Abstract
The rise of Artificial Intelligence (AI) has been exponential over the last five years, whether its schools, offices or personal life, AI has been integrated into every part of our lives. One such sector is the federal court system, although AI has recently become popular, it has been in use by courts since early 2000s. The courts use these AIs to get a recidivism score out of ten, which helps the judges decide on rulings regarding a defendant’s bail, sentencing or parole. These AIs on the backend are powered by enormous amount of data and advanced machine learning algorithms. The data and algorithms combined take a defendant’s information and then make predictions based on historical data and the defendants demographic. The use of AI in a court cases has raised many issues regarding the fairness of the trial, since a person is being judged on their demographics and not as individual. In addition to this, the risk and concerns regarding bias introduced by the AI are high, since the AI is being trained on historical data, it will only be as unbiased as the data, which many consider to be biased. Furthermore, AI can be heavily influenced by the developer who developed it, as they have tremendous control over how an AI is trained. Hence, it is extremely important that we research the use of AI and its effects in federal courts.
My STS research revolved around researching the use of AI in federal courts, and more specifically how the public feels about the use of AI in courts. In the research, literature review of multiple studies was done to gather quantitative data regarding the public’s opinion, the studies included different questions regarding the use of AI in different stages of case, just general questions regarding AI use and also different scenarios where AI could be used. Throughout the literature review it became apparent that the public supports the use of AI in courts when it is used for administrative tasks such as translation, transcription etc. However, it was very clear that public does not want AI to be used for decision making, it is important to note that the public did realize the benefits of AI, such as speed and helping reduce courts backlog. In addition to the literature review, a questionnaire for COMPAS, an AI used in courts, was examined for any questions that may introduce bias into prediction, and it was determined that the questionnaire didn’t only introduce bias, but it was also asking personal questions that had nothing to do with the case itself, this further helped come to the conclusion that the public did not support the use of AI in courts, when it came to decision making.
My technical research involves creating an unbiased AI that can be used in courts for predictions. The research included creating multiple AI models trained using different methods, one model being manually trained to remove any biased features from the data, a second model trained using THEMIS and AEQUITAS, which are tools used to determine if a model discriminates against individuals and also identify any discriminatory inputs, finally a third model will be trained using the IBM AIF360 toolkit which contains 71 bias-detecting metrics and nine bias mitigating algorithms. Due to the nature of AIF360 toolkit, it is fully expected that this model will perform the best and will produce the most unbiased prediction, however it will still be compared with the other models by comparing the precision, recall, true and false positive rates and finally the disparate impact score. Although this model will most likely not be implement by the courts, this research is done to give future researchers a framework to work on and further study.
Looking back at my research, I believe I was able to accomplish a lot. AI is a rapidly evolving field and being able to combine findings from various studies on public opinion, while also contributing new insights through questionnaire analysis, allowed me to write a paper that future researchers may find valuable while doing their research. I hope future researchers are able to conduct more surveys targeting students and the general public and are able to ask direct questions about the public’s support for AI use. Additionally, I was able to develop a detailed proposal for three AI models in courtrooms. While widespread implementation of my models is unlikely in the near future, I hope future researchers can build on this plan and evaluate how such models perform in practice. Ultimately, I hope government officials can use this research as a framework for implementing more unbiased AI in courtrooms.
