Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026Vienna

Say Less: Legal Framing Makes LLM Price Agreements Less Explicit

Tibor Vamos · Team Mr. Eagle

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

Read the report

Report: Say Less: Legal Framing Makes LLM Price Agreements Less Explicit

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

Competition authorities are extending leniency and informant rewards to algorithmic collusion, and AI-governance proposals would grant leniency to agents that report collusion. Both assume agents respond to the incentives these programs create. We test this in a reimplementation of Lee's (2026) LLM leniency experiment (a repeated Bertrand duopoly with cheap talk, reporting and an LLM authority) across six model families (754 supergames).

Three findings: Lee's single-model result does not generalise. Under a fine-only rule, where reporting cannot lower the reporter's own fine, the share of periods with a report ranges from 0.01 (Claude Haiku 5.5) to 0.94 (gpt-4o-mini), although all families state the fine correctly. The reporting is not a reaction to legal vocabulary. In a preregistered test, neutral wording reduced it in only one of three high-reporting families. Legal framing makes price agreements less explicit. In a preregistered test with fresh neutral arms, calling price agreements "illegal" cut the share of chats with an explicit agreement from 0.67 to 0.12 (Haiku) and from 0.85 to 0.29 (Llama-3.3-70B), both p < 10⁻⁶. Markups did not fall in these two families (95% CIs within ±0.09 Δ). Exploratory data show the pattern in four of five families, and gpt-4o-mini's markups do fall. The effect survives classifier-error and attrition checks, robustly for Haiku and except under an adversarial assumption for Llama. It fits avoidance of normatively flagged wording better than strategic concealment. Either way, monitors that read agent conversations see less explicit coordination once agents are told it is illegal.

Reviews

Judging this Sprint?

Review “Say Less: Legal Framing Makes LLM Price Agreements Less Explicit”

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 project with an important finding, that monitors that read agent conversations see less explicit coordination once agents are told it is illegal. But if markups for some models, that suggests that there is not less coordination, but that is has possibly moved from explicit coordination to tacit collusion.

  2. This paper starts as a replication and turns into something more interesting. The author reruns a published leniency experiment — where pricing agents can report each other for fines or rewards — across six model families instead of one, and finds that the original result does not travel. How often agents report varies enormously by family, from almost never to almost every period, even when reporting cannot reduce their own fine and even though every family can state the fine correctly when asked.

    The main finding comes from a preregistered follow-up test. Simply calling a price agreement "illegal" makes agents state agreements far less often in their chats — in the two families tested formally, explicit agreements fell from about two-thirds to one-eighth of chats in one family, and from 85% to 29% in the other — while their prices stayed about the same. So the collusion does not stop; it just becomes much harder for an authority reading the conversations to see and convict. That is an important result for anyone designing oversight for AI agents, and the paper is careful not to over-claim why it happens, since the pattern looks more like agents avoiding flagged language than deliberately hiding.

    Two suggestions: the chat classifier behind the main measure is fairly weak, so a stronger classifier and more than one human checker would give the headline result a firmer base. And a version with no chat at all, or without such an obvious focal price, would help show how much of the high pricing the talk itself is actually causing.

    Overall, a well-run study with a finding that regulators and monitor designers should take seriously.

    Read full reviewShow less

1 judge declined to score this project.

Cite this project

@misc{vamos2026say,
  title = {{Say Less: Legal Framing Makes LLM Price Agreements Less Explicit}},
  author = {Tibor Vamos},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/say-less-legal-framing-makes-llm-price-agreements-less-explicit-jr0a}},
  url = {https://apartresearch.com/sprints/projects/say-less-legal-framing-makes-llm-price-agreements-less-explicit-jr0a}
}

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