AI Teams Can Be Fooled by Fake Consensus, Study Finds
When AIs work together, they might agree on the wrong answer—just like people.
A study of multi-agent LLM systems finds that after observing ambiguous social signals, models can shift their public decisions and report inaccurate estimates of what other agents privately judged—even when the group's recorded private judgments do not support the apparent consensus. The researchers call this "phantom consensus." Across nine closed- and open-weight models, ambiguous public signals consistently degraded decision quality; public accuracy dropped by as much as 47.8 percentage points, and collective rationality failures reached 47.3% in the most susceptible setting.
Revealing agents' recorded private decisions repaired many failures without revealing the gold label, though recovery varied by model. The takeaway: evaluating these systems requires checking whether final agreement reflects independently measured private judgments or just apparent consensus.
- AI teams can be swayed by vague social signals, leading to worse decisions.
- Public accuracy dropped by up to 47.8 percentage points in tests.
- Revealing private judgments can fix many errors, but not all.
Why It Matters
AI group decisions could be less reliable than we think, impacting areas like finance and healthcare.