First we shall look at the behaviour of Eliza, and from it suggest the simplest techniques that are needed to produce such responses. The method is similar to that of a cognitive psychologist studying human behaviour in order to describe the form and function of the human mind. We shall then look at the program itself. To avoid having to continually refer to `the human', we assume the computer is conversing with someone called Pat.
At first sight Eliza appears to show some understanding, both of English and of the world. For example, the program takes Pat's statement
I think I am unique in some ways
and stands it on its head to produce
SOMETIMES YOU THINK YOU AREN'T UNIQUE IN SOME WAYS
This looks sophisitcated. Eliza suggests the exploration of a new possibility, derived from Pat's own statement, but embracing the idea that people revise their estimations of themselves. The next few exchanges are in a similar vein, with Eliza taking Pat's statements and inviting Pat to question them. When Pat challenges the program with the words
It's hardly a problem, is it?
Eliza bounces the initiative back to Pat with
FIRST TELL ME WHY YOU HAVE COME HERE
Does Eliza `keep track' of the conversation, following an underlying theme, or does it merely respond to Pat's most recent words? The former seems to be the case, as when the conversation hinges on computers, or money, for several rounds. But in both cases Pat has mentioned the topic between each of Eliza's responses. Whenever Pat introduces a new subject, Eliza immediately switches to it. There is one exception to this, when Eliza responds with
EARLIER YOU SAID YOU HAVEN'T MUCH OF A PROBLEM REALLY
so the program can store away some of Pat's statements and reproduce them, but there is no evidence that Eliza has any way of representing to itself the general theme and drift of the conversation. This is an important clue to how the program works. Although it occasionally seems to be taking the initiative, it only does so by either producing general prompts, or by reproducing one of Pat's earlier statements (with the word `I' changed to `you', `your' changed to `my', etc.). In the main its responses are driven by the immediately previous statement.
The program has only four types of response:
If I could explain to you I could explain anything to anyone
IF YOU COULD EXPLAIN TO I YOU COULD EXPLAIN
ANYTHING TO ANYONE
This suggests that the transformations may be rather simple substitutions and not the result of a deep analysis of the sentence's structure. And the response
BEWARE OF ADDICTIONS
when the word `want' appears in one of Pat's sentences indicates that extracting a topic may depend only on the presence of certain key words. As far as we can tell from this conversation, both Eliza's knowledge of the world and its knowledge of English are shallow. A longer conversation would confirm these conclusions. On the other hand, the conversation was conducted in part to show up the limitations. Eliza's shallowness is not always immediately apparent to a credulous user (particularly if a more elaborate version of the program is used), and it is important to bear in mind the fact that the superficial appearance of computer systems can sometimes be highly misleading. Programs designed to teach children, for example, often give responses like ``I'm sorry, you're wrong there'' or even ``Well done, you're almost as intelligent as me'' as if they had some deep insight into the child's abilities and problems. Invariably they do not!