AI, funding US subsidiaries and a recipe for bank runs

Banks explore new data techniques to tackle money laundering

Artificial intelligence in tandem with human analysis seen as effective for know-your-customer

Foreign banks in US wary after funding costs rise

Following jump in Libor/OIS spread, many US entities continue borrowing from parents

EU parliament’s moratorium plan billed ‘recipe for bank run’

Proposal allows pre-resolution stay as well as one during resolution provided bank reopens in between

COMMENTARY: In the loop

EU and US regulators, in very different contexts, are pushing banks and other financial institutions to keep a human in the loop when handing over increasing power and autonomy to artificially intelligent software.

As well as in more established applications such as trading models, AI and machine learning are now coming into vogue in a number of new areas – among them anti-money laundering (AML), with automation seen as the only way to deal with the vast amounts of new data that will be gathered under the fifth EU Anti-Money Laundering Directive. It is hoped the technology can reduce soaring compliance budgets by targeting AML efforts more accurately.

But regulators are keen that AI should not be let off its leash. Last year, the US Federal Reserve warned against using machine learning alone to assess contagion risks in model networks, saying the approach lacked transparency. This year, AI-assisted AML will also be governed by the EU’s incoming General Data Protection Regulation (GDPR), which includes a ban on legally consequential decisions taken solely by automated means. This means AI can be used to flag suspicious cases, but the final decision must be taken by a human. 

Controlling the use of black box models has been worrying regulators for some time – as they become more sophisticated, the challenge will only become greater. Increased use of improved AI and machine learning software shines a spotlight on the growing gap between legal responsibility and effective authority.

For how much longer will a human be able to second-guess, fast enough, the decisions of a machine in order to make a difference? That stage has already been passed in trading software – indeed, the whole point is for the machines to act faster than the human brain can comprehend.

The market relies on automated safeguards. But when the consequences of a machine error are more difficult to spot than a price spike or flash crash, both human supervision and automated safeguards may lack impact until long after the damage has been done.

STAT OF THE WEEK

New rules on securitisations introduced by the Australian watchdog this year contributed to an added A$375 million ($284 million) in aggregate capital charges at the country’s four largest banks.

QUOTE OF THE WEEK

“Banks do start realising that certain business models – let me be a bit brutal – based on sub-standard loan underwriting and excessive forbearance will have less of a future, will be less tolerated by markets and supervisors” – Peter Grassman, EC directorate of financial services