Reproduce “Differences in the Moral Foundations of Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Peter Kirgis
Peter Kirgis (Princeton University) investigated the normative ethical judgments of 16 frontier Large Language Models using Moral Foundations Theory (MFT) to characterize their value biases against a human baseline. The study found LLMs generally exhibit a "liberal" moral profile, overvaluing individualizing foundations and undervaluing binding ones, with more capable models showing increased divergence from human moral intuitions.
Key insights
- Moral Foundations Theory (MFT) is a robust and explanatory framework for understanding the moral biases and profiles of Large Language Models.
- LLMs consistently display a "liberal slant" in their moral judgments, prioritizing individualizing foundations (care, fairness, liberty) over binding foundations (loyalty, authority, sanctity) compared to human intuitions.
- As LLMs increase in scale and capability, they tend to exhibit a 'moral drift,' moving further away from the human moral baseline, particularly in their valuation of binding foundations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.