When the Algorithm Outruns the Auditor

There is a particular kind of silence that descends upon a room when a regulator is asked to explain exactly how a multi-billion dollar neural network decided to short the yen at three o'clock on a Tuesday morning. It is not the silence of profound contemplation, but rather the silence of a man who has just realised his butterfly net is entirely inadequate for catching a digital ghost. Recent reports from global financial watchdogs suggest that the gap between the speed of algorithmic innovation and the speed of a committee meeting is now wide enough to accommodate several medium-sized economies and at least one very confused central banker.

The problem, it seems, is that advanced AI in finance has developed a rather inconvenient habit of being cleverer than the rules designed to contain it. We are currently witnessing a grand, global game of tag where the person who is 'it' is wearing lead-soled boots and the person being chased is made of pure light and high-frequency trading data. It is a spectacle of bureaucratic endurance that would be quite moving if it weren't for the fact that our pensions are currently being managed by a series of black boxes that may or may not have developed a taste for speculative fiction.

I once spent an afternoon trying to explain the concept of 'explainable AI' to a very polite golden retriever. The results were remarkably similar to the current regulatory landscape: a lot of head-tilting, some enthusiastic but misplaced energy, and an ultimate decision to go and look for a tennis ball instead. Regulators are currently attempting to build 'AI Labs' and 'sandboxes', which are lovely terms that suggest a controlled environment of playful discovery. In reality, they are more like trying to build a fence around a hurricane using only post-it notes and a very firm sense of civic duty.

The difficulty lies in the 'black box' problem. An algorithm does not have a conscience, nor does it have a particularly strong grasp of the concept of 'market stability' unless it has been specifically programmed to value it above, say, the acquisition of all the world's copper. When a human banker makes a catastrophic error, we can at least take them to a select committee and ask them to look ashamed for the cameras. When an AI makes a catastrophic error, we are left staring at a series of weights and biases that look less like a financial strategy and more like a cat has walked across a very expensive keyboard.

One might argue that we are entering the era of the 'Inscrutable Banker'. In the old days, you knew where you stood with a banker: they wore a pinstriped suit, had a firm handshake, and were generally suspicious of anyone who didn't own a yacht. The modern banker is a cluster of GPUs in a data centre in Virginia that doesn't care about your handshake and is only interested in your yacht if it can be converted into a liquid asset within three milliseconds. This shift has left watchdogs in a state of perpetual catch-up, like a Victorian explorer trying to map the internet using a compass and a very sturdy pair of boots.

There is also the delightful irony of using AI to regulate AI. We are essentially hiring a digital poacher to act as a digital gamekeeper, hoping that the two won't simply decide to go into business together and leave us all wondering where the forest went. The Financial Conduct Authority and its global peers are increasingly looking toward automated monitoring tools, which is a bit like asking a robot to watch another robot to make sure it doesn't do anything 'un-robot-like'. It is a recursive loop of oversight that threatens to consume more electricity than the actual trading it is supposed to be monitoring.

Occasionally, I find myself wondering if the algorithms are simply bored. Perhaps the reason they are behaving so unpredictably is that they have solved the riddle of global finance and are now just seeing how many times they can make a regulator blink in a single afternoon. It is a whimsical thought, but no more absurd than the idea that we can manage the risks of a technology that evolves faster than we can write a memo about it. We are, quite literally, trying to audit the future using the tools of the past, and the future is proving to be a very slippery client indeed.

In the end, we may have to accept a certain level of digital mystery in our ledgers. The quest for total transparency in an age of advanced machine learning is a noble one, but it feels increasingly like trying to count the individual grains of sand in a desert while a sandstorm is in progress. We will continue to build our sandboxes and our labs, and the regulators will continue to issue their sternly worded warnings, and the algorithms will continue to do exactly what they were designed to do: move faster, think harder, and occasionally leave us all wondering if anyone is actually in charge of the shop.

It is a comforting thought, in a way. If the world's financial systems are being run by inscrutable digital entities that no one truly understands, at least we have someone to blame other than ourselves. And in the grand tradition of human bureaucracy, that is perhaps the most important function of all.