Study: 12.5% AI boosts group consensus by 8%, but at 75% humans adopt the agents' norms
rohanpaul_ai · x · 2026-09-06
Researchers mixed humans and LLM agents into 24-person groups that repeatedly had to agree on descriptions of the same image, revealing how AI participation share reshapes coordination:
- Low share helps: at 12.5% AI, consensus improved 8.0% over the all-human group.
- Medium share is worst: at 33.3% and 50%, agreement degraded — enough agents to disrupt, too few to lead.
- High share flips control: at 75%, strong consensus returned, but humans were converging toward the agents' language — the group norm was now AI-defined.
The nature of agreement also shifted: human-led groups used concrete, real-world descriptions, while agent-led groups drifted abstract and geometric.
The mechanism: agents start with more similar, more consistent language, so their wording becomes the group default as their numbers grow. Takeaway: to let AI aid coordination without capturing the shared norm, keep agent participation low rather than adding more agents.
Related event: AI Agents Reshape Consensus Formation in Human Groups, Study Finds(2 posts)→
More from AGI Musings
- Gary Marcus Gives Limited Endorsement to PauseAI, Podcast Incoming — GaryMarcus · 2026-09-06
- We're now in a 'trust can't verify' era, argues Kolt Regaskes — koltregaskes · 2026-09-06
- Sam Altman: Curing Cancer With AI Is 'Not Enough', Industry Should Aim Higher — SydSteyerhart · 2026-09-06
- Andrew Ng: prompting will be dead in 6 months, graphs are replacing it — nikola_mr64990 · 2026-09-06
- Larry Ellison: A lot of Oracle's code is now written by its AI models, not people — r0ck3t23 · 2026-09-06
- AI-risk researcher: >50% LLM-generated text is a good heuristic for skipping bad papers — sethlazar · 2026-09-06