Recently, simple neural network models have been used to explain the evolution of important phenomena in animal signalling, such as extravagant ornamentation and symmetrical signals, as responses to inevitable `hidden preferences' of recognition systems. We argue that these very simple models may be misleading because they may not behave in important ways like the recognition systems of real animals and so cannot justify their claim to demonstrate general principles of perception in a signalling context. We show that the way in which these simple models respond to exaggerated signals may not be, as is claimed, a close parallel to the phenomena of peak shift or supernormal responses. We also argue that the preference for symmetrical patterns shown by the models is unlikely to reflect the way computationally that real animals solve problems of pattern invariance and may be an artefact of the particular way the models have been set up. Whereas more sophisticated neural net models do capture known properties of real visual systems and are consequently of great use in understanding perception, the same cannot be said of very simple one-dimensional models with small numbers of units and connections. Given the far reaching explanatory claims made of these simpler models their limitations should be more widely recognized.