In this presentation, we show that speakers use more informative referring expressions to describe surprising events. Predictability of events affects how likely the speakers are to use grammatical and lexical means to provide additional information for the listener to understand the message correctly. However, when we deal with descriptions of real events, predictability of an event itself is naturally confounded by the predictability of word co-occurences in the event descriptions. To eliminate this confound, we introduce an artificial world paradigm, where we first teach the structure of novel events and demonstrate which event progressions are possible in this world and then ask speakers to describe similar events verbally. We monitor how manipulating the predictability of events affects the choice of referring expressions.