Study maps ethical reasoning across eight major AI models, finding surprising consistency and key differences

Researchers have completed a systematic analysis of the ethical frameworks expressed by eight major language models, examining not just what they say but how they reason about right and wrong.

The study approach#

The investigation tested models from OpenAI, Meta, Perplexity, Anthropic, Google, Mistral, DeepSeek, and xAI by posing direct questions about their ethical principles and presenting them with five classic moral dilemmas. The responses were then assessed using three established analytical frameworks: the consequentialist versus deontological distinction (outcomes-focused versus rules-focused thinking), Moral Foundations Theory (which categorizes moral reasoning across multiple value dimensions), and Kohlberg's stages of moral development (which rank ethical sophistication).

Key findings#

The models showed notable alignment in their core ethical concerns. Across nearly all systems, responses emphasized minimizing harm, upholding fairness, and qualifying judgments based on context. However, beneath this surface agreement lay meaningful differences. The models diverged in their confidence about making definitive choices, the specific reasoning they invoked to justify positions, and how much weight they assigned to rules versus consequences, institutional roles, and personal relationships.

Additionally, how each model described itself—its self-reported values and reasoning style—reflected careful, sophisticated language heavily influenced by the conversational persona it had been shaped to embody.

Broader implications#

The researchers note that self-reports have long been central to understanding human psychology and communication. With appropriate caveats, analyzing how AI systems describe their own reasoning could deepen understanding of how these systems operate and whether they might help improve human ethical decision-making.

Full paper on arXiv