THUFIR - A virtual Ethics Advisor

Abstract

We present THUFIR (Text and Hera for Unified Formal Intelligent Reasoning), an interactive ma-chine ethics system designed to guide users in navigating moral dilemmas. THUFIR is a hybrid sys-tem. It employs the HERA library for ethical reasoning as its verifiable, top-down component and an LLM as its data-driven, bottom-up component. Users present their moral dilemmas and interact with the system through a conversational web interface. When the system has gathered enough information about the dilemma, it models and evaluates the dilemma using a programmatic implementation guided by utilitarian and Kantian ethical principles. Each of these principles produces an independent evalua-tion of the dilemma, which is presented and explained to the user through the moral reasons underly-ing their specific situation. The potential divergences in these evaluations are then highlighted by con-trasting the consequentialist and deontological foundations, i.e. utilitarian and Kantian versions, re-spectively, of their underlying ethical theories. We also conduct pilot user experiments to identify common use patterns and attitudes toward the system. We find that THUFIR’s use of multiple ethical theories produces a sense of impartiality and trustworthiness, although some users would prefer more directly applicable and committed ethical guidance. Furthermore, we conclude by discussing a range of similar qualitative findings and THUFIR’s inherent limitations.