The effects of inert gas narcosis on behavior before unconsciousness are reviewed with particular attention to four issues. The first is whether the qualitative behavioral effects of all inert gases are identical. Evidence is limited but does not contradict an affirmative answer. This is consistent with the unitary hypothesis of narcosis at the physicochemical level. The second issue concerns the relative merits of four approaches to narcosis; (a) the descriptive model, (b) the hierarchical organization hypothesis, (c) the operant paradigm, and (d) the slowed processing model. It is concluded that the latter two are showing some promise. In particular, operant techniques allow more sophisticated measures of narcosis in animals than behavioral end points, such as loss of the righting reflex. The slowed processing model claims that the majority of performance deficits in humans are caused by a single fundamental deficit, slowing of information processing due to decreased arousal. This slowing is usually accompanied by alterations in task strategy. These alterations, in combination with cumulative slowing in working memory, are said to account for the various manifestations of narcosis on complex tasks. The third issue concerns adaptation to narcosis. There is some evidence that adaptation can occur but it is unclear whether the cause is learning specific to narcosis or development of a physiological tolerance. However, adaptation has not always been found and the variables controlling its presence or absence have yet to be identified. The fourth issue concerns the modifying effects of various factors, such as carbon dioxide and anxiety, on narcosis. Methodological and conceptual problems hinder interpretation of the evidence in this area but, contrary to some current views, there appears to be no conclusive evidence that any factor other than ethanol potentiates narcosis. Some implications of these conclusions for diving operations are discussed.