--- - |2+ Abstract Background Child molesters form a heterogeneous group, but one generally shared characteristic is maladaptive, rigid behaviour. Impairments in reinforcement learning may explain these maladaptive tendencies, but this has not been systematically investigated. Further, it is not known if such impairments vary with subtype of child molesters. Aims To investigate the presence of impairments in reinforcement learning among child molesters and to test for differences in patterns of impairment with subtype. Methods A group of 59 child molesters was recruited from several prisons in a two‐stage screening process, the first using records and the second interview; a comparison group of 33 offenders who had never committed a sex offence and who denied paedophile ideation was similarly recruited; 36 nonoffender comparison men were recruited by social media and word of mouth. Each was asked to perform a probabilistic reversal learning task, in which stimulus‐outcome contingencies had to be learned. Results Child molesters, as a group, made significantly more errors on the probabilistic reversal learning task than the nonoffenders; the comparison offenders and the nonoffenders gained similar scores, although findings may have been confounded by older age in the child molester group. Nonpaedophilic child molesters had significantly worse scores than paedophilic child molesters. Conclusions Child molesters, especially those not diagnosed with paedophilia, have deficits during both the acquisition and reversal of contingencies, suggesting reinforcement learning deficits that may undermine their capacity to benefit maximally from therapy without preliminary work to repair those deficits, possibly in conjunction with extending the offender programmes. Testing before programme entry would enable accurate targeting of scarce resources in this respect. - 'Criminal Behaviour and Mental Health, EarlyView. '