Who is responsible when AI gets it wrong

The South African Revenue Service (SARS) is placing artificial intelligence at the centre of its institutional overhaul, according to reporting by Moneyweb, with a senior official quoted as saying technology must strengthen human capability rather than replace it. The move signals that AI tools are being built into how SARS administers tax collection and compliance at scale. For South Africans, the development raises practical questions about how automated systems will affect taxpayer interactions with one of the country's most consequential public institutions.

South African company boards are increasingly using AI systems to support decisions on investments, risk, and strategy, but the Companies Act has not kept pace, according to a research paper published in The Conversation Africa. Company law researcher Rehana Cassim finds that current law leaves key questions unanswered: whether directors meet their legal duties when they rely on AI-generated advice, how far they can delegate decisions to automated systems, and who bears legal responsibility when an AI recommendation causes harm. Cassim argues that the Companies Act should be updated to require human oversight of AI, a minimum level of AI literacy among directors, and a designated officer responsible for the ethical use of these systems within companies.

President Cyril Ramaphosa has cautioned against a reckless approach to artificial intelligence (AI), according to ITWeb. The statement signals that South Africa's head of state is engaging publicly with how the country should adopt and govern AI, a question with direct consequences for policy, business, and citizens. The full detail of his remarks was not available in the provided text, so the specific measures or context he outlined should be confirmed from the source before publication.

Business Day's editorial board argues that South Africa needs an AI governance framework as a matter of urgency, citing growing international concern about the risks the technology poses. The piece notes that prominent global voices, including Pope Leo and Barack Obama, have warned of serious dangers, and uses this as context for why a local policy response can no longer wait. For South African readers, the editorial signals that the absence of a clear national framework is itself a policy choice with consequences.

Writing in Business Day, Tamsin MacKay argues that as new rules governing AI-generated content take effect, human oversight is becoming the practical safeguard that determines whether those rules hold in practice. The piece frames the development around South African firms, which now face growing pressure to put accountable review processes in place as regulators expect companies to be able to explain and stand behind content their AI systems produce. For local businesses using AI to generate text, images, or decisions at scale, the implication is that compliance is less a technical checkbox and more an organisational question about who is responsible when something goes wrong.

Lesotho's government is planning a R100-billion deal combining AI infrastructure (the computing facilities needed to run AI systems at scale) and hydropower development, according to Business Day. The publication reports that Public Works Minister Matjato Moteane holds a shareholding in a local firm the government intends to invite into the deal, raising conflict-of-interest questions. Given Lesotho's longstanding role supplying water and power to South Africa, a project of this scale could have material implications for regional energy and computing capacity.

Business Day columnist Tom Eaton examines whether the AI sector is a financial bubble or something more deliberately misleading, pointing to last week's market pullback and its timing as grounds for suspicion. The question has relevance for South African investors and institutions with exposure to AI-sector assets, even though the piece is opinion rather than reported analysis.

South African retailers face a new challenge in online commerce as AI-powered shopping agents (software that can search, compare, and complete purchases on a shopper's behalf) increasingly bypass traditional search and brand websites to find products, according to an opinion piece by Salesforce's Africa country manager Linda Saunders, published by Hypertext. Saunders cites Salesforce's State of Ecommerce Report, which found that traditional search traffic to brand-owned channels fell 15% while newer discovery channels, including AI assistants, grew 38% over the same period. For South African retailers, the practical implication is that product data, including pricing, inventory, and descriptions, needs to be accurate and well-structured, because an AI agent works from back-end data feeds rather than browsing a page the way a person does, and a wrong recommendation quietly teaches the agent to stop suggesting that retailer.

TechCentral publishes a commentary by Simone Santana arguing that AI agents (software that can take actions on a user's behalf, such as sending emails or querying databases) should be treated as privileged users in enterprise systems, because the access they hold is what makes prompt injection attacks (where a malicious instruction tricks the agent into acting against its owner's interests) dangerous. The piece is relevant to South African organisations that are deploying AI agents with access to internal systems and credentials, where the same security principles apply.

Writing in Business Day, South African commentator Heath Muchena argues that financial markets are misreading where the real constraint in AI development lies, comparing the technology's strategic importance to oil. The piece is relevant to South African investors and business leaders weighing where value in the AI sector will actually accrue, and which assumptions about AI growth may not hold.

Cape Town-based startup support organisation 22 On Sloane has launched KUMii, a platform that uses AI-powered matching (software that automatically connects users to relevant opportunities based on their profiles) to help African entrepreneurs find funding, customers, mentors and business tools across the continent's fragmented support landscape. The organisation also announced plans to raise R1 billion through its Sloane Capital fund to invest in startups and small businesses across Africa, according to TechCabal. For South African entrepreneurs and small businesses, the platform offers a single point of access to resources that are currently spread across disconnected programmes, funders and procurement systems.

Datatec, the Johannesburg-headquartered technology group, says it intends to position itself to benefit from growing demand for AI infrastructure (the hardware, networks, and data-centre capacity needed to run AI systems), according to Business Day. The company is also weighing structural changes as part of a broader international turnaround. As one of South Africa's largest technology businesses with operations across multiple continents, how Datatec repositions itself in this market is worth watching for what it signals about local corporate appetite for AI-related investment.

Using AI (artificial intelligence) tools to generate marketing content risks producing material that is culturally tone-deaf or offensive, according to Silas Matlala of the Gordon Institute of Business Science (GIBS), as reported by Moneyweb. Matlala warns that AI systems can produce output that comes across as insulting or disrespectful, a particular concern in South Africa's diverse linguistic and cultural landscape. For local businesses adopting AI-generated marketing, the finding points to a practical limit: automated content generation may need human oversight to avoid reputational harm.

Business Day carries a sponsored podcast in which Standard Bank describes how it is combining human judgement with machine-driven analysis to develop and release customer products more quickly. The item is labelled sponsored content, so the claims reflect the bank's own account rather than independent reporting. South African readers may find it useful as a signal of how a major retail bank is framing its use of AI internally, though the substance behind the claims is not independently verified here.

The Bill & Melinda Gates Foundation has announced a $1bn investment aimed at preventing artificial intelligence from deepening the divide between wealthy and poor populations, according to Business Day. The foundation's concern is that without deliberate intervention, AI's benefits will concentrate among those already well-resourced, leaving lower-income countries and communities further behind. For South Africa, where inequality is already acute, the commitment signals growing philanthropic attention to how AI is distributed and governed across developing economies.

Civil society groups have appealed the City of Cape Town's approval of a data centre by Equinix, a company that builds and operates large facilities housing the computer servers that power cloud computing and AI services, according to Moneyweb. The activists argue the project was approved without adequately considering its effect on nearby communities. The case raises questions about how South Africa weighs community interests when approving the physical infrastructure that underpins AI and digital services.

TechCabal profiles Samuel Segun, a Nigerian philosopher and AI safety researcher whose work is substantially rooted in South Africa: he completed his PhD at the University of Johannesburg, held a postdoctoral fellowship at Stellenbosch University, and worked at Praelexis AI, a South African machine learning company. Segun argues that the most pressing AI safety problems are not distant, science-fiction scenarios but systems already in use today, including hiring and credit algorithms that treat people unequally, and facial recognition tools that perform poorly on people of colour. His research into an Afro-ethical framework for AI, and his work developing algorithmic auditing guidelines for INTERPOL and a United Nations crime and justice institute, centre the question of whose values are built into these systems and whether African societies have the institutional knowledge to assess and govern them.

Microsoft has committed to a set of rules restricting how pupil data may be used across its education products, according to Business Day. The agreement bars the company from using that data to train AI systems (software that learns patterns from large amounts of information), to serve advertising, or to develop new products. South African schools that use Microsoft platforms are covered by the same data flows, so the commitment has direct bearing on how the personal information of local pupils is handled.

Hypertext reports from Salesforce's Dreamforce 2026 conference, where Harmony Public Schools in Texas described using AI to connect previously inaccessible records held in old or incompatible systems, a problem the company calls 'trapped data'. The school network credited this approach with a 34 percent drop in families turning down allocated school places and a 210 percent rise in application volumes. The Harmony and Salesforce teams told Hypertext that any similar effort in South Africa would need local expertise and a locally tailored solution, though they noted Salesforce's existing presence here and a philanthropic funding model that could support cash-constrained public school systems.

Anthropic, the US company behind the Claude AI assistant, says seven Chinese technology firms including Alibaba, DeepSeek and Moonshot AI secretly used nearly 190 million conversations with Claude to train their own AI systems, in some cases routing real customer requests through Claude without users knowing. The practice, known as model distillation (where a more capable model's responses are used to teach a less capable one), is not inherently unusual, but Anthropic alleges these campaigns used fraudulent accounts and proxy networks to bypass its restrictions, and that sensitive user data including names, corporate information and internal credentials was exposed in the process. Memeburn flags a direct warning for South African organisations: businesses that enter source code, financial data or customer information into any AI service are trusting not just the model they see, but the infrastructure behind it, which may route conversations to third parties without disclosure.

Several of the world's largest AI companies – Anthropic, OpenAI, Google DeepMind, and Microsoft – have publicly called for slowing the pace at which AI systems are made more capable, citing risks that could become difficult to manage or reverse. Anthropic chief executive Dario Amodei led the shift with a lengthy essay arguing that commercial competition is pushing the industry toward outcomes that are dangerous; OpenAI's Sam Altman, Google DeepMind's Demis Hassabis, and Microsoft's Satya Nadella each voiced support within hours, according to Ars Technica. If this coordinated posture holds and leads to binding standards or governance bodies, it would reshape the regulatory and commercial environment that South African businesses, institutions, and policymakers must plan around.

Microsoft has published a 37-page code of conduct for how it develops and deploys AI, stating that 'people matter more than AI' and that its AI systems are not conscious and should not be designed to appear so, according to The Verge. The document arrives as Anthropic chief executive Dario Amodei called for a coordinated slowdown in AI development, following researcher warnings that progress in AI systems may be outpacing the ability to deploy them safely and keep their actions under human control. For South Africa, where Microsoft operates and where global governance norms increasingly shape how AI is built and used locally, the commitments major technology companies make at this level set a practical floor for what responsible AI deployment looks like.

OpenAI chief executive Sam Altman has confirmed the company will not list on a public stock exchange in 2026, telling Fortune magazine that the timing is wrong given unresolved questions around AI safety and the need for closer cooperation between AI companies and governments, according to Memeburn. The decision coincides with Anthropic chief executive Dario Amodei publishing an essay calling on leading AI laboratories to slow the rate at which they make their most capable models more powerful, so that safety research can keep pace – a position Altman and Elon Musk both said they supported publicly. For South African businesses, universities, and developers that rely on these tools, Memeburn notes the effects would likely arrive indirectly: a genuine slowdown could mean longer gaps between major model releases and more scrutiny before increasingly autonomous systems reach users, and could also influence how regulators outside the United States eventually govern powerful AI systems imported from abroad.

A new report from the nonprofit Basel Action Network warns that electronic waste (discarded hardware such as servers, cables, and cooling equipment) generated by the global AI industry has been significantly underestimated, and could reach the equivalent of 23 million shipping containers by 2050 once all data centre infrastructure is counted. Previous estimates focused narrowly on servers alone; the broader accounting reveals a much larger waste stream, according to The Verge. As South Africa expands its own data centre capacity to support growing AI workloads, the findings point to waste management and environmental governance questions that will become more pressing locally over time.