In psychology, a common way of investigating human information processing is by response time experiments. Suppose with Prolog that you had some way of measuring the time between adding a fact and getting the prompt, and between asking a question and getting the answer. How could you determine which kind of inference is going on?
If you don't like thinking about the psychology of computers, put yourself in the place of a psychologist who has just discovered live Martians. They speak a kind of English: to be more exact, they behave just as Prolog would if it worked on speech not text. Our psychologist has discovered that the Martians make the same kind of inferences as Prolog: but is this by forward or backward chaining?