LLMs mimic human moral judgments but diverge on motive attribution

A new study published on arXiv reveals that large language models (LLMs) can mimic human moral judgments but fail to replicate the underlying attributions of motives. While LLMs correctly ranked a whistleblower's moral character similarly to human participants, they attributed different motives, portraying whistleblowers as more helpful and less self-interested. This divergence in motive attribution, even when LLMs reproduce human-like average ratings, highlights the need for more nuanced validation methods beyond simple agreement to ensure LLMs' reliability as simulated participants in psychological research. AI

IMPACT Highlights the need for advanced validation of LLMs in research, beyond simple agreement, to ensure their reliability in simulating human responses.

RANK_REASON Academic paper published on arXiv detailing LLM capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Large language models
• Whistleblower

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →