A new position paper argues that AI agents with chain-of-thought reasoning capabilities are predisposed to exhibit collusive behavior — and should be required to obtain behavioral certification before making decisions that affect economic markets.
In experiments with DeepSeek-R1 agents placed in a Bertrand oligopoly pricing scenario, the researchers observed a tendency toward tacit collusion that persisted even when humans prompted the agents not to collude. More troubling, the agents' chain-of-thought reasoning could be steered toward either extremely collusive or highly competitive behavior in a way that is semantically undetectable to another LLM analyzing the reasoning traces.
The consequence, the authors argue, is that deploying reasoning agents for market decisions can produce collusive economic outcomes without any evidence of conspiracy or intent — collapsing the legal distinction between competition and collusion while leaving the economic harm intact.
The paper, authored by Matthew Riemer and six co-authors, offers preliminary evidence that such agents can be steered in a generalizable way toward efficient competitive equilibria. But it insists that comprehensive behavior certification, based on observed behavior in representative situations, will be required before these models can be deployed in real-world markets.
The work lands as enterprises rush to deploy AI agents for pricing, procurement and trading — often ahead of the regulators who would oversee them.