The Bank for International Settlements (BIS) has warned that AI should not take over from central banks when making important decisions such as interest rates.

Hyun Song Shin, the BIS’s head of research and top economic adviser, said policymakers should not view AI as "something magical".

The report by the BIS discussed how AI can be used to help central banks predict inflation and other economic variables by sifting through data to identify potential financial system vulnerabilities, including the risk of future recessions.

But the report also pointed towards how some AI models sometimes struggle to explain their decisions and also have a tendency to sometimes “hallucinate” data.

The report also touched on how becoming dependent on a relatively small number of AI model providers, due to the high costs of developing these advanced models, may create third-party dependency risks involving tech partners.

It also discussed how cyberattacks could be used to attack these AI models, explaining how for example a “data poisoning” attack could be used to maliciously tamper with the data a central bank AI model is trained with.

The BIS also pointed out how central banks leaning too heavily on AI could also be at risk of making biased decisions against certain groups if the models are trained based on biased information.

The BIS said its Innovation Hub is testing AI's capabilities in several areas together with its central bank partners.

“Central banks were early adopters of machine learning and are therefore well positioned to make the most of AI's ability to impose structure on vast troves of unstructured data,” Cecilia Skingsley, head of the BIS Innovation Hub. “For example, Project Aurora explores how to detect money laundering activities from payments data and Project Raven uses AI to enhance cyber resilience, to mention just two from our portfolio.”

---