Abstract
The use of Artificial Intelligence (AI) in the financial sector spans across various use cases, which are broadly in financial crime and compliance management; customer insight and relationship management; credit risk management; and customer service and operations. From the literature, it is evident that the use of this technology is accompanied by certain risks and challenges.
The leading risks and prominent challenges involved with the application of AI in the banking sector were assessed, followed by different regulatory approaches taken by some of the global leading countries in AI. Through literature, five leading risks were discovered with the use of AI in banking, four prominent challenges, and eight different regulatory approaches that are driven by some of the AI leading countries in the world, that is China, Germany, France, and Canada. A total of 15 experts in AI in the South African financial industry were then asked to rate their agreeability to the discovered risks, challenges, and regulatory approaches through a questionnaire. Hierarchal cluster analysis and thematic analysis were then performed on the responses from the experts.
It was found that the most significant risk was AI violating civil rights, that is, profiling, identifying, and tracking people in a way that influences their conduct, thus limiting their freedom of expression. The risk of discriminatory or biased datasets being used in AI algorithms for automated decisioning was also found in the same cluster as the most significant risk, and was also rated as highly significant. The most significant challenge which experts could not agree on was ensuring outputs from AI models are explainable, verifiable, interpretable and reliable. Due to the infancy of AI in South Africa, there seemed to be no consensus from the experts regarding the challenges posed by AI. This is also shown in the significant levels of disagreeability between experts on the challenges posed by AI. It was found from the hierarchal cluster analysis that the regulatory approaches belonged to different clusters. In this instance, it was noted that the hierarchical cluster analysis was not effective in determining larger group clusters as initially expected. Moreover, the most significant regulatory approach found was to align laws, subordinate acts and technical standards to AI. This was followed by creating public awareness, civil activism, education and training on AI
3 | P a g e
as a regulation approach. The expects only found agreeability on these two approaches as the most significant out of the eight posed to them.
Through the thematic analysis performed on the open-ended questions, one expert raised a new risk that AI is also prone to various types of attacks such as poisoned and evasion attacks. The experts felt that the current regulations in South Africa were not sufficient in regulating AI in banking and that due to South Africa’s previous history of racial profiling, there were concerns that these would be reinforced through AI. Moreover, some experts suggested that there are system vulnerabilities in the South African financial sector, and many customers are not well experienced to protect themselves against these. One expert commented that there might be a need to conduct an overall impact assessment on the use of AI in South African banking.