Exploring bias analysis on judicial data using machine learning techniques

The use of data driven automation is not new, but it has gain a lot of attention recently with the wide-spread understanding that it is the solution to all problems in terms of ‘fair’ and ‘non-bias’ classification. This is not different in the law area, where ‘artificial intelligence’ became a ‘magic word’. However, using historic data is a very tricky job which can quite easily propagate discrimination in a very efficient way. Thus, this work is aimed to analyse data from legal proceedings looking for evidence related to the occurrence of bias in the judges' decision-making process, considering mainly the gender or social condition of the convicts. Supervised and unsupervised machine learning techniques, preceded by data analysis and processing procedures, were used to explain and find explicit data behaviour. Our results pointed to the fragility of the techniques to identify biases but suggest the need to improve data pre-processing and the search for more robust classification techniques.