Studies of social learning suggest that many animals are disproportionately likely to adopt the behaviour of the majority, and that this conformist transmission hinders the spread of novel behavioural variants. However, novel learned behaviour patterns regularly diffuse through animal populations. We propose a hypothesis, termed the 'social release hypothesis', that resolves these apparently conflicting findings by suggesting that animals are released from conforming to traditional behaviour in the absence of demonstrators. We investigated the role of pretrained, female demonstrator guppies, Poecilia reticulata, in influencing the escape response of untrained females to an artificial predator. Naïve 'observer' guppies were given the opportunity to follow trained demonstrators through an established escape route or escape independently via an alternative route. In the presence of demonstrators naïve fish overwhelmingly preferred to escape via the route taken by the demonstrators and escaped more quickly than fish in a control group with sham demonstrators. However, once the demonstrators were removed, the naïve fish were equally likely to use either escape route, despite once again escaping significantly faster than the controls. These findings are consistent with the social release hypothesis. Conformist social learning ensures that the behaviour of animals within a population is similar, and allows individuals rapidly to acquire locally adaptive information. However, innovations may spread as a consequence of social release.