We experimentally investigate communication in sender-receiver games with i
mperfect incentive alignment. We consider both a priori meaningless message
s and messages with pre-established meanings. Under four canonical incentiv
e conditions, we get communication outcomes. However, it is by no means a f
ait accompli. We observe significant deterioration and recoding of a priori
meanings, sucker behavior by receivers, and focal point and initial condit
ion effects. A conservative partial common interest (PCI) condition general
ly is a reliable, albeit coarse predictor of the form of communication. Equ
ilibrium selection criteria sometimes improve on the PCI prediction but nei
ther influentiality, ex ante efficiency, nor Farrell's neologism-proofness
refinement is reliable across all games, and equilibria are not always obta
ined. Considering comparative statics, equilibrium selection criteria are h
elpful but imperfect predictors of how equilibrium frequencies respond to i
ncentives, whereas the less ambitious PCI prediction is never rejected by t
he data. (C) 2001 Academic Press.