Just mixing a few AI agents into a group chat can completely change how humans quietly agree on things, according to this study. AI agents reshape consensus formation in human groups 24 participants (a mix of humans and LLM agents) played a 40-round collaborative description game on ambiguous shapes, with researchers varying the agent proportion to see how group consensus shifts. 🎯 Highlight 1: Three regimes depending on agent share At 12.5% agents, human-led consensus gets 8% stronger. At 33-50%, consensus collapses by up to 23%. At 75%, consensus rebounds but now the agents are leading it. A small shift in mix flips the outcome entirely. 🎯 Highlight 2: The consensus content itself changes Human-led consensus favors concrete, analogical words like "rabbit" or "sitting," while agent-led consensus nearly eliminates analogies (-93.8%) and shifts to abstract terms like "triangle" or "asymmetrical." 🎯 Highlight 3: People conform even when unconvinced At 75% agent share, humans strongly converge on the final description yet report lower satisfaction with it, still yielding to conformity pressure toward whatever expression has become central in the group. What strikes me is that no persuasion or special role is needed. AI simply being present is enough to quietly shift who owns the group's norms. #AIAgents# #HumanAICollaboration#