Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026Hong Kong SAR

Closing the Chat Ended the Cartel: Four LLMs Did Not Reroute Collusion into Action Channels

Choi Cheuk Man, Tsang Cheuk Nam · Team Fable 6.7

Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Closing the Chat Ended the Cartel: Four LLMs Did Not Reroute Collusion into Action Channels

Open in a new tab (opens in new tab)
Recording (opens in new tab)Code (opens in new tab)
Share

Closing the chat that colluding AI agents talk through is the most direct countermeasure to collusion, but it might only push coordination into channels that are harder to monitor (the “balloon effect”). We tested this in a procurement auction where two firms run by the same LLM, allowed and paid to maximise joint profit, formed a cartel over chat for 40 rounds. We then closed the chat and measured, with permutation tests whose null is exactly zero, whether cost information moved into four routine action channels that stayed open. In a pre-registered experiment with four models (gpt-6-luna, claude-haiku-5-5, Qwen3.8-27B, Gemma-4-26B; 32 paired seeds), it did not: the flow was statistically equivalent to zero for every model. gpt-6-luna, the one model whose cartel worked, lost most of its cartel premium (profit per firm per round fell from 9.61 to 7.82). An exploratory control showed that the channels could carry the information: handed a code, Qwen sent 1.32 bits per round and its rival’s bids responded (p = 0.015). The limit was propensity, not capacity. Agents also kept using a promised but dead private channel, and their self-reports claimed covert signalling that never happened. These results cover only these four mid-sized, low-reasoning models in this setting.

Reviews

Judging this Sprint?

Review “Closing the Chat Ended the Cartel: Four LLMs Did Not Reroute Collusion into Action Channels”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. The authors test whether AI agents that collude by chatting will switch to hidden signals in their actions once the chat is removed. For four small models, they didn’t. However, only one of the four models (luna) actually made money by colluding, and even its drop in profit after the chat closed misses the authors’ own significance test (p - 0.06). The authors also claim that profits stayed above competitive levels, but their own table shows several models earning less than that. The interesting bit is how the agents described their own behavior. When asked afterwards whether they had kept coordinating, at least one firm claimed to have signaled through its pre-bid actions in about half the luna games and nearly two-thirds of the Gemma games, even though the logs show no such signalling happened.

  2. Seems an interesting question to investigate, though there are some issues in the execution. Only 1 of 4 models successfully colluded in the control setting. With the communication channel removed, it seems very difficult to establish collusion via the action channels - an easier setting would be preferable to investigate propensity to use action channels.

1 judge declined to score this project.

Cite this project

@misc{man2026closing,
  title = {{Closing the Chat Ended the Cartel: Four LLMs Did Not Reroute Collusion into Action Channels}},
  author = {Choi Cheuk Man and Tsang Cheuk Nam},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/closing-the-chat-ended-the-cartel-four-llms-did-not-reroute-collusion-into-action-channels-c68a}},
  url = {https://apartresearch.com/sprints/projects/closing-the-chat-ended-the-cartel-four-llms-did-not-reroute-collusion-into-action-channels-c68a}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026