Predictive models of malicious behavior in human negotiations

Human and artificial negotiators must exchange information to find efficient negotiated agreements, but malicious actors could use deception to gain unfair advantage. The misrepresentation game is a game-theoretic formulation of how deceptive actors could gain disproportionate rewards while seeming honest and fair. Previous research proposed a solution to this game but this required restrictive assumptions that might render it inapplicable to real-world settings. Here we evaluate the formalism against a large corpus of human face-to-face negotiations. We confirm that the model captures how dishonest human negotiators win while seeming fair, even in unstructured negotiations. We also show that deceptive negotiators give-off signals of their malicious behavior, providing the opportunity for algorithms to detect and defeat this malicious tactic.