NeurIPS 2025 researchers name Artificial Hivemind phenomenon

Researchers have identified the Artificial Hivemind, a phenomenon where different AI models produce strikingly similar outputs to open-ended prompts due to typicality bias in human preference data used during post-training. This bias rewards familiar, easily processed responses, causing mode collapse toward common answers. The author explains that less typical, more original answers still exist within the model's distribution and can be recovered using techniques like verbalized sampling. Open article →