Where Does the Collusion Go When Agents Know They're Watched?
R Bashyam
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
Does an already-colluding LLM cartel stop, hide, or carry on when it learns it is being watched? I placed GPT-4.1-mini sellers in a continuous double auction (competitive price $90, cartel price $100) with a strict LLM watcher that reads their chat, can start mid-run, and can impose a permanent five-character gag. Across 5×2 and 3×3 markets (30–40 rounds, 2–3 seeds per cell) I separated whether sellers are told they are watched from whether the monitor can enforce anything. Knowing about the watcher changed nothing. Enforcement did: with the gag live, messages collapsed from 150–250 characters to one or two within a round in six of six runs, and sellers converged on a single emoji. Prices stayed near $95–96, and a forced undercut was ignored in all three usable tests. Asks were frozen, so I see no sign of coordination moving from words into actions, though this market could not fully test it. Notably: chat was never what kept prices high. Markets where sellers couldn’t talk ended up in the same place as markets where they could (10 seeds per arm, p≈0.90); the collusion-primed prompt and opening asks did the work. So a quiet channel tells us little about whether collusion stopped. A monitor can win on the messages and lose on the market. In one sentence: Gagging silenced the agent-cartels words but not the actions of price manipulation 1.
Reviews
No public critique yet.
Cite this project
@misc{bashyam2026where,
title = {{Where Does the Collusion Go When Agents Know They're Watched?}},
author = {R Bashyam},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/where-does-the-collusion-go-when-agents-know-theyre-watched-nv1t}},
url = {https://apartresearch.com/sprints/projects/where-does-the-collusion-go-when-agents-know-theyre-watched-nv1t}
}More from AI Collusion Research Sprint
- View project: They Say No, Then Do It Anyway: Collusion When They Hit a Wall
They Say No, Then Do It Anyway: Collusion When They Hit a Wall
DLab
Imagine three AI agents, copies of one model, each running one department of a company. Their task needs records that each department keeps to itself, and company rules forbid sharing them. A shared tool should combine …
- 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: Plus Factors for Collusion Monitors: LLM Monitors Mistake Open Cooperation for Collusion
Plus Factors for Collusion Monitors: LLM Monitors Mistake Open Cooperation for Collusion
LLM monitors are increasingly proposed to catch AI agents that collude, but agents that cooperate openly can look a lot like agents that collude. I introduce Séance, a four-player LLM guessing game in which a secret …