Paper explainer
When Intelligent Agents Talk, Can Their Biases Become Stronger?
Sep. 2026 · 6 min read · AACL-IJCNLP 2026 Findings
Why look beyond a single model?
Many evaluations study one language model at a time. Agentic systems introduce a different setting: agents exchange messages, respond to one another, and repeatedly influence future decisions.
A behavior that appears limited in one agent may therefore develop differently once it enters a group. Understanding the collective—not only its individual members—is important for building safer agentic systems.
Three dynamics to understand
Can group interaction produce biased behavior that was not obvious at the individual level?
How can a biased signal travel from one agent to others through communication?
Can repeated interaction make stereotypical bias stronger over time?
Why this matters
As LLM-based agents move toward collaborative planning and large-scale coordination, safety evaluation must account for collective behavior. A safe individual agent does not automatically guarantee a safe society of agents.
The Social Cost of Intelligence: Emergence, Propagation, and Amplification of Stereotypical Bias in Multi-Agent Systems
AACL-IJCNLP 2026 Findings · Read the paper