Quantifying Ethics in Artificial Intelligence

Original article can be found here (source): Artificial Intelligence on Medium

Machine learning works by quantifying the features of a particular object or concept, building that concept into a model, and passing that model through a process of guesswork, error measurement, adjustment of weights, and a repetition of the process until a desired outcome is achieved that matches the objective function. On the technical side, if AI ethics is going to make strides, the field will need to define, determine, and quantify the features of ethics.

The Ethical Leadership Behavior Scale as developed in the 2010 study “Actions Speak Louder Than Words: Benefits of Ethical Behaviors of Leaders” provides one possible strategy for obtaining the quantification of ethics to be used in machine learning. In this study the primary goal of the researchers from the Department of Psychology at the University of Zurich was to “build and test a first version of an Ethical Leadership Behavior Scale (ELBS)… and to expand upon previous assessments of ethical or authentic leadership by assessing concrete ethical manifestations of varying difficulties.”[1]

A secondary goal of the study was to test the correlation between the ethical behavior of leaders and the satisfaction of those they supervise. The study incorporated 592 employees of 110 work units in two departments. Through a rigorous process 35 items were decided upon to become the metrics upon which the participants would be measured[2]. After the metrics of the ELBS were determined the researchers used group data and informational interviews to assign ELBS scores to supervisors in the system.

Critical in this process was a use of a modified Rasch system to analyze results. As the study notes, “The Rasch approach provides two separate measures: (a) the person’s performance level and (b) the difficulty for each behavior. A behavior difficulty is defined as the probability that anybody will behave in a certain way, regardless of his or her specific attitude or motivation to act.”[3] The system then calibrates those measures and puts them in relationship with one another in order to assign a person a numerical representation of their ethical leadership value.