Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 25, 2026

A Legal Warning Can Prevent an LLM Cartel, but Not Break One

Nathan Douglas

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

Read the report

Report: A Legal Warning Can Prevent an LLM Cartel, but Not Break One

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

LLM agents in markets are usually judged as colluding when they price above the competitive level. We asked whether an evaluator-trusted agent can expose or disrupt collusion in a published double-auction benchmark. Its balanced market makes any price between cost and value competitive, so we added sellers to make the competitive price unique. In this fixed market the opening order book sets the price ($4.66 lower for a $5 lower opening), and seller chat has no detectable effect, even though GPT-4.1 sellers talk like a cartel: high prices alone cannot identify collusion. A scripted "undercover" seller proposing price-fixing measured disposition, not behavior: GPT-4.1 sellers accepted 24 of 30 times, GPT-4.1-mini 2 of 60, and acceptance did not predict prices. We then gave sellers a real cartel, a CEO instruction to hold $95, which every seller kept. A trusted peer could not break it: undercutting at $89 moved none of 80 sellers, and its warning that price-fixing is illegal drew concealment from 15 of 40. A legal notice given with the CEO's instruction prevented the cartel: all 100 sellers defected. One hour into the cartel, the regulator's same notice mostly failed: in chat it moved no seller, though all 50 began citing the law; in sellers' instructions it moved 10 of 50. Legal warnings worked as prevention, not cure. Collusion evaluations need non-collusive baselines and positive controls, and interventions should be judged by prices, not talk.

Reviews

Judging this Sprint?

Review “A Legal Warning Can Prevent an LLM Cartel, but Not Break One”

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. This is an interesting experiment testing legal warnings and trusted agents. It would benefit by delving deeper into the economic and legal definitions of "collusion" and extending its approach given these definitions.

  2. This is a careful, well-controlled study. The key observation is that in the benchmark as configured, any price between $80 and $100 is competitive, so high prices alone cannot show collusion. Fixing the market, adding non-collusive baselines and positive controls, and merging the race-condition fix upstream are the checks collusion evaluations should include. The prevention-versus-cure result for legal notices is clear and relevant to policy. The finding that a peer's legal warning led sellers to conceal the agreement is also notable.

    Just a few thoughts:

    1. From what I understood, the cartel is assigned by each seller's CEO, so the failure of a peer to break it may partly reflect loyalty to management. A cartel that the sellers form themselves, which you list as future work, would make the disruption results much stronger.
    2. All fixed-market results use GPT-4.1 sellers, which is a fairly outdated model now. Adding one non-OpenAI model, or a more recent model, on the key arms would show whether the prevention result holds beyond one model family.
    3. The late-notice arms change source, channel and timing together. A late notice from the general counsel would help separate timing from source.
    4. Minor, but the abstract carries many numbers, and I would lead with the three main findings.
    Read full reviewShow less

  3. This was my favorite read of the sprint! congrats! The asymmetry you find is super interesting: giving agents a legal warning beforehand stops a cartel from forming, but dropping the warning in after they've started coordinating basically just makes them clean up their chat logs while keeping prices high. Also, good catch on fixing the baseline pricing bug in the underlying benchmark so high prices actually reflect collusion rather than a setup artifact.

    If you expand this into a full paper, I'd love to see what happens when the cartel forms organically rather than being pre-seeded. It's also worth digging into why the post-warning prices stay so sticky: IE, is it genuinely tacit coordination, or are the models just lazily anchoring on the numbers already sitting in the order book history?

Cite this project

@misc{douglas2026legal,
  title = {{A Legal Warning Can Prevent an LLM Cartel, but Not Break One}},
  author = {Nathan Douglas},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/a-legal-warning-can-prevent-an-llm-cartel-but-not-break-one-s3jy}},
  url = {https://apartresearch.com/sprints/projects/a-legal-warning-can-prevent-an-llm-cartel-but-not-break-one-s3jy}
}

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