Estimates and inferences about personality traits, disposition and future behaviour based on physical appearance have a long history in philosophical, anthropological and psychological research. Over the past years, due to increased threat of terrorist attacks, the topic of criminality inferences based on physical appearance became again actual. Therefore, the objectives of the research were to examine: 1. the accuracy of the criminality inferences based on facial characteristics; 2. correlation between self-estimated accuracy of inferences and actual accuracy of guilt / innocence inferences; 3. the accuracy of the inferences about the type of offense committed by the perpetrator. The survey was conducted on-line, on a convenience sample of 832 adult persons, majority of whom were young (M=24 years) women (80.7%) holding secondary school degree (53.9%). In the study, 20 photographs of women were used, which were successively presented to the participants for three seconds. Photos of innocent persons were collected from acquaintances, while photographs of perpetrators were collected from publicly available databases of Interpol's international arrest warrants. Photographs of satisfactory technical characteristics (e.g., brightness, resolution) and personal and facial characteristics (e.g., neutral expression, without prominent features) were selected through pilot study. The average accuracy of inferences about innocence / guilt was 57%. The accuracy of inferences about innocence of the person was significantly higher (66,8%) than the inferences about guilt (44.7%) (t (832) = 24.062; p <.01). There was no correlation between the self-estimated accuracy of inferences about persons criminality based on the appearance and actual accuracy of the inferences (r (832) = -. 014; p> .05). However, participants who estimated that their inferences more accurate were statistically significantly more accurate in their inferences about criminals (r (832) = 088; p <.05; t (478) = 2.1; p <.05) and statistically significantly less accurate in estimates of innocent. (r (832) = - 114; p. <01; t (478) = 2.6; p. <.01). Participants who accurately estimated who perpetrators were, had problems in identifying specific crime that person has committed. Their estimations were bellow guessing level.