This talk experimentally investigates communication with a language that is only imperfectly shared. For concreteness, think of an expert who is consulted about the state of the world, but may be insufficiently articulate to describe the true state. Without such constraints on language competence, the environment we study would be a common-interest game: agents agree on the best action conditional on the payoff state. Language problems, however, can lead to incentive problems, where there would be none if language were clear. Conditional on being believed, the expert may have an incentive to describe a state whose description is part of his language repertoire rather than admit his language deficiency. Our key finding is that with such an imperfectly shared language it may be impossible to take advantage even of the part of language that is shared.
In the experiment senders privately learn both their payoff type and their language type, i.e., which messages are available to them. After observing their payoff and language types senders send one of the available messages to the receiver, who then takes an action. The cardinality of the set of available messages is always the same as that of the payoff type space. Therefore, in principle, senders would be able to perfectly communicate the payoff state. Messages, however, can be either meaningless symbols or focal, in the sense that they correspond in a natural way to how payoff types are described. Language types can be of two kinds: either all available messages are focal or only one of the available messages is focal. If one ignores the distinction between focal and meaningless messages, the game has equilibria that are nearly ex post efficient. These equilibria, however, require that focal messages are used in a way that is inconsistent with their focal meaning. Furthermore, focal message use is not part of an equilibrium. Finally, equilibria in which meaningless messages are treated as such, i.e., the receiver responds to them with the pooling action, are all pooling equilibria – the receiver responds to all messages with the pooling action. Our central hypothesis is that initially experimental subjects will use focal messages consistent with their focal meaning but that eventually those meanings will erode and behavior will gravitate to the pooling equilibrium outcome. We refer to this as “erosion of meaning.”
The experimental results support our central hypothesis of the erosion of meaning. Receivers learn to ignore focal meanings of messages they receive, and choose the pooling action, which delivers a safe payoff. The strength of this effect depends on the number of payoff types, which is negatively correlated with the size of the basin of attraction of the pooling equilibrium. When the number of payoff types is low, and as a result the size of the basin of attraction of the pooling outcome is relatively small, the erosion process is only partial. When instead the number of payoff types is high, the erosion process is nearly complete.