What we know about the world is difficult to set down precisely; but if we can't get that knowledge inside a computer, it cannot converse about any and all subjects as HAL does. The real question is, then, How much does a computer need to know in order to talk about something coherently? As long as chess is the example, the knowledge needed is well defined and understanding is possible. But as soon as the domain of required knowledge is not well defined, things get very complex very fast.

Early AI work relied on chess-playing programs as a kind of "quick hit." Success was relatively easy, and all of a sudden computers seemed pretty smart. The problem is that what early AI researchers took as evidence of being smart was illusory.

The essence of the natural language problem is not language at all. A researcher can input definitions into a computer for decades, and still never give it the ability to understand human experience. People who argue against the possibility of machine intelligence often make similar statements, of course. Some of them might interpret this as a denial that a computer like HAL -- that is one with HAL's language abilities -- could ever exist. However, the problem is not one of language, but of knowledge and the acquisition of knowledge. A computer would need to know a great deal to engage in even simple dialogue. The writers of 2001, and most other lay people writing about AI, simply assumed that an intelligent machine would talk and respond as any human would. HAL talks just like a human being because the scriptwriters wrote dialogues for a human actor to speak. But would an intelligent computer sound just like a human?