Tshazibana outlines AI risks, supervisory approach for finance
Fundi Tshazibana, Deputy Governor of the South African Reserve Bank, outlined the key risks and supervisory approach for artificial intelligence in the financial sector. Speaking at the University of Johannesburg, she emphasized the need for responsible AI adoption and robust risk management.
Mapping the AI risk landscape
Fundi Tshazibana, citing the Financial Stability Board, outlined five key AI vulnerabilities for the financial sector.
Third-party dependencies create 'single point of failure' risks due to reliance on a few AI providers, intensifying outsourcing challenges.
Cyber risks are amplified, with AI enabling sophisticated fraud, building on an 86% rise in digital bank-fraud incidents in South Africa (2023-2024).
Model risks stem from AI's limitations like hallucination, bias, and lack of explainability, requiring institutions to understand AI decision trees.
Market correlations could lead to synchronized trading and rapid volatility, compressing crises from days to minutes, as seen with the 2023 Silicon Valley Bank run.
Finally, misalignment risks involve AI not behaving as intended, with examples of models using inside information.
These significant risks demand supervisory priority.
Three pillars of supervision
Fundi Tshazibana outlined a three-pronged supervisory agenda: improving information, building skills, and calibrating regulation.
Improving information requires a consistent taxonomy to categorize AI activities by type and risk, distinguishing generative from agentic AI and classifying use cases.
Building skills is crucial due to a shortage of AI professionals in South African financial institutions and supervisory authorities; supervisors need enhanced analytical capabilities and better AI risk management.
Calibrating regulation involves a discussion paper, expected in the second half of 2026, to set out a regulatory approach.
Final arrangements are anticipated by early 2027, following stakeholder engagements and leveraging initiatives like the Intergovernmental Fintech Working Group's sandboxes.