This paper will review some of the arguments for and against the use of semantics in MT. It will start by presenting a view of computational semantics as it is currently being investigated in the LRE FraCaS project. We will then consider three kinds of translation: sentence-sentence translation, paragraph-paragraph translation and dialogue translation aids as envisioned in the Verbmobil project. The latter two as we currently conceive them seem clearly to need support from semantics. In the case of sentence translation the issue is perhaps not so clear. We will examine arguments that semantics is not only not needed but also gets in the way. We will then examine arguments that have been advanced that semantics is necessary. We will support the view that transfer needs to involve the integration of several levels of information and that transfer also involves negotiation. We will also support the view that the representation of underspecified information is of key importance. That is, we need to develop techniques for representing the semantic information that is useful for translation without representing the semantic information that gets in the way. We will conclude by presenting some of the demands that translation may place on computational semantics.