This paper investigates the following proposal about machine intelligence: that behaviour in which a habitual response that would have been inappropriate in a certain unfamiliar situation is overridden and replaced by a more appropriate response be considered evidence of intelligence. The proposal was made in an earlier paper (Sterrett 2000) and arose from an analysis of a neglected test for intelligence hinted at in Turing's legendary 'Computing Machinery and Intelligence'; it was also argued there that it was a more principled test of machine intelligence than straightforward comparisons with human behaviour. The present paper first summarizes the previous claim then looks at writings about intelligence, or the lack of it, in animals and machines by various writers (Descartes, Hume, Darwin and James). It is then shown that, despite their considerable differences regarding fundamental things such as what kinds of creatures are intelligent and the relationship between reason, instinct and behaviour, all of these writers would regard behaviour that meets the proposed criterion as evidence of intelligence. Finally, some recent work in employing logic and reinforcement learning in conjunction with 'behaviour-based' principles in the design of intelligent agents is described; the significance for the prospect of machine intelligence according to the proposed criterion is discussed.