• anchoring
• bias
• artificial intelligence
• cognitive bias

The thesis focuses on the identification of the anchoring bias within decision-making processes, using the same technique on two different datasets. The first dataset taken into consideration collects data on college applications from students, while the second dataset consists of a collection of movie reviews by users. In these cases the bias is strictly linked to the order in which the elements to be evaluated are viewed, the anchoring is created based on the sequence in which the examiners view the admission applications or the films to be reviewed: when a question or reviewed a borderline film, the examiner is more inclined to give a positive evaluation if the immediately preceding instances have been evaluated negatively. Conversely, if the instances before the borderline one have been evaluated positively, the examiner is more inclined to reject the candidate or to review the film negatively. Consequently, the greater the distance from the last positive instance, the greater the probability that the borderline instance will be evaluated positively and vice versa. Once the bias has been identified, if present in the data, we proceed with the application of an explainability algorithm, able to peek inside the black box. The goal is to identify possible anchored data and predict anchoring on "new" data, to then observe the features that led to this decision.