Moral Asymmetries in LLMs

Human moral reasoning is complex and is associated with a range of biases and asymmetries. The present study investigates two such moral asymmetries in the current version of ChatGPT: asymmetry in the attribution of intentionality and asymmetry in the attribution of causality. Both phenomena pertain to the fact that not only do our judgments of intentionality and causality affect our judgments of whether behavior is morally permissible, but our perception of the moral status of the behavior itself can affect how we attribute intentionality and causality to the agents involved. An analysis of data points collected from ChatGPT-4 using measures typically employed with human participants revealed that ChatGPT-4 mirrors two moral asymmetries found in human moral reasoning. The analysis also confirmed that ChatGPT-4 frequently misinterprets simple text, shows hypersensitivity to linguistic variation, and demonstrates low sensitivity to reasons for rule violations.