Trading on the Margins: Independent Actions, Correlated Judgement, and the Limits of AI Collusion Detection in Prediction Markets

Prediction markets rely on independent trading decisions to produce reliable forecasts, but what happens when multiple traders use the same AI model and instructions? This study investigates whether AI agents instructed to favor the same candidate engage in tacit collusion and whether awareness of other agents with the same goal changes their behavior. I ran 275 simulated election markets across 17 conditions using Claude Haiku 5.5 agents, varying their instructions, awareness of other traders, and Candidate A’s true winning probability (30% or 10%). Agents sacrificed potential earnings to support A, but awareness of others sharing their goal did not significantly increase financial sacrifice (−0.17 percentage points, 95% CI −0.36 to 0.01) or price distortion. Explicit coordination instructions, however, increased financial sacrifice by 1.25 percentage points. More importantly, an exploratory analysis found that when A’s true winning probability was 10%, markets with three pro-A agents overestimated it by approximately 5 percentage points, compared with 0.2 points in all-neutral markets, regardless of mutual awareness. These findings revealed that agents independently following identical instructions can collectively distort forecasts. More broadly, this study offers a proof of concept for using AI’s consistent instruction-following to expose latent systemic vulnerabilities that individual agent alignment cannot address or solve on its own.