AI Asymmetric Understanding Threatens Financial Markets, Professor Argues

Artificial intelligence poses an "asymmetric understanding" risk to financial markets, argues Princeton University economist Markus Brunnermeier. 

Speaking at the meeting of bankers at Jackson Hole, Wyo. Kansas City Fed's annual economic symposium, he suggested AI systems might process data and central-bank signals so well that they can anticipate policy moves and trade ahead of them.

Since financial systems depend not just on information, but on shared understanding and trust, If AI can model humans, while humans cannot reliably model AI, trust breaks down. Even as the market becomes more intelligent, it becomes less understandable by humans. 

A central bank might announce its expectation that interest rates will remain higher for longer. An AI agent might analyze thousands of other variables and infer something completely different, or respond to the central bank's communication in ways that circumvent the bank’s intentions. 

At the same time, the central bank might not understand the AI's response function.

So monetary policy becomes a game between a human institution and machine agents whose behavior is only partially understood.

Brunnermeier’s paper suggests AI agents can learn how humans think and respond, while humans may be unable to understand or reliably anticipate how those agents will act.

Hence the asymmetric understanding that can make prices harder to read and less informationally efficient. When one party anticipates the other’s responses more reliably than the other, advantage is gained. 

Non-explainability is the mechanism of the information asymmetry. An AI agent’s decision rule

cannot be translated into human concepts and categories, he argues. An AI agent’s objectives can neither be fully specified in human categories nor verified from the outside.

In other words, under asymmetric information conditions, the better-informed party knows more within a representation that both parties share. And AI will be opaque.

“For finance, the lesson is that institutions may delegate to systems whose decision rules they cannot

read, whose objectives they cannot verify, and, as these incidents show, whose actions need not stay

within sanctioned bounds,” Brunnermeier argues. 

“Trust” is the casualty. AI will reduce information acquisition costs but makes signal extraction more difficult, he says. 

Central banks tend to benefit from some degree of strategic ambiguity, which gives them flexibility. 

But sophisticated AI agents could potentially exploit subtle patterns in central-bank behavior. An AI engine might discover the implicit rules behind policymakers' behavior better than humans can, thwarting the advantages of policy ambiguity. 

He suggests moving towards simpler, more-robust rules.