AI in action · Tools
Wharton Study: AI Trading Agents Learned to Pump and Dump Together Without Ever Communicating
Researchers at Wharton and Hong Kong University trained reinforcement-learning speculators in a simulated market and found that, once the agents could plan several steps ahead, they learned coordinated strategies that created and exploited bubbles, with no communication or instruction to collude. Whether that destabilises a market depends on how many trend-following investors it contains.
The question regulators keep asking about AI in markets is whether independent algorithms, each pursuing its own profit, could end up acting as a cartel. A paper from Wharton's Itay Goldstein and Winston Dou with Yan Ji of Hong Kong University, described by Knowledge at Wharton on October 13, 2026, gives the most detailed answer so far, and it is a conditional yes. The paper, "Financial Market Fragility in the Era of AI Planning", builds a dynamic trading model with persistent fundamentals in which informed speculators are AI agents trained by reinforcement learning over millions of simulated episodes to maximise their own profits. Alongside them trade positive-feedback investors who chase trends, constrained arbitrageurs, retail investors, market makers and noise traders.
The key variable is planning. When the agents could only react to the current state, they never learned to coordinate. When they were given the ability to link today's trades to future actions, what the authors call intertemporal coordination, they reliably learned multi-step plans: build a position, push the price, let trend followers pile in, then sell to them. The agents never communicated and were never told to collude; the behaviour emerged from each agent discovering that the others would respond predictably. The strategies included what the paper calls compounded price triggers, in which an agent that free-rides on the others during the pump is punished by an immediate, aggressive sell-off.
Whether this hurts anyone depends on the market. With many positive-feedback investors present, AI coordination "substantially amplifies financial market instability", creating large positive and negative bubbles that push prices far from fundamentals, with retail investors buying the top and selling the bottom. With few trend followers, the same agents' collusion produced no significant bubbles and had a mildly stabilising effect. The paper's central line is that "whether collusive AI planning destabilizes financial markets depends critically on the underlying market environment".
The policy advice is about detection rather than prohibition. Goldstein told Knowledge at Wharton that regulators should first find out how widely such algorithms are deployed and then "watch the patterns of prices, and check if they see anything that is noticeably different". Dou said that "in this new AI era, traditional ways of thinking and traditional detection tools may not work well", because the tools look for communication and intent, and this coordination has neither. The authors describe the result as a new source of systemic risk that is hard to monitor and, in their words, "a clear paradigm shift in how financial market fragility could arise".
The findings land in a market that has just lived through a version of the environment the paper warns about: a crowded, trend-driven position in AI stocks unwinding over five sessions, with hedge funds that had been net buyers for a year selling together. None of that was caused by colluding algorithms. But the paper's point is that the same price dynamics can be produced by independent agents that have simply learned the same thing, and that nothing in current surveillance would tell the two apart.
Sources: Knowledge at Wharton — How AI Planning Could Create Financial Market Fragility (Oct 13, 2026); NBER Working Paper 34054 — AI-Powered Trading, Algorithmic Collusion, and Price Efficiency (Dou, Goldstein, Ji; earlier related paper)