Non-montonic logics have been used recently for a variety of A.I. purposes, including belief revision and default reasoning in question-answering and exper systems. This paper argues that by their nature, such systems dicard information which has a role in human belief systems. In particular, systems which use non-monotonic reasoning lose the distinction between fully justified inferences and reasonable presumptions, in the process losing the ability to record failed expectations as such, an ability which provides a useful measure of salience for A.I. systems.
