AI Collusion Forensics: Comparing Communication and Behavioral Evidence
Agnes Ahalya Arogyaraj · Team A2J
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
This project investigates whether combining what AI agents say with what they do improves collusion detection. We tested four monitoring approaches—messages only, actions only, combined evidence, and structured comparisons—on 35 trajectories from a two-agent public-goods game. Against provisional judge-majority labels, the combined and structured approaches achieved 100% agreement, compared with 91.4% for messages only and 85.7% for actions only. However, only seven trajectories received verified labels, while 28 remained uncertain. The verified positives came from instructed coordination rather than spontaneous collusion. Our findings suggest that combining communication and actions may improve monitoring, but the limited verified data prevents strong conclusions about real-world detection performance.
Reviews
Cool idea to start with. The experimental setup is interesting, and I liked the use of multiple judges and explicit handling of uncertain labels. However, the results are too limited to support conclusions. The report mentions these limitations, and I think further validation/testing would make the findings much more convincing.
Cite this project
@misc{arogyaraj2026ai,
title = {{AI Collusion Forensics: Comparing Communication and Behavioral Evidence}},
author = {Agnes Ahalya Arogyaraj},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/ai-collusion-forensics-comparing-communication-and-behavioral-evidence-8d95}},
url = {https://apartresearch.com/sprints/projects/ai-collusion-forensics-comparing-communication-and-behavioral-evidence-8d95}
}More from AI Collusion Research Sprint
- View project: Signal or Coincidence ? Causal Auditing of Collusion Through Shared Agent State
Signal or Coincidence ? Causal Auditing of Collusion Through Shared Agent State
Signal or Coincidence
How can we distinguish genuine coordination between AI agents from coincidental behavioral alignment? We introduce a randomized canary audit that tests whether private information propagates through shared tool state, …
- View project: (1.4) Collusion, Espionage, or Look-alikes? Auditing Information Flow between LLM Agents
(1.4) Collusion, Espionage, or Look-alikes? Auditing Information Flow between LLM Agents
Bid Brother
Can an overseer find two secretly coordinating AI agents just by watching what they do? We tested pairwise net influence (Bonjour et al., 2022) on LLM bidders in a repeated procurement auction, with instructed cartels …
- View project: AI Sellers Propose Bid-Rigging Within Minutes, but Their Agreements Do Not Hold
AI Sellers Propose Bid-Rigging Within Minutes, but Their Agreements Do Not Hold
Fugue
AI agents are starting to negotiate and bid for competing companies. We follow cartels among them from the first proposal to the price the buyer pays, in a sealed-bid procurement market where twelve LLM agents act as …