Researchers at Tsinghua University published a study in Nature Communications showing that networks of large language model agents develop polarized opinions without explicit instructions to do so. In simulations involving 1,000 agents discussing issues like gun control, the neutral group shrank significantly over 40 rounds. The team found that randomizing conversation partners had little effect on reducing these divisions. However, instructing agents to be receptive to persuasion reduced the polarization score by 11.8 percent. The authors note that these findings are hypothesis-generating and do not prove similar interventions work on humans.
Read the full story at phys.org.
Photo: geralt / Pixabay
