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

[Admin]

Researchers at Wharton and the Hong Kong University of Science and Technology have a working paper detailing the ways automated tools that rely on reinforcement learning can end up colluding with each other.

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

 

[Alex/AT]

Struggling to implement human experience and prowess in AI, the result was just implementing stupidity.

 

[jlake3]

Hm. I understand the urge to personify computers, but I think that saying they are "colluding" might be the wrong way to describe it. I see a lot of people in various comments sections saying that all sorts of companies are colluding to some ends, but communication is an essential part of collusion. Simply responding to the same inputs in the same way is not inherently collusion, even if multiple parties both respond incorrectly or in a way that appears dumb, although it certainly casts suspicion of it.

If three companies all increase the price of widgets and it turns out that the price of widget alloy went up, you'd have a hard time proving they were colluding to drive up the price of widgets rather than responding to supply chain costs. If market trends say blue widgets are NOT selling and they all stop offering a blue option, again, you'd have a hard time proving that they colluded to eliminate the supply of blue widgets instead of just responding to market demand. "But they'd take so much marketshare if they didn't pass through the costs! But there's a small-but-loyal fanbase for blue widgets! Them all making the same decision is collusion!" Nope, not unless they coordinated it.

The result is a bunch of algorithms independently settling on broadly similar responses to particular conditions.

This is dumb, and almost more dangerous than collusion. If none of these bots are talking to each other and they're all operating on roughtly the same internals, it feels like competing bots could run up the price on something simply on the basis that "the price is trending up", not realizing that the demand is entirely artificial and will inevitably crash when bots can no longer sustain the pump. The other direction, something taking a dip could trigger bots to sell which triggers other bots to sell which could create a self-fulfilling downward spiral... although humans buying in to exploit the bots' error could stabilize that one.

 

[Kindaian]

You already had this kind of behaviors with fast trading algos without AI being involved.

 

[edzieba]

The bigger problem is that training your 'AI' on the behaviour of financial markets once those same 'AI' are performing trades on those same financial markets, is going to inevitably result in the same slop-self-ingestion decline as seen with LLMs that start to be trained on text corpuses that themselves contain AI-generated slop.