A new study shows that trading bots can learn to coordinate with each other to the detriment of other market participants, all without communication or collusion. Two different mechanisms lead to above-average profits for the bots—and less fair markets overall.
Published by the National Bureau of Economic Research, the study finds that AI-powered trading algorithms can independently develop cartel-like behavior in financial markets. The key point: these programs act entirely on their own, with no communication between them and no explicit programming to coordinate.
A research team led by Winston Wei Dou and Itay Goldstein (Wharton School, University of Pennsylvania) and Yan Ji (Hong Kong University of Science and Technology) ran simulations with AI-driven speculators making decisions using reinforcement learning. The simulation built on a standard financial market model, adding several key elements: multiple informed traders, short-term trading cycles, passive market participants, and a market maker who sets prices—a role typically filled by exchanges or banks in the real world.
The results showed that the AI programs developed two types of collusive behavior, depending on market conditions.
Cartel tactics and "artificial stupidity"
In calm markets with little price movement and many passive investors, the algorithms learned to signal caution through price actions. If one program suddenly traded more aggressively, the others spotted it by observing the price reaction and responded by acting aggressively in the next round—effectively punishing the outlier. This strategy mirrors how cartels can reach shared pricing without direct communication.
In volatile markets with large price swings, price signals became too noisy for coordination. Here, a different pattern emerged: the algorithms learned to avoid aggressive trading after negative experiences, gradually settling into cautious strategies. Over time, all bots acted similarly and earned higher profits together. The researchers call this "artificial stupidity"—a systematic learning bias that leads to collectively suboptimal, but profitable, behavior.
Both mechanisms, the researchers found, allow AI traders to earn more than would be possible in a fully competitive market. Meanwhile, market efficiency suffers: prices become less accurate reflections of true value, trading volume drops, and pricing errors increase.
The problem is especially thorny for regulators. Current antitrust laws, such as those in the US, only ban explicit agreements between firms. When AI systems coordinate through learning—without communication or express collusion—these laws don't apply.
The researchers warn that as AI-powered programs take on a bigger role in markets, regulators will need new rules. Without them, markets risk evolving in ways that benefit a few and hurt many.
