Researchers find automated financial traders will collude with each other through a combination of 'artificial intelligence' and 'artificial stupidity'

If it looks like a duck, swims like a duck, and quacks like a duck, would it be fair to call it a duck? Or, in the case of a working paper from researchers at Wharton and the Hong Kong University of Science and Technology, how closely does the behavior of "AI-powered trading" have to resemble collusion before it would be fair for financial regulators to start treating it as such?

The working paper (PDF) is titled "AI-Powered Trading, Algorithmic Collusion, and Price Efficiency" and was published by the National Bureau of Economic Research. It effectively looks to answer the question raised above by conducting experiments with algorithmic trading agents that use reinforcement learning to determine when they should buy and sell assets based on the broader market's history, trends, and forecast.

The authors, Winston Wei Dou, Itay Goldstein, and Yan Ji, found that "AI collusion in securities trading can robustly emerge through two distinct algorithmic mechanisms: one based on price-trigger strategies, and the other driven by over-pruning bias in learning" — which they not-so-charitably labeled "AI collusion driven by 'artificial intelligence' [... and] AI collusion driven by 'artificial stupidity,'" respectively.

The paper features enough undefined proper nouns and mathematical symbols for me to caveat that I am primarily relying on the authors' introduction and conclusion here. My investment strategy has been — and look, I'm trusting y'all, so please don't share this with anyone else — to be too poor to have to worry about some hedge fund's algorithm having an impact on my net worth. (Google's algorithms, though...)

But I am familiar enough with how algorithms like this "make decisions" — if it can be called that — to be unsurprised by the paper's findings. Tools like this are meant to figure out the best way to maximize the probability of a number going up and minimize the probability of that number going down. The result is a bunch of algorithms independently settling on broadly similar responses to particular conditions.

"This highlights a fundamental insight about AI: algorithms relying solely on pattern recognition can exhibit behavior that closely resembles logical and strategic reasoning," the authors wrote, adding that the over-pruning bias they found "is not the result of specific, nonstandard algorithmic assumptions or limitations, but a generic feature of [reinforcement learning| that persists even in sophisticated settings."

The problem, they explained, is that regulators attempting to solve the "artificial intelligence" problem could exacerbate the "artificial stupidity" problem in the process. The former occasionally prompts algorithmic traders to make potentially risky moves; the latter mostly has them adopt conservative trading strategies. How does one discourage aggressiveness without further encouraging timidity?

My favorite example of this comes from a 12-year-old paper (PDF) describing a bot that was taught how to play games on the Nintendo Entertainment System. The bot was great at "Super Mario Bros." but terrible at "Tetris"—so it ultimately decided the best way to "win" at the game, which technically only ends when the player loses and therefore lacks a true win condition, was to pause the game before it lost.

What's the quickest way to avoid the appearance of collusion, regulatory scrutiny, and potential fines resulting from a bunch of algorithms making aggressive trades? Teaching 'em not to make aggressive trades, which is the same behavior encouraged by "artificial stupidity." This won't be an easy problem for companies developing these algorithmic traders or the financial regulators overseeing the market to solve.

Note that this paper doesn't prove AI collusion via artificial intelligence or stupidity is already occurring in financial markets; the findings were based on how different algorithmic traders behaved in simulated markets created as part of this research. But if it looks like a duck, swims like a duck, and quacks like a duck in a simulated pond...

Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.

Nathaniel Mott is a freelance news and features writer for Tom's Hardware US, covering breaking news, security, and the silliest aspects of the tech industry.