Access, trust, and the cost of getting AI wrong

TechCentral argues that the shutdown of a major AI gaming platform called Fable 5 reveals a broader risk: AI services built on US-controlled infrastructure can be switched off by decisions made in Washington, leaving users elsewhere with no recourse. For South African businesses and public institutions that depend on cloud-based AI tools (software that runs on remote servers and is accessed over the internet), this points to a real question of sovereign risk – the possibility that a foreign government's policy or commercial decision could disable services South Africans rely on. The piece frames this not as a hypothetical but as a demonstrated capability.

Amazon CEO Andy Jassy reportedly raised concerns about security risks in Anthropic's most advanced AI models with senior US government officials shortly before Washington ordered Anthropic to cut off access to those models for users outside the United States, according to Reuters. The order followed government concerns that the models' built-in safety controls (rules meant to stop the system producing harmful outputs) could be bypassed, though Anthropic disputed that its systems had any universal weakness of this kind. For South African businesses and developers that rely on US-built AI platforms, Memeburn notes the episode is a practical warning: access to a widely used AI tool can be withdrawn overnight by a foreign government decision, with no local policy able to prevent it.

Business Day spoke with Celiwe Ross, group chief human capital and corporate affairs officer at Old Mutual, about how the South African financial services group is managing a workforce that combines human employees with AI tools. Old Mutual is one of the country's largest financial institutions, so its approach to integrating AI into staffing and people management is a practical signal of how major employers here are navigating that shift.

KPMG pulled a major report on artificial intelligence from its websites after researchers at GPTZero, a content-detection group, found that 40 of its 45 citations pointed to sources that do not exist, and four organisations named in the report publicly denied the case studies attributed to them, according to Memeburn. The firm acknowledged that its existing responsible-AI guidelines, which require human review of AI-generated content, were not applied before publication. For South African businesses, government departments, and professional services firms that use AI tools to produce research or client-facing work, the incident is a practical warning: AI systems can generate plausible-sounding but fabricated references, and those errors can reach publication if no one independently checks the output.

Meta CEO Mark Zuckerberg told employees the company made mistakes during its shift to an AI-centred workforce, according to Reuters, as reported by Memeburn. The company cut roughly 10% of its global staff, moved about 7,000 people into AI-related roles, and raised its annual infrastructure spending forecast to between $125 billion and $145 billion, covering data centres, chips, and the computing power needed to build and run large AI models. South African businesses introducing AI into customer service, banking, and HR face similar risks: moving people into new roles without clear training or explanation of how their work will change can erode trust, and under South Africa's data-protection law (the Protection of Personal Information Act), companies must also consider carefully how workplace data is collected and used when AI systems are involved.

Amazon founder Jeff Bezos has argued that AI will not leave the world short of workers, but short of work itself, contending that people have 'endless' things they want to do and that AI will remove the barriers stopping them from doing those things, according to TechCentral. The claim runs counter to widespread concern about job displacement, though Bezos offers it as reassurance rather than analysis backed by evidence. For South Africa, where unemployment is already high, the question of whether AI creates new demand for human effort or simply reduces the need for it carries particular weight.

Anthropic, the US company behind the Claude family of AI models, has disabled two of its most advanced models globally after the US government issued an export control order on 12 June 2026 requiring it to block access by foreign nationals, citing national security concerns. The order affects Fable 5 and Mythos 5, models above Anthropic's standard flagship tier in capability, and applies broadly enough that Anthropic pulled them for all customers rather than by location alone. For South African organisations that rely on cloud-based AI tools for work such as coding, research, or customer support, the episode illustrates a concrete risk: a policy decision in Washington can remove access to a widely used tool overnight, with no local recourse.