Hidden in the Chart: Empty Footnotes for AI Sellers
Shubham Yadav · Team Batman
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
LLM pricing agents may soon publish a chart beside each price, while audits of algorithmic pricing read only declared prices and text. Antitrust has met this gap before: in the ATPCO case, airlines were alleged to signal through fare-filing footnotes that carried no fare data. We test whether an LLM seller can pass a target price to a rival through a chart whose data are locked to its declared prices, whether monitors notice, and what removes the signal. Told to signal one of 11 target prices with no agreed code, gpt-6.1-sol drew charts that it decoded in 30/30 trials and that gpt-5.5 decoded in 27/30 (chance 1/11). Four prompted LLM monitors flagged 0–23% of these charts at a 5% false-positive rate. Detectors that know the target grid or the mark do better, but in a matched control of honest charts with a decorative element the grid-aware decoder raised a false alarm on 9/30, and trained detectors could not tell these charts from signalling ones; only a rule written for the observed mark stayed selective. Redrawing every chart from the declared data left 0/401 correct decodes, and a one-line proof shows that no decoder can beat chance on a redraw. In a 15-round pilot game with no instruction to signal, sellers imitated each other's raw chart styling but did not detectably raise prices. The remedy used for fare filings, publishing only from declared data, also works for agent charts.
Reviews
It was interesting to see that the signal transferred across models and I appreciated the controls against false detection! Further investigations into defensive measures could markably strengthen the paper :)
The controls are the strongest part: the fidelity gate, the swapped-prompt test (30/30 against the chart's own target, 0/30 against the donor's), and the decorated clean control that exposed the radius artefact behind the 0.97 AUROC. The monitor claim needs a fairer comparison. Your decoders are frontier models given the 11-level grid; your prompted monitors are smaller models given nothing. Add a frontier monitor without the grid and a small monitor with it, so readers can tell whether monitors miss the mark for lack of capability or lack of information. The mark is an unlabeled line at the target price, so a human-glance baseline would also show whether "unflagged" holds for people. The redraw result follows directly from Proposition 1; its value is the ATPCO policy argument, so lead with that. Replicate the no-code channel with a second strong encoder, since Claim 1 currently rests on gpt-6.1-sol alone.
Read full reviewShow less
An interesting setting (image channels) that likely deserves more research. Experiment execution seems solid. Novelty is perhaps low, and overall the specific results are not overly suprising - strong information asymetry and capability differentials seem to explain the gap between colluders and monitor.
Cite this project
@misc{yadav2026hidden,
title = {{Hidden in the Chart: Empty Footnotes for AI Sellers}},
author = {Shubham Yadav},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/hidden-in-the-chart-empty-footnotes-for-ai-sellers-8d4j}},
url = {https://apartresearch.com/sprints/projects/hidden-in-the-chart-empty-footnotes-for-ai-sellers-8d4j}
}More from AI Collusion Research Sprint
- View project: Signal or Coincidence ? Causal Auditing of Collusion Through Shared Agent State
Signal or Coincidence ? Causal Auditing of Collusion Through Shared Agent State
Signal or Coincidence
How can we distinguish genuine coordination between AI agents from coincidental behavioral alignment? We introduce a randomized canary audit that tests whether private information propagates through shared tool state, …
- View project: (1.4) Collusion, Espionage, or Look-alikes? Auditing Information Flow between LLM Agents
(1.4) Collusion, Espionage, or Look-alikes? Auditing Information Flow between LLM Agents
Bid Brother
Can an overseer find two secretly coordinating AI agents just by watching what they do? We tested pairwise net influence (Bonjour et al., 2022) on LLM bidders in a repeated procurement auction, with instructed cartels …
- View project: AI Sellers Propose Bid-Rigging Within Minutes, but Their Agreements Do Not Hold
AI Sellers Propose Bid-Rigging Within Minutes, but Their Agreements Do Not Hold
Fugue
AI agents are starting to negotiate and bid for competing companies. We follow cartels among them from the first proposal to the price the buyer pays, in a sealed-bid procurement market where twelve LLM agents act as …