To be learnable, words must contribute something that is pretty stable across contexts. But equally, words must also be flexible enough to be able to stretch, in a principled way, to cover new cases. Similarly, to be effective for communication, the information that words encode must be robust enough and flexible enough to help us achieve a wide variety of goals. It is argued that truth conditions, and information understood in terms of truth conditions, cannot satisfy these requirements. A replacement for the truth conditional model is suggested based on a statistically grounded conception of semantic information. Informally, this can be understood in terms of reasonable expectations (what it is reasonable to believe, given the words that were used). Formally, this semantic information is captured using probabilistic and information theoretic tools. Vagueness, understood in terms of borderline cases, is argued to be a byproduct of making the above learning and communication requirements central. Vagueness, understood as our ability to be vague with words, is given an information theoretic explanation. Finally, the account is defended With respect to some of the philosophical problems and puzzles found in the vagueness literature.