Consider a toy deduction system, e.g. ZOOKEEPER from [Winston, 1992]. It has rules such as the following:
It is a trivial fact that such systems don't know the meaning of the symbols that make up their sentences. So, for example, the ZOOKEEPER cannot answer questions such as:
Does ?x touch the ground while it flies?
Where are the eggs before ?x lays them?
even
Does ?x lay eggs?
These questions can be answered by any five year-old. They do not require any extra knowledge beyond the simple comprehension of the meanings of the terms such as ``fly'' or ``lay''. So failing this simple test, we can say that the ZOOKEEPER does not understand what these terms mean. Actually for all it is concerned, the rule could have been:
And looking from its perspective, you wouldn't have any advantage over the ZOOKEEPER to answer questions such as:
Qbrf ?x gbhpu gur tebhaq juvyr vg syvrf?
Jurer ner gur rttf orsber ?x ynlf gurz?
However, you could, in principle, manipulate the symbols of this language, to make the inferences ZOOKEEPER makes. Am I repeating the Chinese room argument? [Searle, 1980] No. I am just stating a simple fact that this small system doesn't have any understanding of the symbols it is using.
What would a program be able to do, if it did have any understanding of its
symbols? The Cyc book [Lenat and Guha, 1990] has many nice examples:
