Position: Artificial Hivemind Evidence Does Not Support Societal Homogenization Claims

Jiang et al. (2026) in the NeurIPS 2025 Best Paper, report that large language models produce strikingly homogeneous responses, termed the Artificial Hivemind effect. They suggest that this effect could lead to the ``long-term homogenization of human thought and behavior’’. We argue that this societal claim is not supported by their evidence because it relies on a user-agnostic evaluation, where every model receives identical prompts without user-specific context. As a simple counterexample, we prepend user information, persona, to the prompt, allowing the model to personalize its responses, which substantially reduces the reported homogeneity. Our findings suggest that measured AI homogeneity is highly sensitive to evaluation assumptions and that claims about the societal impact of AI homogeneity should be based on evaluation protocols that reflect realistic deployment settings, including personalization and heterogeneous users.

preprint