I think a lot of people are missing the point. The idea is not that the computer DOES what you want ahead of it, the idea is that the computer becomes more aware of itself and the user's habits so that it can set aside resources appropriately, and better understand the commands from the user (whatever the input). This is not so important for dedicated input systems such as touch, keys, and mouse where there is a very ridged meaning behind each command, but it is absolutely necessary for better speech control, and 'combined commands'.
The idea is that you could tell the computer to check a document for errors and send it to someone. The computer would be able to recognize that you mean spell and grammar check, export to PDF or some other document type that the other user can accept (because cognitive understanding can be implemented for other 'users' that are not the host user), and then send that document in a way that the receiver would prefer (email, file transfer, FTP, etc.). It means high level control of a machine instead of low level control, it does not mean that the computer is sentient, or makes 'moral' decisions, or any decisions outside of the realm of the task at hand.
It also means that the computer can be better self-aware. If you are doing a specific activity the computer will be better able to cache resources for said activity and predict the user's intent. It does not activate it until the command it given, but it would be a better end-user expierence for your computer to act one way during the work day, and annother way during hours that you would normally have as recreation. You could still do activities that the computer is not expecting, but at least it would be more effective at guessing your needs more of the time than not guessing at all.