Is AI Necessary In Deciding Whether an Offender Is Likely To Recidivate, With & Without the Effect of Protected Characteristics

Our research topic addresses the issue of whether or not an AI algorithm can be developed to forecast whether or not a criminal would recidivate by removing protected characteristics from models in terms of both accuracy, precision and recall and their equality across different groups, as well as if such a tool is necessary

Our research topic addresses the issue of whether or not an AI algorithm can be developed to forecast whether or not a criminal would recidivate by removing protected characteristics from models in terms of both accuracy, precision and recall and their equality across different groups, as well as if such a tool is necessary. The background of this issue is that there is always an underlying human prejudice present in the criminal justice system when it comes to making judgments, thus it is crucial to determine whether or not this human bias is required while making these decisions and whether there is similar kind of bias with machine learning methods. There are various ways to approach and respond to this topic, but the strategy we used involved examining the key variables that might influence prejudice, such as gender or ethnicity, and comparing how well the AI system can predict whether someone is likely to recidivate with or without incorporating these protected characteristics. The research's significant findings are that, after utilizing a number of different models, logistic regression was used to determine whether or not an offender was likely to recidivate. The accuracy of the predictions for our logistic regression model was 67.74 percent without the removal of any protected characteristics that hold risk of being discriminated against unfairly, but it increased to 68.16 percent when race and gender were taken into account. The random Forest model revealed that, after accounting for the gender characteristic, the mean absolute error—the percentage of variance between each measurement and the quantity's actual value—was 34.06 percent along with an accuracy score of 65.94 percent. The main results drawn from this demonstrate that utilizing various models and measuring the precision, recall, and accuracy of each race and gender are comparable with or without protected characteristics. Although the accuracy of this still isn't the finest, it could demonstrate that bias may not be as significant a factor in decision-making as it may first appear. This, however, can also be contested because the dataset utilized to train these models could be encoding real world bias such that race and gender get encoded into other variables such as prior offenses; this would mean that even if we removed race and gender data it would still be present in our model through other variables. The question of whether or not we should remove protected characteristics is required and needed remains unanswered as a result.

Explore More!

Research Paper