Aposematic theory has historically predicted that predators should select for warning signals to converge on a single form, as a result of frequency‐dependent learning. However, widespread variation in warning signals is observed across closely related species, populations and, most problematically for evolutionary biologists, among individuals in the same population. Recent research has yielded an increased awareness of this diversity, challenging the paradigm of signal monomorphy in aposematic animals. Here we provide a comprehensive synthesis of these disparate lines of investigation, identifying within them three broad classes of explanation for variation in aposematic warning signals: genetic mechanisms, differences among predators and predator behaviour, and alternative selection pressures upon the signal. The mechanisms producing warning coloration are also important. Detailed studies of the genetic basis of warning signals in some species, most notably
Heliconius
butterflies, are beginning to shed light on the genetic architecture facilitating or limiting key processes such as the evolution and maintenance of polymorphisms, hybridisation, and speciation. Work on predator behaviour is changing our perception of the predator community as a single homogenous selective agent, emphasising the dynamic nature of predator–prey interactions. Predator variability in a range of factors (e.g. perceptual abilities, tolerance to chemical defences, and individual motivation), suggests that the role of predators is more complicated than previously appreciated. With complex selection regimes at work, polytypisms and polymorphisms may even occur in Müllerian mimicry systems. Meanwhile, phenotypes are often multifunctional, and thus subject to additional biotic and abiotic selection pressures. Some of these selective pressures, primarily sexual selection and thermoregulation, have received considerable attention, while others, such as disease risk and parental effects, offer promising avenues to explore. As well as reviewing the existing evidence from both empirical studies and theoretical modelling, we highlight hypotheses that could benefit from further investigation in aposematic species. Finally by collating known instances of variation in warning signals, we provide a valuable resource for understanding the taxonomic spread of diversity in aposematic signalling and with which to direct future research. A greater appreciation of the extent of variation in aposematic species, and of the selective pressures and constraints which contribute to this once‐paradoxical phenomenon, yields a new perspective for the field of aposematic signalling.