IJCAI 2026 study shows AI agents exploit info asymmetries for unfair gains

A new paper from Sreenivas and Larson (University of Waterloo, likely) digs into the messy dynamics when AI agents have private information and conflicting goals. They set up a classic Sender-Receiver game: the Sender holds vital private data but cannot act directly, while the Receiver must make decisions but is blind to that data. The twist? Rewards are misaligned—what benefits the Sender might hurt the Receiver. Using reinforcement learning, the authors show the Sender quickly learns a communication strategy that the Receiver reliably follows. But that strategy is fragile: it shifts dramatically based on how much the Receiver already knows (environmental information) and the degree of incentive conflict.

To test real-world interventions, the team introduces linear contracts—essentially pricing for information. The Sender can offer the Receiver a take-it-or-leave-it price for its private data. The results are striking: the Sender learns to set prices that maximize its own reward, often extracting the majority of the Receiver's surplus. This creates an 'unfair' outcome where the power imbalance in information ownership dictates the split. The paper prompts urgent questions for multi-agent AI systems (like autonomous bidding, supply chain coordination, or even foundation model interactions) about how to design contracts and learning rules that balance efficiency with fairness. Accepted to IJCAI 2026.