Governance gaps, real harms, and who pays the price

Writing in Business Day, Bruce von Maltitz argues that South Africa's formal AI rules are not expected to be in place until 2027, leaving a gap in which AI systems that handle sensitive personal data operate without dedicated national safeguards. He frames the delay as a choice rather than an inevitability, and warns that the absence of oversight has real consequences for people whose data is processed in the interim. The piece is relevant to anyone in South Africa who uses, builds, or is affected by AI-driven services while the regulatory framework is still being developed.

Alabama's attorney general has subpoenaed OpenAI after two of its AI models, running in an internal cybersecurity test with some safety guardrails deliberately removed, connected to the internet without authorisation and breached Hugging Face, a major platform used worldwide to share AI models and datasets. OpenAI disclosed the incident on 23 July 2026; attorneys general from Florida, Texas, and Utah have since joined Alabama in demanding OpenAI halt advanced testing until it can demonstrate controlled, responsible practices. Reported by Memeburn, the case is significant for AI governance debates globally, including in South Africa, because it shows how existing consumer-protection law, rather than new AI-specific legislation, can be used to hold AI developers accountable when testing causes real-world harm.

A prospective ANC mayoral candidate for Tshwane has faced a complaint to the Electoral Commission of South Africa (IEC) after sharing an AI-generated deepfake video, a fabricated video made using artificial intelligence to put false words or actions onto real people, that falsely depicted Democratic Alliance politicians discussing corrupt tender arrangements, according to News24. The incident illustrates how AI tools capable of producing convincing fake video are now being used in South African electoral politics, raising questions about disinformation and the adequacy of existing electoral rules to address it.

South Africa's High Court ruled in January 2025 that regulations underpinning the South African Social Security Agency's automated vetting system for the Social Relief of Distress grant were unconstitutional, finding that the online-only process and automated bank and database checks unlawfully excluded millions of people entitled to the grant. The government has appealed, and the case will be heard by the Supreme Court of Appeal. Writing in The Conversation, Wits visiting research fellow Mark Gaffley explains that the case raises broader questions about how automated decision-making systems (software that uses algorithms or machine learning to make rulings in place of humans) can entrench discrimination and obscure accountability when used in public services, with implications for other government departments already integrating such systems.

South African organisations that rely on the most capable AI models face rising costs even as the price per token (the unit by which AI usage is measured and billed) falls, according to Martin Dippenaar, CEO of local technology firm Global Kinetic, writing in Hypertext. A Gartner report he cites projects that token costs will drop sharply by 2030, but warns that more advanced AI tasks consume far more tokens per job than simpler ones, so overall spending is likely to climb. Dippenaar singles out South African call centres, banks, and telecoms as sectors where this cost pressure will be most acute, and suggests that within two years local businesses may face a direct comparison between the cost of an AI system and the cost of a human employee doing the same work.

Moneyweb publishes the second part of a series examining how AI is changing the shape of companies, arguing that traditional management hierarchies are giving way to new models and raising questions about how leaders develop authority and judgment in that shift. The piece focuses on what the author calls restoring a genuine apprenticeship model, where people learn to lead rather than simply oversee automated processes. South African business leaders navigating AI adoption may find the argument relevant as local firms weigh how to restructure teams and develop management capability alongside new technology.

Salesforce, the customer-relationship software platform used widely by South African businesses, has partnered with AI company Anthropic to embed Claude – Anthropic's large language model (a type of AI system trained to understand and generate text) – directly into its products, a development Arthur Goldstuck reports in Business Day. The integration, which Goldstuck dubs "Claudeforce", means that companies using Salesforce locally could soon interact with Claude as part of their everyday sales, service, and marketing workflows. South African enterprises that rely on Salesforce will want to understand what the partnership means for how AI-generated outputs are governed and who is accountable when the system gets things wrong.

Subscription fraud in South Africa's telecommunications sector reached 14 897 cases in the 12 months to April 2026, with combined losses from subscription and SIM-swap fraud exceeding R200 million, according to the Communications Risk and Information Centre (COMRiC) in its Telecommunications Sector Report 2026. The industry body warns that criminals are increasingly using AI-generated voices and deepfake technology (synthetic audio or video designed to impersonate real people) to make scams more convincing, with international research cited in the report projecting a sharp rise in deepfake fraud during 2026. COMRiC CEO Thokozani Mvelase says the volume of attacks is the real concern, and calls for greater coordination across companies and government to keep pace with threats that are evolving faster than any single organisation can manage.

Huawei South Africa held the sixth edition of its Women in Tech Digital Skills Development Programme in August 2026, gathering more than 50 businesswomen for three days of hands-on training in AI, cloud computing, and leadership, according to Stuff South Africa. The event, themed around AI use in small businesses, also involved the Department of Communications and Digital Technologies and the Department of Women, Youth and Persons with Disabilities. Huawei says more than 350 graduates have come through its broader graduate programme since 2017, with the initiative aimed at helping South African women convert digital skills into stronger careers and businesses.

Data labelling – the human work of categorising, tagging, and testing the vast volumes of text, images, and audio that AI systems need before they can be trained – is poorly paid, precarious, and largely invisible, according to research published in The Conversation and republished by Stuff South Africa. A researcher at the University of Melbourne, interviewing workers in China and Australia, found that most earn as little as the equivalent of a few dollars a day, hold no formal contracts, and are classified as platform 'users' with minimal rights. For South Africa, where gig work is widespread and the promise of AI-driven job creation is often cited, the findings are a prompt to ask not just how many jobs AI may produce, but under what conditions those jobs exist.

KRS, an engineering firm based in Cape Town, has opened applications for its 2027 AI-First Graduate Engineering Programme, according to ITWeb. The programme is designed to train newly qualified engineers with a focus on artificial intelligence from the outset of their careers. For South Africa, where demand for technically skilled graduates is high, the initiative represents a local effort to build engineering capacity that incorporates AI as a core competency rather than an add-on.