Can Mathematical Models Be Weapons of Mass Destruction?

Abstract

With the widespread adoption of digitalization, mathematical models have become indispensable in every field. Predictions are made, processes are evaluated and optimized, and future forecasts are developed using mathematical models. The usefulness of these models has accelerated their proliferation. Their application in nearly every aspect of life has triggered a new phase of modeling, where the output of one model can now serve as the input for another, enhancing overall efficiency. Consequently, models are no longer discrete but interconnected, encompassing and influencing human life. At this point, understanding how models operate is critically important for grasping how decisions affecting us are made. Therefore, in this study, mathematical models and algorithms are examined in detail based on Cathy O'Neil's (2016) book ‘Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy’. It is emphasized that a model does not encompass everything related to a given field, and therefore prioritizes aspects of the field, assigning weights externally during this prioritization. As a result, every model provides only an approximation for the field, meaning that elements not measurable within the model risk losing value over time. The biases present in the dataset used by a model can lead to biased outputs, thereby reproducing existing inequalities in society. It is particularly emphasized that the fact that models now serve as inputs for one another weakens the possibility of correcting biased outputs and increases the risk of further deepening inequalities. This risk is expected to grow significantly, especially with the widespread adoption of artificial intelligence technologies. Therefore, the study recommends adopting a participatory management approach during the development phase of mathematical models, enabling the involvement not only of domain experts but also of representatives of all stakeholders directly affected by the model. This approach could help prevent the use of biased assumptions and datasets in the models, thereby mitigating the potential negative impacts caused by these models.