chatbots for election information during the recent provincial campaign.
INSIGHT AND ANALYSIS
The quality of those answers is the problem. A Brennan Center audit, which describes itself as exploratory, found that 32.5% of chatbot responses on election topics contained at least one factual error. In California, testers found Anthropic's Claude wrongly saying a Los Angeles mayoral candidate had lost a primary she in fact advanced from, ChatGPT pointing voters to the wrong ballot measures, and Gemini omitting tens of millions of dollars in campaign spending, although the companies say such errors are a small share of answers. In Hungary, ChatGPT and Gemini said they could not advise on how to vote, then ranked parties anyway, and gave different answers to identical prompts.
Local elections expose the weakness most sharply. Chatbots learn from what is published, and national parties are written about far more than an independent candidate in a small ward. The Quebec case carries a further warning. Because almost all Quebec newsrooms block chatbots from collecting their content, 39% of ChatGPT's voting advice in the final weeks of the campaign came from an error-filled hobby website built by AI. In South Africa, trust in news has already fallen to 50%, and 51% of online news users get news through social media, messaging or email, in a survey of English-speaking online users that is not nationally representative.
I have previously written about this when arguing that voters cannot spot deepfakes and will not punish the parties that use them. Chatbots are a different problem with the same lesson: voter vigilance alone will not protect the ballot.
IMPLICATIONS
The strongest counter-argument is that chatbot use for political information in South Africa is still modest, that WhatsApp groups and door-to-door canvassing matter far more in local races, and that any attempt by the state to control what AI may say about politics is a step towards censorship. That last point deserves the most weight. The answer is not for government to write chatbot answers. It is for AI companies to direct election questions to the Electoral Commission, publish how their systems handle South African election queries, and allow independent testing before campaigns begin. None of this can yet be compelled, so it starts with a voluntary agreement that the Electoral Commission is well placed to negotiate.
Boards have a nearer task. Banks, telcos, insurers and retailers run customer chatbots, and customers will ask them political questions. A brand's assistant that recommends a party, misstates a voting rule or invents a candidate's record is a reputational and regulatory risk. Before 4 November, every such system should be tested on election prompts, set to decline political advice and point users to official sources, and logged so that its answers can be reviewed.
CLOSING TAKEAWAY
South Africa has begun to govern AI in its elections, and that deserves credit. The code, however, governs only the people asking for votes. The adviser that sits in millions of pockets, answering fluently about candidates it barely knows, answers to no one here. Before the next national election, the Electoral Commission, the Information Regulator and the AI companies should agree on how chatbots handle South African election questions, and test that agreement in public. Democracy depends on informed choices, and an informed choice needs a source someone is willing to stand behind.
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