Algorithm

bias, power, and ethics in the age of Artificial Intelligence

Abstract

Through the exploration of a real database, we conducted an empirical study to try to answer the research questions: whether discriminatory biases exist and can be identified in AI, and, in a second stage, whether mitigation is possible. The methodological procedure adopted includes a bias identification phase, through the application of three types of processing techniques to the database (pre-processing, in-processing, and post-processing), in which the results were compared and the decisions made by the model were verified, resulting in an Explainable AI, whose data can be understood, scrutinized, and challenged according to the values described in our 1988 Constitution. In the future phase of proposing actions, in order to bring constitutional ethics closer to algorithm development, we intend to "train the machine" to include the foundations, values, and principles of the Brazilian Federal Constitution and see if the system makes the database less biased. The immediate linguistic objective is to deconstruct the belief in the infallibility of automation and propose future improvements for system operators. To better understand the biases/discriminations reproduced by AI, the theoretical framework of Critical Discourse Analysis and Decolonial Theory served as a compass. The contributions are multiple and fall within the scrutiny of a discourse that was imagined to be free from social ills and in bringing machine operators closer to a more tangible ethic, foreseen in a reliable Document. The results are still preliminary and only account for the bias identification phase, as we were able to demonstrate their existence and point them out in the chosen database. The offering of coping tools will be left for the opportunity to complete the postgraduate course. For now, we share the certainty of abuses produced in technological means, in line with others existing in other areas of our society.