"Heuristic" theories of autism postulate that a single mechanism or process underpins the diverse psychological features of autism spectrum disorder. Although no such theory can offer a comprehensive account, the parsimonious descriptions they provide are powerful catalysts to autism research. One recent proposal holds that "noisy" neuronal signalling explains not only some deficits in autism spectrum disorder, but also some superior abilities, due to "stochastic resonance". Here, we discuss three distinct actions of noise in neural networks, arguing in each case that autism spectrum disorder symptoms reflect "too little", rather than too much, neural noise. Such reduced noise, perhaps a function of atypical brainstem activation, would enhance detection and discrimination in autism spectrum disorder but at significant cost, foregoing the widespread benefits of noise in neural networks.