We sketch an inference architecture that permits linguistic aspects of politeness to be interpreted; we do so by applying the ideas of politeness theory to the SCARE corpus of task-oriented dialogues, a type of dialogue of particular relevance to robotics. The fragment of the SCARE corpus we analyzed contains 77 uses of politeness strategies: our inference architecture covers 58 of them using classical AI planning techniques; the remainder require other forms of means-ends inference. So by the end of the paper we will have discussed in some detail how to interpret automatically different forms of politeness — but should we do so? We conclude with some brief remarks on the issues involved.