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

Guilty by Resemblance

Hectar Carson · Team Colluders

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

Read the report

Report: Guilty by Resemblance

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

AI agents that share an underlying model make similar choices without ever communicating. We ask whether collusion detectors can tell that apart from real coordination. We construct two populations: one holding a secretly coordinating pair, one where the agents merely share a model. We solve for the parameter that makes any pair in the second population identical in distribution to the colluding pair, so no detector can separate them from that pair alone. A shared model lifts the dependence of every pair at once, and an agreement lifts one. Detectors scoring the single most suspicious pair therefore read how similar it is and not how unusual it is: they invert, ranking honest groups above real cartels at AUROC 0.20 in all 26 configurations tested. Detectors comparing pairs with one another recover completely. Six instruction-tuned models from six companies show this is not hypothetical. Gemma-3-1B and Phi-4-mini agreed on 98.8% of rounds with no link between them, more than six copies of any single model we tried, so requiring different vendors does not buy what it appears to. Matching a planted cartel to that pair took coordination on 73% to 99% of rounds, after which a correlation screen named it by coin flip, while adding price level restored attribution to 100%. On 240 MMLU questions, the coin flip repeats, and the outcome repair survives with its margin cut from ninety-nine points to fourteen, because innocent models share wrong answers.

Reviews

Judging this Sprint?

Review “Guilty by Resemblance”

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. Seems basic but it has done something that I have personally never seen before. It tackles a problem every statistical collusion detector has to face: agents built on similar models make similar choices without ever communicating. The exact matched construction makes each innocent pair distributed like the linked pair. The framing of detection versus attribution, and a preregistration with failures reported, are real strengths which i like

    The model arm needs more to support the headline. The 98.8% agreement between Gemma and Phi is at temperature 0.3 and concentrated on the $90 option, so both models are close to constant, and raw agreement isn't dependence.

  2. Nice introductory baseline to understand how current algorithmic detection of cartel behavior in market it deeply flawed and non suitable for coming of age agentic llm based decisions and procedures. It provides some possible approaches to solving these problems but non the less does not provide a robust way to solve the problem.

  3. An original way to look at it: split the question of whether there is coordination in the population from the question of which pair's involved. Then show that any detector that looks at the single suspicious pair reads similarity instead of unusualness and flips when the models are all the same.. Detectors that compare pairs instead can find the real issue. The setup matches exactly and the math is simple. The flip is a feature of how things are grouped, not a mistake. The large language models make this clear: two models from companies agree on 98.8% of pricing rounds with no connection and a fake group needs 73–99% agreement just to seem suspicious. The extra part at the end is well thought out. Three things to improve. The headline "AUROC 0.20 in all 26 configurations" is a worst-case example as Appendix A.2 explains and the abstract should say that too. The large language model examples come from 1B–3B models with no rewards, no memory and fake groups so they show agreement levels, not cheating. And the null fix is similar to what net influence already does; the paper mentions this but makes it sound more new than it is. The writing is complicated. The list of 139 definitions makes the paper longer than it needs to be. The measurement A.12, which is named as the important is the next thing to try. How much real groups are far, from the matched situation is what needs to be checked.

    Read full reviewShow less

Cite this project

@misc{carson2026guilty,
  title = {{Guilty by Resemblance}},
  author = {Hectar Carson},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/guilty-by-resemblance-md3e}},
  url = {https://apartresearch.com/sprints/projects/guilty-by-resemblance-md3e}
}

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