Voters Cannot Spot the Deepfake and Will Not Punish It

With South Africa’s local elections weeks away, new research suggests the real deterrent must come from the rulebook, not the ballot box.

Sign up for my Substack daily AI newsletter here.

Article link:

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

Five years ago, in a column about engineering consent, I imagined a hostile state hacking electronic voting machines to steer a foreign election. South Africa votes on paper, and that threat never arrived. What did arrive is quieter and cheaper: generative tools that can fabricate a candidate’s voice in minutes, and community WhatsApp groups where a fabrication can circulate unchecked. South Africans vote in local government elections on 4 November, less than three weeks from now. New research points to an uncomfortable conclusion for that vote. Voters struggle to spot a deepfake, and they do not punish the parties that use one. If that holds, the safeguard has to come from somewhere else.

CONTEXT AND BACKGROUND

The scale of this election makes it unusually exposed. The Electoral Commission has certified 136,790 candidates to contest 10,526 council seats across 257 municipalities, about 43% more than the 95,427 certified in 2021. Where coalitions decide who governs a metro, small margins carry large consequences. University of Johannesburg vice-chancellor Letlhokwa George Mpedi argues, in an opinion piece, that a deepfake dropped into one ward WhatsApp group shortly before polling need only shift a few hundred votes to change control of a hung council.

The Commission has responded with its first rule aimed squarely at AI. The Disinformation Code, gazetted on 4 September, requires parties and candidates to label synthetic content clearly, correct false or misleading claims within 36 hours, avoid fake accounts and bot networks, and report disinformation to the Real411 platform. Breaches carry sanctions that range from fines and campaign restrictions to cancellation of a party’s registration in serious cases.

INSIGHT AND ANALYSIS

The case for such a rule rests on two findings. The first concerns detection. In a June 2026 study by Utah Valley University’s Herbert Institute for Public Policy, 632 American participants watched either real or AI-generated videos about a fictional ballot measure. Only 16.4% correctly identified a deepfake after viewing it, the synthetic videos shifted opinions as effectively as the authentic ones, and self-described familiarity with deepfakes did not help.

The second concerns consequences. Across three preregistered studies involving more than 7,600 American respondents, researchers writing in Political Communication found that the public strongly disliked deceptive uses of AI in campaigns. Yet parties that used them suffered no significant drop in favourability among supporters, opponents or independents.

Taken together, these studies describe a failure of the ordinary democratic check. In these experiments, voters could not reliably see the deception, and when they learned of it they did not withdraw their support. According to the IEC’s latest Voter Participation Survey, the share of respondents confident in recognising fake news fell from 62% in 2023 to 52% in 2025/26, and only 18% regularly verify election information before sharing it.

I have previously written about this, arguing that deepfakes were pushing governments away from abstract ethics debates and towards concrete offences, enforcement powers and tougher penalties. Where voters cannot act as the check, the rules have to.

IMPLICATIONS

The strongest counter-argument deserves its due. Both studies used American samples under controlled conditions, not the closed, high-trust WhatsApp groups where much South African campaigning happens, and the Institute for Security Studies notes that South Africa has not yet experienced a major election crisis linked to deepfakes. Voter education helps, but the evidence suggests it cannot carry the load alone.

Enforcement has its own weakness: speed. A fabricated voice note can saturate a ward within hours, while a correction takes up to 36 hours and a sanction takes longer still. The Code is only as strong as the willingness of the IEC and the Electoral Court to act within a campaign’s compressed timeline. A penalty imposed months after the results are declared deters nobody.

The risk cuts both ways. A July 2026 brief from Morocco’s Policy Center for the New South calls synthetic disinformation in Africa a multiplier of existing tensions, and warns that laws meant to protect public debate have at times been used to criminalise satire, dissent and journalism. Enforcement must be firm without becoming a tool of censorship.

For boards, the lesson travels well beyond elections. An organisation that relies on staff to spot a cloned voice or fabricated video relies on the same weak control these studies expose. Verification and consequences have to be designed into process, through call-back rules, approval thresholds and clear accountability, rather than left to training alone.

CLOSING TAKEAWAY

Engineering consent no longer requires a state to hack a machine. It requires a cheap tool, a closed group and an electorate that cannot tell the difference. On 4 November, South Africa will learn whether its institutions can enforce rules quickly, visibly and fairly, because the research suggests voters will not be the safeguard. Boards should draw the same conclusion: when detection fails and consequences are absent, governance is the only control that remains.

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