Policy gaps, access cuts, and who is accountable

South Africa's draft national AI policy has been officially retracted, according to ITWeb. The withdrawal pauses the country's formal process for setting rules and direction around artificial intelligence, leaving businesses, civil society, and government bodies without a settled policy framework to work from or respond to.

The United States government has ordered Anthropic, the company behind the Claude family of AI models, to block all foreign nationals from accessing its two newest models, Fable 5 and Mythos 5, citing national security concerns. The directive, reported by South African technology outlet Hypertext, affects South African users, developers, and businesses that rely on these tools, with no timeline given for when or whether access will be restored. Anthropic has said the government has so far provided only verbal evidence of a security vulnerability, and that the same vulnerability appears to exist in other widely available models, including those from OpenAI.

A court has ruled that Google can be held legally responsible for inaccurate information produced by its AI-generated search summaries, according to TechCentral. Google says it will appeal the decision, arguing that errors are isolated and that most summaries are accurate. The ruling is significant for South Africa because it tests a question that local courts and regulators will eventually face: who is accountable when an AI system gives a user wrong or harmful information.

Business Day's podcast features Shane Cooper, head of digital advisory at Forvis Mazars South Africa, discussing what company boards should be examining when it comes to AI oversight in 2026. The conversation centres on governance: how boards can test whether their organisations are using AI responsibly and managing the risks it introduces. For South African companies, board-level accountability for AI is a growing area of practical concern as the technology becomes more embedded in business decisions.

Visa has announced a partnership with OpenAI to connect its global payment network to ChatGPT, allowing AI agents (software that can act on a user's instructions, not just answer questions) to help complete purchases directly within the chatbot. Memeburn reports that users would be able to link a Visa card to ChatGPT and set controls such as spending limits and merchant restrictions, with Visa handling payment security through tokenisation (a method that replaces card details with a secure code, similar to how mobile wallets work). For South African consumers, the outlet notes that while the technology could simplify tasks like booking flights or reordering groceries, it raises real questions about consumer protection, fraud liability, and the risk to people managing tight budgets if an AI agent misreads an instruction or triggers an unexpected recurring payment.

AI tools are already in South African classrooms, both through officially approved school platforms and through informal use by learners and teachers, according to IT News Africa. The publication argues that the pressing challenge is no longer whether AI belongs in schools but how to move past broad enthusiasm toward practical guidance: clear implementation frameworks and consistent support that let schools use these tools effectively. For South Africa's youth, the piece suggests, the quality of that policy and institutional response will shape whether AI in education delivers real benefit.