Bias Amplification in Multi-Agent AI Systems
A new study on arXiv (2605.28098) examines how biases in individual AI agents can amplify system-wide unfairness in multi-agent systems. The researchers propose a metric called Favor Bias Strength (FBS) to quantify bias alteration, distinguishing between favored-group uplift and disfavored-group suppression. Using prompts to expose agents to group-favoring bias, they tested multiple agent designs, benchmarks, and large language models. Results show that uniformly biased agents can elevate system-wide bias beyond the additive sum of individual biases, highlighting challenges for fairness preservation in collaborative AI.
Key facts
- arXiv paper 2605.28098 examines bias amplification in multi-agent systems.
- Proposes Favor Bias Strength (FBS) metric to measure bias alteration.
- FBS decomposes bias into favored-group uplift and disfavored-group suppression.
- Study uses prompts to expose agents to group-favoring bias.
- Tests multiple agent designs, benchmarks, and large language models.
- Uniformly biased agents can elevate system-wide bias beyond additive sum.
- Fairness preservation through bias reduction remains challenging.
- Multi-agent systems are increasingly deployed for various tasks.
Entities
Institutions
- arXiv