928 mation required to automatically transcribe speech to written text is simply insufficient for the larger task of creating a machine conversationalist that “understands.” For example, if someone says something that contradicts his or her earlier statement, we would expect a plausible machine conversationalist to spot it. Without some structure and memory, however, it is hard to see how a system could check statements for consistency. One could never expect to learn to do that simply from data: We just do not see or hear enough sentences to have previously encountered all the inconsistencies that we could spot immediately. At the moment, people encounter machine conversationalists only in recreational chatbots on the Web, or in simple phone transactions such as ordering travel tickets. But research systems are already much better than that, and the range of projects expected to deliver usable prototypes has expanded in recent years. These efforts range from the Defense Advanced Research Projects Agency’s Cognitive Assistant that Learns and Organizes project (8) to the European Commission’s new Companions project (9) to create a long-term conversational partner (see the figure). Such a Companion would learn its person’s likes and dislikes, carry out Web-related tasks accordingly, and prompt reminiscences about the person’s photo collection so as to build up his or her life story through conversation (10). Researchers generally agree that although these large goals need more research, speech recognition technology is still not accurate enough to build a reliable machine partner capable of understanding what we say, unless it has a considerable amount of stored knowledge to enable it to understand; mere reactive chatbots will be no more help than ELIZA was. The current paradigm split in research is about how it will be possible to capture and store knowledge and language experience in large enough detail and volume to build such assistants, outside of very small domains such as recording a complicated pizza order. A long-term assistant to an astronaut on a voyage to another planet, or one to help elderly people recover their past through conversation and organize it in text and images, is a much larger goal, and one that will require better machine learning techniques than have been deployed so far.