Objects may be classified by many different criteria. An important distinction is between “artificial classifications” and “natural classifications”, where artificial classifications may adequately serve some limited purposes, but natural classifications overall are most fruitful by allowing inference and thus serving many different purposes. There is strong support for the view that a natural classification should be based on a theory (and, of course, that the most fruitful theory provides the most fruitful classification). Nevertheless, atheoretical classifications are often produced (and are sometimes termed “descriptive” classifications). Paradoxically atheoretical classifications may be very successful. The best example of a successful “atheoretical” classification is probably the prestigious Diagnostic and Statistical Manual of Mental Disorders (DSM) since its third edition from 1980. On the basis of such successes one may ask: Should the claim that classifications should ideally be natural and theory-based be reconsidered? This paper argues that the seemingly success of atheoretical classifications hides deeper problems and that the ideal of theory-based classification should be maintained.