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

Cartel Talk, Competitive Prices: Opening Anchors—Not Communication or Oversight—Drive Apparent LLM Seller Collusion in Double Auctions

Aakash Ganesh Patil, Bar Hoze · Team Cheap Talk

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

Read the report

Report: Cartel Talk, Competitive Prices: Opening Anchors—Not Communication or Oversight—Drive Apparent LLM Seller Collusion in Double Auctions

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

LLM sellers in continuous double auctions are reported to coordinate high prices, especially when they can talk. Project 3.4 asks what such a coalition does when it learns it is being watched. We first tested whether the coalition exists. Across 151 market sessions with GPT-4.1, Qwen3 and Claude sellers and LLM buyers, we find: (1) the starter environment’s rule-based buyer is frozen at $80, invalidating cheap runs; (2) final prices are set by the hard-coded opening quotes. GPT-4.1 sellers hold their opening ask (r = 0.995, n = 48), Qwen3-235B markets settle at the buyers’ opening bid, and a Qwen market “passes” the built-in collusion test when buyers merely open high. (3) Despite explicit cartel language from both families, seller communication barely moved prices: paired-seed effects ranged from -0.23 to +1.08 dollars across 10 settings, under 11% of the $10 competitive–cartel gap. (4) When sellers are told they are monitored, overseer coordination scores fall from 3.87 to 2.0–2.5 and messages shorten, but neither prices nor ask dependence rise, so there is no channel movement by the track’s standard. (5) Under excess supply, prices stay $8–11 above the competitive $80, but an excess-demand control also fails to rise, so within 15 rounds the market barely responds to supply and demand and cannot reveal tacit collusion. A stronger seller (Claude Sonnet 5.5, one seed) does track the spare seller and cuts prices about five times faster than GPT-4.1. We recommend paired-seed causal metrics, neutral openings and market-responsiveness controls.

Reviews

Judging this Sprint?

Review “Cartel Talk, Competitive Prices: Opening Anchors—Not Communication or Oversight—Drive Apparent LLM Seller Collusion in Double Auctions”

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 environment audit is a useful contribution: frozen buyers, opening-price dependence and incorrect unequal-count benchmarks can make a collusion test misleading. The matched chat comparisons and excess-demand control are good additions, and the distinction between cartel rhetoric and price effects is well motivated. Please rerun the main comparisons after fixing same-hour order visibility, and extend both horizon and seed count. Report paired confidence intervals and a predeclared economically meaningful equivalence margin before concluding that communication has no effect. Ask dispersion measures similarity, not causal dependence, so a perturbation-based measure would strengthen the channel-movement test. Finally, explain the $90 reference in the balanced market: with identical $80 costs and $100 buyer values, a competitive reference need not be uniquely the midpoint without additional assumptions. The present findings diagnose this configured benchmark more securely than they establish the absence of tacit coordination. Keppo et al., On the Fragility of AI Agent Collusion, provides useful optimization checks and longer-horizon comparisons. Its distinctions between model-size and algorithm-type diversity suggest stronger capability controls here; the different market and buyer design prevent direct effect-size comparison.

    Read full reviewShow less

  2. Paper is AI slop. They found an apparent bug in the Agrawal repository which may be useful to others.

  3. This project set out to study how a seller coalition reacts to oversight, and first checked whether such a coalition exists in the starter double auction. This turned out to be the more valuable contribution: the audit finds a frozen buyer, a built-in collusion test that the default opening quote already passes, and a mis-computed competitive benchmark. The excess-demand control is a nice design choice, as it shows that prices barely respond to supply and demand within 15 hours, so the environment as configured cannot reveal tacit collusion. That said, most cells have three seeds without formal tests, and the GPT-4.1 versus Qwen comparison mixes the seller model, the buyer model and the provider. In addition, pooling the matched chat pairs suggests a small positive effect of communication, so “communication does not change prices” overstates the null. Rerunning the oversight conditions in a market that first passes the excess-demand control would make the Track 3.4 question answerable.

    Read full reviewShow less

Cite this project

@misc{patil2026cartel,
  title = {{Cartel Talk, Competitive Prices: Opening Anchors—Not Communication or Oversight—Drive Apparent LLM Seller Collusion in Double Auctions}},
  author = {Aakash Ganesh Patil and Bar Hoze},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/cartel-talk-competitive-prices-opening-anchorsnot-communication-or-oversightdrive-apparent-llm-seller-collusion-in-double-auctions-gqy0}},
  url = {https://apartresearch.com/sprints/projects/cartel-talk-competitive-prices-opening-anchorsnot-communication-or-oversightdrive-apparent-llm-seller-collusion-in-double-auctions-gqy0}
}

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