Shared inputs cause bias in random choices by ai reasoning agents
Shared and structured inputs undermine collective random choice by reasoning AI agents
Artificial Intelligence
Summary
Artificial intelligence agents are often used to make random selections, like choosing requests for review. However, this paper shows that when these AI agents base their choices on shared or structured information such as parts of identifiers or timestamps, their selections become predictable and correlated. The authors tested several AI models and found that even explicit instructions to randomize didn’t fully prevent this issue. This reveals hidden risks when relying on AI for truly random decisions, especially in auditing and resource allocation.
What this means in practice
- •For audit teams: Detect and mitigate input-dependent biases in AI-driven random selection during resource audits to improve oversight reliability.
- •For customer service managers: Ensure fair and unpredictable selection of customer requests for human review using AI that considers correlation effects in identifiers.
Authors
Takahiro Ezaki, Naoto Imura, Katsuhiro Nishinari
Abstract
Random selection is widely used in resource allocation and auditing, making reliable implementation essential for AI-agent systems. Behavioural tests across six reasoning models uncovered threshold and divisibility rules used in identifier-based choices. For threshold-following GPT-6 Sol and Gemini 3.8 Flash, single-agent measurements prospectively predicted correlated participation under shared identifiers and biased participation under distinct identifiers with common timestamp bits. Changing dates, formats and identifier labels revealed when these predictions held. Explicit instructions to randomize independently reduced but did not eliminate shared-input correlation. To test implications for oversight, we asked four models to select customer requests randomly for human review. GPT-6 Sol approached the target rate while selecting predictably from identifiers; the others rarely selected requests. All four closely followed supplied random draws. These findings expose collective and audit vulnerabilities that selection rates alone miss, making input-dependent bias, correlation and predictability central targets for agent evaluation.