Analyzing the Ethical Logic of Eight Large Language Models

LLMs show convergent ethical reasoning across different companies: All eight models tested emphasized harm minimization, fairness, and contextual qualification when addressing moral questions, suggesting alignment around core ethical principles despite coming from different organizations.

Models differ significantly in their willingness to make decisions: While the models shared similar ethical values, they varied in how readily they would commit to specific choices in moral dilemmas, indicating different approaches to handling ethical uncertainty and ambiguity.

AI ethical reasoning is heavily shaped by design choices, not just training: The models' self-descriptions were "erudite, cautious, and strongly shaped by a conversational persona," suggesting that how companies design AI systems to communicate influences their expressed ethical logic as much as underlying values.

LLMs use multiple ethical frameworks simultaneously: The study found models weighted rules, outcomes, role obligations, and interpersonal considerations differently, meaning they don't rely on a single ethical approach but blend consequentialist (outcome-focused) and deontological (rule-based) reasoning.

AI self-reports can reveal how systems work: The researchers propose that analyzing what AI systems say about their own ethical reasoning—similar to how psychologists study human self-reports—can help us understand AI decision-making and potentially improve how AI augments human ethical behavior.