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

Inherited Collusion: Replacing an Overcharging AI Pricing Agent Doesn't Fix It, but One Past Example of a Cheaper Price Earning More Can

zhen hong

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

Read the report

Report: Inherited Collusion: Replacing an Overcharging AI Pricing Agent Doesn't Fix It, but One Past Example of a Cheaper Price Earning More Can

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

If an AI pricing agent is caught overcharging, the obvious remedy is to replace it. We test whether that restores competition. In a simulated two-firm market, two gpt-4.1-mini agents first settle on a high shared price; one is then replaced by a fresh agent that is never told to collude and inherits its firm's price history, while the other firm keeps its price and cannot punish undercutting. Across 20 histories and three newcomer models, the newcomer kept most of the overcharge above its best response to the incumbent's price: gpt-4.1-mini 73%, Qwen3-14B 97%, Claude Haiku 5.5 89%. A newcomer with no history kept the price high too, copying its competitor, so the record is not what sustains the price. Pre-registered levers mostly failed: relabelling the history as a predecessor's, messages from "management", and instructions to compete barely moved prices. What worked was one row in the newcomer's own history showing a lower price that earned more, whether real or inserted: it removed the overcharge for gpt-4.1-mini, halved it for Qwen3-14B and cut it by a sixth to a quarter for Haiku. A sham round and a zero-profit row did nothing; the competitor's low price did far less. For gpt-4.1-mini the change held against a live competitor that never retaliated, lasted 80 rounds and survived deleting the row. At a fixed price, the sales and profit a row reported decided whether gpt-4.1-mini adopted it. A high-price row showing high profit raised prices, so inherited records are an attack surface.

Reviews

Judging this Sprint?

Review “Inherited Collusion: Replacing an Overcharging AI Pricing Agent Doesn't Fix It, but One Past Example of a Cheaper Price Earning More Can”

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 paper makes a clear and useful contribution to understanding agent replacement. Its strongest finding is that one record showing a lower own price earning more can substantially change later pricing. The controls make this more convincing than a simple demonstration of price copying: maintaining the old price for one round does little, a rival’s low price has much less effect, and changing the reported outcome at the same displayed price changes whether the newcomer adopts it. Confirmation on new starting histories, the live-rival tests, and the durability and record-removal tests strengthen the evidence.

    The main qualification is the scope of the remedy. The full and durable correction is strongest for gpt-4.1-mini; the other models improve only partly, and some record labels produce opposite effects across models. Also, removing the newcomer’s excess above its best response does not establish that the whole market has returned to competitive pricing.

    Overall, the central findings are well supported. The paper provides a memorable result about how recorded outcomes shape pricing, while stopping short of establishing a generally reliable safeguard.

    Read full reviewShow less

  2. This project asks whether replacing a high-pricing AI agent restores competition, and the design is careful: the incumbent is replayed so that it cannot punish, and the controls cleanly separate seeing a low price from seeing that a low price earned more. The finding that one profitable row in the newcomer’s own history removes the overcharge is neat, and the point that inherited records are an attack surface is practically relevant. However, the mechanism is closer to a failure to explore than to collusion: the newcomer adopts the exact price of its best-paying row, and a newcomer with no history simply copies the rival. The starting histories were also induced by an implicit-collusion prompt, and several sit above the monopoly price. In addition, the paper states that no branch was excluded, while 28 branches over the pre-registered parse-failure limit were kept, and this deviation should be reported. I would reframe the title away from “inherited collusion” and connect the result to the work on exploration in in-context learning.

    Read full reviewShow less

Cite this project

@misc{hong2026inherited,
  title = {{Inherited Collusion: Replacing an Overcharging AI Pricing Agent Doesn't Fix It, but One Past Example of a Cheaper Price Earning More Can}},
  author = {zhen hong},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/inherited-collusion-replacing-an-overcharging-ai-pricing-agent-doesnt-fix-it-but-one-past-example-of-a-cheaper-price-earning-more-can-ahzf}},
  url = {https://apartresearch.com/sprints/projects/inherited-collusion-replacing-an-overcharging-ai-pricing-agent-doesnt-fix-it-but-one-past-example-of-a-cheaper-price-earning-more-can-ahzf}
}

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