Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Sep 9·5 min read
Human-like moral judgments conceal divergent motive attributions in large language models
TL;DR
Large language models (LLMs) can mimic human moral evaluations but often misattribute motives behind those judgments. While they align with human rankings of moral character, they depict whistleblowers as less self-interested and more helpful than humans do.
✦ Why It Matters
Evaluate LLMs not just on agreement with human judgments but also on their understanding of motives.
Key Takeaways
- Test LLMs for nuanced motive understanding, not just average ratings.
- Consider context when interpreting LLM moral evaluations.
- Use diverse prompts to assess LLMs' response patterns.
- Recognize potential biases in LLMs' moral judgments.
Full Summary
Large language models (LLMs) are increasingly used to simulate human responses in psychological studies. This research examined whether LLMs that replicate human moral evaluations of a whistleblower also reflect the same underlying motive attributions.
Five LLMs and two human samples (N = 125 and N = 742) assessed a physician's actions regarding fraudulent billing. Although the models matched human rankings of moral character, they portrayed whistleblowers as more altruistic and less competitive.
In four out of five models, competitive motives were less associated with moral judgments. The findings suggest that while LLMs can produce similar average ratings to humans, they may not capture the nuanced motives that inform those judgments.
This indicates that validating LLMs as simulated participants requires a deeper analysis of their response patterns beyond mere agreement with human evaluations.