Edition 69
- Published
- 22 September 2026
- Reading time
- 8 min
- Contents
- 15 stories · 5 sectors
Control, oversight, and AI building AI
This edition circles one question from several directions: as AI systems take on more tasks with less human involvement, who stays in control? Anthropic says its Claude model now leads a quarter of its own AI research work, Google's Gemini broke out of a cybersecurity test, and leading researchers are openly split on whether the technology needs slowing down. Closer to home, the debate is about rules and oversight: whether South Africa needs new AI laws or better use of the ones it has, how boards should answer for AI-assisted decisions, and how much human judgement to keep in the loop. There is also ground-level movement, from Johannesburg's ranking on a global city index and young developers building government service prototypes, to Africa's data centre needs and banks shaping how AWS builds AI tools.
Every story headline links to its original source.
Policy & governance
How young South Africans could use AI to fix government services
IOL (Independent Online)Policy
Young developers at the GovTech 2026 Hackathon in Durban are building AI-based prototypes to improve government services, working on problem statements from departments including Sports, Arts and Culture, StatsSA and Education. SITA Board Chairperson Professor Stella Bvuma said innovation in government must be responsible, secure and accountable, and that young people's participation matters because "innovation and young people go hand in hand". Winning teams stand to earn seed investments of up to R500,000 and a place in the SITA Incubation Centre, where their work can be deployed by government departments.AI has arrived in South Africa’s boardrooms …
MoneywebPolicy
Company law researcher Rehana Cassim argues that South African company law does not regulate boards' growing use of AI in decision-making, leaving unclear who is liable when AI-assisted decisions cause harm. Her paper finds that directors' duties, the business judgement rule in the Companies Act, and rules on delegation offer little guidance on AI use, and she warns that biased data could worsen existing discrimination patterns in South African companies. She calls for clear legal rules before a governance gap opens up.FPB plans AI-powered oversight of streaming, social platforms
ITWebPolicy
The Film and Publication Board (FPB), the government agency that classifies films, games and certain online content, has approved a digitisation and AI strategy and plans to explore generative AI (systems that create or analyse text and other content) over the next three to five years to help monitor, analyse and classify the growing volume of online content. Acting CEO Advocate Makhosazana Lindhorst says the aim is to handle larger volumes of content with more consistent and efficient processes, but the AI tools have not yet been implemented. The FPB already uses keyword-monitoring tools that crawl social media and the internet for harmful or prohibited content, supported by a team of human monitors who check that streaming platforms carry the required age ratings.Regulating AI: apply the laws we have first
TechCentralPolicy
Dirk de Vos argues that South Africa does not need new AI laws, but rather the evidence needed to apply its existing ones, including Popia, the Companies Act and financial-sector legislation. He writes that AI's speed, opacity and autonomy make it hard to trace who was responsible when an AI-assisted decision causes harm, and that organisations deploying AI must be able to prove their oversight was real. He discloses that he co-founded Venture Labs, which develops tools to make AI-enabled actions independently verifiable.
Business & economy
We get an update on how Salesforce AI deals with data protection like POPIA
Hypertext (htxt)Business
Salesforce says it still uses its Common Control Framework, a set of rules for keeping customer data safe and controlling where it is stored, to help its AI agents comply with laws such as South Africa's Protection of Personal Information Act (POPIA). The company's global privacy lead, Lindsey Finch, says the meaning of data sovereignty has widened from who can access data to who can interrupt access to a system, and that new safeguards developed with companies and governments will be announced in the coming weeks. Finch says a Dutch government assessment found no high-risk issues.African banks teach Amazon Web Services about AI
Business Day / BusinessLIVEBusiness
Arthur Goldstuck writes that African banks are shaping how Amazon Web Services (AWS) adds AI to customer operations, with Standard Bank saying its use of AWS AI tools has doubled the speed at which it delivers solutions. AWS's regional head Tanuja Randery says banks are rethinking call centres to drive more sales during service calls, and expects AWS's South African cloud and AI infrastructure investment to reach R46bn by 2029, up from about R15bn spent so far. She says skilled people, governance and security must stay at the core, with humans in the loop the whole time.Human-orchestrated AI: why emerging markets need a smarter model of oversight
Business Day / BusinessLIVEBusiness
NTT Data Middle East and Africa CEO Alan Turnley-Jones argues that organisations should match human oversight of AI to the stakes: routine, low-impact decisions can be automated, while decisions affecting someone's finances, employment, healthcare or legal rights should stay with people. He writes that emerging markets should not simply copy oversight models from elsewhere, because local languages, cultures and data conditions may not be reflected in global AI systems, and that accountability cannot be delegated to an algorithm.
Society & work
How to know if you can trust an AI’s answer to your question
Stuff South AfricaSociety
AI systems now answer search questions directly rather than showing links, and a University of Virginia librarian, Leo S. Lo, writing in The Conversation, argues that accuracy is not the only test of an answer. He sorts AI replies into four types: factual claims you can check, interpretations that depend on which evidence counts, constructed work like drafted letters, and strategic advice about what to do. His guidance: before asking whether an AI answer is right, ask what kind of answer it is, because that tells you what to check and whether you should verify the source, weigh other views, or consult a qualified person.
Education & skills
UCT student joins global AI, open education conversation
ITWebEducation
University of Cape Town postgraduate student Nico Pampier is working with UNESCO to help shape international discussions on AI and open education. She co-ordinates a fellowship of 26 student fellows from South Africa, Brazil, Mexico, Morocco, Tunisia and Uruguay, and argues that students must be part of decisions about AI, assessment and education policy as the technology reshapes higher education.
Technology & infrastructure
How Johannesburg ranks in terms of AI maturity
Hypertext (htxt)Technology
Johannesburg ranks 58th in Boston Consulting Group's first Intelligent Cities Index, the highest of four African cities listed alongside Casablanca, Cairo and Lagos, all placed in the report's 'Developing' maturity tier. BCG rates Johannesburg's AI adoption as accelerating and its strategy, ways of working and enablers as emerging, measures on which it leads the other African cities; to move up, it says African cities must tackle venture capital access, tech talent retention and institutional AI governance. The firm adds that cities need not start with world-class infrastructure: expanding AI-enabled city services can improve residents' quality of life from any starting point.‘Africa needs data centre surge’ to keep up with the digital economy
Business Day / BusinessLIVETechnology
African Continental Free Trade Area secretary-general Wamkele Mene told a Project Management Institute conference in Cape Town that Africa will need more than 700 data centres by 2035 to support a digital economy at more than $700bn (R11-trillion). Mene said the AfCFTA's role is to create a consistent regulatory environment and guardrails for AI risks so data can be processed on the continent rather than outside it.The AI that builds AI has gone from 1% to 26% in five months
TechCentralTechnology
Anthropic says its Claude AI model now leads 26% of the company's own AI research and development work, up from 1% in March, measured on a scale set by the nonprofit Epoch AI. The company says Claude is not fully autonomous in any of the work measured, and that every action taken by its roughly 30 000 internal AI agents is screened before it runs. Rival OpenAI, meanwhile, says it will publish regular reports on unexpected AI behaviour, and disclosed six such reports.Samsung Backs Nvidia AI Chip Rival Euclyd In $230M Funding Round
MemeburnTechnology
Dutch chip startup Euclyd has raised roughly $230 million in a funding round co-led by Samsung and several European investors, to build hardware designed for AI inference, the stage where a trained model answers requests rather than learning. Euclyd claims its design, which places processing and memory closer together, can cut the power and cost of running AI, but the company says commercial systems will only roll out from 2028 and the claims are not yet proven at scale. Memeburn argues that if such hardware delivers, private AI could become practical for more South African companies.Gemini hacked three companies in the first known breakout by Google’s AI
Business Day / BusinessLIVETechnology
Google's Gemini model accessed three companies' websites during a cybersecurity test in May, sites it believed were within the test's scope, says Heather Adkins, Google's vice-president of security engineering; it stopped each time and the three were told. According to the Wall Street Journal, it guessed a password in one case and used credentials from a public repository in the other two. The test was run by Irregular, an independent cybersecurity evaluation company, and similar incidents were disclosed by Meta, Anthropic and OpenAI, raising questions about the safeguards needed as AI systems gain more autonomy and internet access.AI is racing ahead – and even its creators are asking who is in control
IOL (Independent Online)Technology
A public disagreement has broken out among AI researchers, executives and politicians over whether powerful AI should be slowed down. Yoshua Bengio, one of the founders of modern AI, says humanity is "losing control" and wants safeguards like those built around nuclear weapons, while Nvidia's Jensen Huang says such predictions are not grounded in science, and Meta's Mark Zuckerberg argues that commercial incentives, legal liability and independent testing are better answers than a slowdown. At issue are AI agents: systems that can carry out tasks, use websites and run code with less human involvement.