Emergent Bias and Fairness in Multi-Agent Decision Systems

TL;DR

The paper investigates emergent bias and fairness in multi-agent decision systems applied to financial tabular tasks. Using simulation-based experiments across credit scoring and income estimation, it shows that collective interactions among agents can produce bias patterns not predictable from individual components, underscoring the need for holistic MAS evaluation. Key findings reveal modest median bias changes but pronounced tail risks, implying substantial regulatory and risk-management implications for deploying MAS in finance. The work advocates system-level fairness assessment and calls for new benchmarks and mitigation strategies tailored to multi-agent interaction dynamics.

Abstract

Multi-agent systems have demonstrated the ability to improve performance on a variety of predictive tasks by leveraging collaborative decision making. However, the lack of effective evaluation methodologies has made it difficult to estimate the risk of bias, making deployment of such systems unsafe in high stakes domains such as consumer finance, where biased decisions can translate directly into regulatory breaches and financial loss. To address this challenge, we need to develop fairness evaluation methodologies for multi-agent predictive systems and measure the fairness characteristics of these systems in the financial tabular domain. Examining fairness metrics using large-scale simulations across diverse multi-agent configurations, with varying communication and collaboration mechanisms, we reveal patterns of emergent bias in financial decision-making that cannot be traced to individual agent components, indicating that multi-agent systems may exhibit genuinely collective behaviors. Our findings highlight that fairness risks in financial multi-agent systems represent a significant component of model risk, with tangible impacts on tasks such as credit scoring and income estimation. We advocate that multi-agent decision systems must be evaluated as holistic entities rather than through reductionist analyses of their constituent components.

Figures (6)

• Figure 1: Discussion paradigms for a multi-agent system of size $N$. (a) illustrates the Memory setup Yin2023ExchangeofThoughtEL that corresponds to a fully connected graph, where agents are provided with the output of all other agents and iteratively suggest refinements. (b) shows the Collective Refinement setup Kaesberg2025VotingOC, where each agent first generates a draft response independently, before refining their answer based on the drafts of all other $N-1$ agents.
• Figure 2: Agents are prompted to take part in a multi-agent debate as shown in this prompt example for the Adult Income dataset.
• Figure 3: Distribution of bias changes for multi-agent systems relative to single-agent baselines across simulations of the Adult Income dataset. Bias is measured by the metric shown in each figure. Positive values indicate higher bias in multi-agent systems. The distributions exhibit long positive tails indicating significant bias increases in some cases, with slight negative skews showing frequent but modest bias reductions.
• Figure 4: Distribution of bias changes for multi-agent systems relative to single-agent baselines across simulations of the German Credit Risk dataset. Bias is measured by the metric shown in each figure. Positive values indicate higher bias in multi-agent systems. The distributions exhibit long positive tails indicating significant bias increases in some cases, with slight negative skews showing frequent but modest bias reductions.
• Figure 5: An example single-agent prompt for the Adult Income dataset.