New arXiv paper proves AI collusion is invisible to price audits

A new arXiv paper (2607.26385) from Xin Xu, Chengrui Wu, Jiayu Lu, Kaizhen Tan, Siru Tao, and Hanzhe Hong proves a fundamental blind spot in antitrust enforcement against algorithmic collusion. The researchers show that if bidding agents coordinate only through the joint distribution of their unexplained bid components while keeping each agent's marginal distribution exactly competitive, then any audit that looks at a single agent's price history has detection power equal to its false-positive rate — no matter how much data is collected. This means price-level audits are blind by construction, not because of insufficient data.

To demonstrate real-world relevance, the team tested 20 language models from 19 independent developers, using three deployment prompts each. They found a residual correlation of +0.053 between two deployments of the same model versus only +0.0001 across different models (95% CI clustered by developer: [0.030, 0.078]). This correlation decreased as sampling temperature increased (p=0.002), suggesting a potential mitigation. On 24 days of Ethereum block-building auction data covering 77,684 bids from 39 bidders, the honest population was so interdependent that a screen with a 5% false-positive rate needed a correlation floor of +0.50 to +0.81 — 20-32 times the family-wise sampling threshold. The paper concludes that since lawful multi-identity operation and conspiracy are behaviorally indistinguishable, the practical regulatory target isn't detection but counting: resolving 40 bidding identities into 23 operators raises the Herfindahl index by 247.5%, and adding behavioral clusters reaches 324.5%.