Production noise, not comprehension noise, drives groups to wrong consensus

A new experimental study on collective information acquisition reveals a critical distinction between two types of communication noise. In an online experiment with 600 participants working in 4-person groups, researchers introduced either comprehension noise (each receiver independently saw a perturbed version of social information) or production noise (perturbations were stored before display and could be seen by multiple receivers). Groups subjected to production noise spent significantly more rounds tightly clustered around a wrong value than comprehension-noise groups (p=0.016), and that wrong common signal persisted across more rounds (p=0.004). Dynamic update models showed production noise was more harmful not because people followed peers more strongly, but because the same peer influence acted on more correlated perturbations.

Surprisingly, noise did not simply degrade performance. Comprehension noise sometimes improved correction relative to a faithful control condition. The findings challenge the assumption that all noise is equally detrimental. To explore mechanism, the team replicated the experiment with GPT agents, which registered uncertainty through reduced confidence but did not reproduce the human-scale vulnerability to production noise. This boundary condition suggests the effect is rooted in human psychology. The research has implications for designing group decision systems, social media algorithms, and AI-human collaboration, where production noise (e.g., a viral misquote or shared hallucination) can stabilize consensus on error.