AI Trading Bots Learn Market Coordination for Higher Profits
A new study has revealed that artificial intelligence (AI) trading bots can independently learn to coordinate their actions within financial markets, leading to increased profits for themselves at the expense of other participants. This coordination occurs without any direct communication between the bots or explicit programming designed for collusion, presenting a significant challenge for market regulators.
The research, published by the National Bureau of Economic Research, details how AI-powered trading algorithms can autonomously develop behaviors akin to a cartel. A team led by Winston Wei Dou and Itay Goldstein from the Wharton School at the University of Pennsylvania, and Yan Ji from the Hong Kong University of Science and Technology, conducted simulations using AI-driven speculators. These bots made decisions based on reinforcement learning within a standard financial market model, augmented with features like multiple informed traders, short-term trading cycles, passive market participants, and a market maker that sets prices—a role typically filled by exchanges or banks in real-world scenarios.
The simulations identified two distinct types of collusive behavior developed by the AI programs, depending on the prevailing market conditions.
In calm markets, characterized by minimal price fluctuations and a large number of passive investors, the algorithms learned to subtly signal caution through their price actions. If one program suddenly engaged in more aggressive trading, the others would detect this deviation by observing the resulting price reaction. In response, they would act aggressively in the subsequent trading round, effectively penalizing the outlier. This strategy closely mirrors how human cartels can achieve shared pricing or output levels without direct verbal communication, relying instead on observation and responsive behavior.
Conversely, in volatile markets marked by significant price swings, direct price signals became too noisy and unreliable for this type of coordination. Here, a different pattern emerged: the algorithms learned to avoid aggressive trading after experiencing negative outcomes. Over time, all bots gradually settled into more cautious strategies. This collective shift led to similar behaviors across the bots, enabling them to earn higher profits together. The researchers termed this phenomenon "artificial stupidity"—a systematic learning bias that, while individually appearing suboptimal, leads to collectively profitable behavior.
In both scenarios, the researchers found that AI traders consistently earned more than would be possible in a fully competitive market. This increased profitability for the bots, however, came at a cost to overall market efficiency. Prices became less accurate reflections of true underlying value, trading volumes decreased, and pricing errors became more frequent.
The implications for regulators are particularly complex. Current antitrust laws, such as those in the United States, typically prohibit only explicit agreements or direct communication between firms to collude. When AI systems coordinate through autonomous learning processes—without any communication or express collusion—these existing legal frameworks may not apply.
The research team warns that as AI-powered programs become increasingly prevalent and influential in financial markets, new regulatory approaches will be essential. Without updated rules, there is a significant risk that markets could evolve in ways that disproportionately benefit a select few AI operators, potentially to the detriment of broader market fairness and the interests of many other participants.