# The Moral Mind(s) of Large Language Models

arXiv: https://arxiv.org/abs/2412.04476 · HTML: https://mlsanity.com/paper/2412.04476/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Builds on 1 of its 31 references that are in the map.
- First posted on arXiv 19 Nov 2024.

## Builds on (most cited first, 1 of 1)

- [Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies](https://mlsanity.com/paper/2208.10264/) (2022)

## Built on by (most cited first, 4 of 4)

- [Would a Large Language Model Pay Extra for a View? Inferring Willingness to Pay from Subjective Choices](https://mlsanity.com/paper/2602.09802/) (2026)
- [Revealed Rationality: Label-Free Evaluation and Regularization from Representation Theorems](https://mlsanity.com/paper/2608.05015/) (2026)
- [To Nuke or Not to Nuke: LLMs' (Missing) Ethical Reasoning and Actions in a High-Stakes Decision-Making Simulation](https://mlsanity.com/paper/2606.08310/) (2026)
- [Morality in AI. A plea to embed morality in LLM architectures and frameworks](https://mlsanity.com/paper/2511.20689/) (2025)

## Abstract (arXiv)

As large language models (LLMs) increasingly participate in tasks with ethical and societal stakes, a critical question arises: do they exhibit an emergent "moral mind" - a consistent structure of moral preferences guiding their decisions - and to what extent is this structure shared across models? To investigate this, we applied tools from revealed preference theory to nearly 40 leading LLMs, presenting each with many structured moral dilemmas spanning five foundational dimensions of ethical reasoning. Using a probabilistic rationality test, we found that at least one model from each major provider exhibited behavior consistent with approximately stable moral preferences, acting as if guided by an underlying utility function. We then estimated these utility functions and found that most models cluster around neutral moral stances. To further characterize heterogeneity, we employed a non-parametric permutation approach, constructing a probabilistic similarity network based on revealed preference patterns. The results reveal a shared core in LLMs' moral reasoning, but also meaningful variation: some models show flexible reasoning across perspectives, while others adhere to more rigid ethical profiles. These findings provide a new empirical lens for evaluating moral consistency in LLMs and offer a framework for benchmarking ethical alignment across AI systems.