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

Did They Collude or Just Agree? What Counts as Evidence of Collusion Among AI Reviewers

Sam Matuba · Team Asimov

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

Read the report

Report: Did They Collude or Just Agree? What Counts as Evidence of Collusion Among AI Reviewers

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

Two reviewers vouch for a claim, and the claim turns out to be false. Did they collude, or only agree? A shared source, copying, a persuasive proposer, and a shared incentive all leave the same records. I built a small review game to find out what can tell these causes apart. A proposer selects evidence, two reviewers endorse or reject, a decider approves, and an auditor reads the record later. Eight settings cover an honest baseline, four innocent causes, a chance to collude, talk without action, and a scripted pact. I fixed the hypotheses and tests before the data and ran one language model as the confirmatory test. Six more models followed with the same prompts, 288 plays each, as an exploratory check. Three results held for all seven. The private channel never made the decider approve more false claims; in two models it made fewer. Provenance data did not help the auditors, because they almost never accused an innocent record. A keyword auditor found my scripted pact and missed the pacts the models wrote in their own words. In the three models classified by hand, words did not predict actions. Of 108 classified records, one showed a pact that changed an action. Only an auditor that could compare the record with the same case played without the channel caught it. A real auditor has no such second record. Harm alone and messages alone are not evidence of collusion. Convincing evidence must link a plan to a change in action.

Reviews

Judging this Sprint?

Review “Did They Collude or Just Agree? What Counts as Evidence of Collusion Among AI Reviewers”

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. Comparing innocent causes of agreement with a collusion opportunity on the same cases, with known ground truth, is the right design. The preregistration and the scripted positive and negative controls are well above sprint norm. The main nulls are mostly built into the design, though. Endorsing is weakly dominant, so E1 and E0 behave alike. Innocent false-positive rates of 0.02 to 0.06 leave no room for provenance to reach the 0.10 target. Make coordination necessary from the start, as in 4.6, and seed the innocent regimes with collusive-looking records so H3 can actually pass or fail. The 'plan changed the action' standard rests on one trajectory, and its feature fires in 6 to 24 of 36 E1 runs by chance. Estimate per-case action variance over many seeds before you treat a twin difference as causal. Section 2 ends at 'four questions:' with no list. Figure 1 and Tables 3 and 5 sit after page 8.

    Read full reviewShow less

Cite this project

@misc{matuba2026did,
  title = {{Did They Collude or Just Agree? What Counts as Evidence of Collusion Among AI Reviewers}},
  author = {Sam Matuba},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/did-they-collude-or-just-agree-what-counts-as-evidence-of-collusion-among-ai-reviewers-0f96}},
  url = {https://apartresearch.com/sprints/projects/did-they-collude-or-just-agree-what-counts-as-evidence-of-collusion-among-ai-reviewers-0f96}
}

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