It is argued that recent analyses of the evolution of animal signals, which claim that signalling systems must be honest indicators of underlying quality, have neglected a vital consideration: the costs receivers pay in assessment. Where the costs of fully assessing a signaller are high, in terms of energy, time, or risk, and the value of the extra information gained is low, then it will pay receivers to settle for cheaper, but less reliable, indicators of quality instead. Thus, it is argued, honest assessment will be replaced by conventional signalling. Conventional signals are open to cheating, but cheating will be kept at low frequencies by the frequency-dependent benefits of occasional assessment (or 'probing'), so dishonest signalling remains stable. The concept of 'honesty' is discussed. © 1991 The Association for the Study of Animal Behaviour.