Correlation Metrics for Safe Artificial Intelligence

There is a growing need to provide AI risk management models that can assess whether AI applications are safe and trustworthy, to make them responsible. To date, there are a few research papers on this topic. To fill the gap, in this paper we extend the recently proposed SAFE framework, a comprehensive approach to measure AI risks across four key dimensions: security, accuracy, fairness, and explainability (SAFE). We contribute to the SAFE framework with a novel use of the coefficient of determination ($R^2$) to quantify deviations from ideal behavior not only in terms of accuracy but also for security, fairness, and explainability. Our empirical findings shows the effectiveness of the proposal, which leads to a more precise measurement of risks of AI regression applications, which involve the prediction of continuous response variables.