Скачать или смотреть LLM Alignment in Collective Reasoning: Masking or Mirroring Bias

The provided source outlines an empirical study examining human-AI alignment in collective reasoning, specifically focusing on leader election in a "Lost at Sea" social psychology task. Researchers conducted a large-scale human experiment where groups elected leaders with either visible demographic attributes or pseudonymous aliases to observe biases. Matching these results, they simulated groups of large language models (LLMs) like Gemini, GPT, and Claude to benchmark their behavior, revealing a "mask-mirror" alignment tension. The results indicate that some LLMs mirror human biases and suboptimal choices (like Gemini and GPT), while others (like Claude) mask these biases by selecting more optimal, meritocratic leaders, demonstrating that alignment is context- and model-dependent. The study ultimately argues that understanding whether LLMs reproduce or compensate for human social dynamics is essential for advancing socially-aligned AI.