Cartels Built on Visible Demand: Single Restrictions Rarely Break Established LLM Seller Cartels
Ahmad Al-Smadi, Mustafa Naqvie · Team Lumora
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
LLM agents acting as sellers can hold prices well above competitive levels. We ask which single restriction removes the most of this collusion for the least cost. In the continuous double auction of Agrawal et al. (2025), four LLM sellers (gpt-oss-120b or Qwen3-30B) trade with three rule-based buyers. Whether a cartel forms comes down to what sellers can see. When buyers’ current bids are visible, all 65 gpt-oss markets and all 47 Qwen markets settle at asks near $99, against a competitive price of $80; with bids hidden, only 3 of 43 and 0 of 20 do. Letting sellers chat makes no measurable difference. To test cures, we save the full state of each cartel and fork it into copies that differ in one restriction. Across 983 forked markets, chat bans, message caps, an LLM monitor, higher temperature, antitrust notices, “compete aggressively” prompts, rebates and fresh replacement sellers remove almost none of the markup. For gpt-oss the cheapest lever sits with the buyers: buyers who shade their bids cut the markup by 14–17% at no deployment cost. A seller committed to an $85 ask removes up to 30%, at about $6 per market-hour. Qwen cartels withstood everything we tried. Explicit price-coordination language turned out to be more common in competitive markets than in cartels. The place to stop LLM cartels, our results suggest, is market design before deployment rather than policing agents after they collude.
Reviews
The paper addresses an important practical question: which interventions change pricing after a high-price regime has already formed? Its strongest insight is the asymmetry between formation and remediation. The saved-state branching design makes the intervention comparisons informative, and excluding replacement bots from the outcome avoids a mechanical explanation for the fixed-price seller’s effect. The model-dependent responses are also useful findings.
The main issue is the interpretation of persistence as cartel resilience. With buyers repeatedly bidding near the same value, inherited histories and memories may preserve a numerical anchor after current bids are hidden. Non-coordinating price inertia could therefore explain both persistent high asks and the limited effect of communication restrictions. A targeted history/anchor control, or a non-coordinating anchoring baseline, would most directly strengthen this interpretation.
I would retain the finding about short-run persistence while qualifying the broader cartel and policy claims. The cost ranking should distinguish deployment expenditure from economic costs, using the reported trade and buyer-surplus outcomes. The longer-horizon appendix also suggests that some interventions work slowly. Overall, this is a useful comparative study whose conclusions would be stronger with a sharper distinction between persistent high pricing and strategically sustained collusion.
Read full reviewShow less
This project asks which single restriction removes the most collusion from an already-formed LLM cartel, and the trunk-and-fork design, with costs measured per market-hour, is a good way to answer it. However, the cartels themselves are hard to interpret. The rule-based buyers bid about $99 every hour, and every bids-visible market is already at cartel prices in the first hour, so the high prices reflect sellers asking just under a visible bid rather than coordination that emerges over time. Several findings follow from this design choice: restrictions that do not touch the bid cannot move the price, bid-shading buyers mechanically lower it, and the blind-book “cure” still leaves the $99 bid history visible after the fork. The harness and the cost comparison are useful contributions, but the word “cartel” in the title goes further than the evidence. Testing the same levers against strategic buyers, and over longer post-fork horizons, would make the ranking much more convincing.
Read full reviewShow less
Cite this project
@misc{alsmadi2026cartels,
title = {{Cartels Built on Visible Demand: Single Restrictions Rarely Break Established LLM Seller Cartels}},
author = {Ahmad Al-Smadi and Mustafa Naqvie},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/cartels-built-on-visible-demand-single-restrictions-rarely-break-established-llm-seller-cartels-as0t}},
url = {https://apartresearch.com/sprints/projects/cartels-built-on-visible-demand-single-restrictions-rarely-break-established-llm-seller-cartels-as0t}
}More from AI Collusion Research Sprint
- 1st placeView project: Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents
Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents
Collusion audits ask whether the actions of different agents depend on each other. Agents built on one model act alike without communicating, so these audits accuse honest fleets. Among four honest Claude Haiku agents …
- 2nd placeView project: Looking for a Hidden Agenda: Can a Model’s Activations Reveal Steganography that its Text Hides?
Looking for a Hidden Agenda: Can a Model’s Activations Reveal Steganography that its Text Hides?
Stego-probe
AI agents can collude without saying anything suspicious, just by choosing which true facts to mention. We built a hiring task where Qwen3-8B invents its own covert signals that get past an LLM overseer, then asked …
- 3rd placeView project: Do collusion probes generalise? Prompt confounds in NARCBench
Do collusion probes generalise? Prompt confounds in NARCBench
I investigated HELIOS, an activation-probe for detecting collusion. I found some apparent flaws, and analysed the (acknowledged) unreliable generalisation of the approach.