Camera traps are a powerful and increasingly popular tool for mammal studies. However, like all survey techniques, camera traps have their limitations. While researchers recognise constraints associated with experimental design or the cameras themselves, photo mis-identification is often neglected. Identifying mammals from camera trap photos can be unreliable and potentially influenced by the expertise of the observer, yet few studies have examined this source of error. We evaluated the effects of observer experience, confidence, animal body size and uniqueness on the accuracy of mammal identifications from camera trap images. Between 20 December 2016 and 22 February 2017, an internet survey was conducted, consisting of 20 questions regarding observer experience and 60 camera trap images to be identified. Images were sourced from surveys in northern Australia and included 25 native mammal species, ranging in size from the delicate mouse (< 15 g) to dingo (> 10 kg). A total of 176 people participated in the survey with experience ranging from none to > 40 years of mammal trapping. There was surprisingly little relationship between the accuracy of mammal identification and levels of experience or confidence. Accuracy was highest for unique species (such as the echidna) and lowest for small indistinct species (such as rodents and dasyurids). There was also a positive relationship between accuracy and body mass. Our results highlight that development of computer-assisted identification is needed, particularly for studies where similar-looking mammals co-occur. However, in the interim, integration of camera traps with conventional survey techniques (e.g. trapping) can aid positive identifications.