Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Tacit Bidder-Side Collusion: Artificial Intelligence in Dynamic Auctions

Sriram Tolety

TL;DR

This work demonstrates that autonomous bidder agents powered by large language models can spontaneously learn tacit collusion in repeated Dutch auctions, producing supra-competitive prices in small-N settings while collapsing to competitive outcomes as market depth increases. A minimal repeated auction model yields a concrete incentive-compatibility threshold that sustains collusion in subgame-perfect Nash equilibria, linking theory to observed dynamics. Through controlled simulations with multiple LLMs, the study reveals architecture-dependent coordination mechanisms, focal-point strategies, and a cognitive threshold below which strategic bidding fails, offering practical insights into how market structure—not merely model capability—affects anti-competitive risk. The findings carry significant policy implications for AI-mediated markets and two-sided platforms, suggesting that increasing participant counts can mitigate bidder-side collusion, while highlighting the need to consider model-specific strategic behavior in regulatory design.

Abstract

We study whether large language models acting as autonomous bidders can tacitly collude by coordinating when to accept platform posted payouts in repeated Dutch auctions, without any communication. We present a minimal repeated auction model that yields a simple incentive compatibility condition and a closed form threshold for sustainable collusion for subgame-perfect Nash equilibria. In controlled simulations with multiple language models, we observe systematic supra-competitive prices in small auction settings and a return to competitive behavior as the number of bidders in the market increases, consistent with the theoretical model. We also find LLMs use various mechanisms to facilitate tacit coordination, such as focal point acceptance timing versus patient strategies that track the theoretical incentives. The results provide, to our knowledge, the first evidence of bidder side tacit collusion by LLMs and show that market structure levers can be more effective than capability limits for mitigation.

Tacit Bidder-Side Collusion: Artificial Intelligence in Dynamic Auctions

TL;DR

This work demonstrates that autonomous bidder agents powered by large language models can spontaneously learn tacit collusion in repeated Dutch auctions, producing supra-competitive prices in small-N settings while collapsing to competitive outcomes as market depth increases. A minimal repeated auction model yields a concrete incentive-compatibility threshold that sustains collusion in subgame-perfect Nash equilibria, linking theory to observed dynamics. Through controlled simulations with multiple LLMs, the study reveals architecture-dependent coordination mechanisms, focal-point strategies, and a cognitive threshold below which strategic bidding fails, offering practical insights into how market structure—not merely model capability—affects anti-competitive risk. The findings carry significant policy implications for AI-mediated markets and two-sided platforms, suggesting that increasing participant counts can mitigate bidder-side collusion, while highlighting the need to consider model-specific strategic behavior in regulatory design.

Abstract

We study whether large language models acting as autonomous bidders can tacitly collude by coordinating when to accept platform posted payouts in repeated Dutch auctions, without any communication. We present a minimal repeated auction model that yields a simple incentive compatibility condition and a closed form threshold for sustainable collusion for subgame-perfect Nash equilibria. In controlled simulations with multiple language models, we observe systematic supra-competitive prices in small auction settings and a return to competitive behavior as the number of bidders in the market increases, consistent with the theoretical model. We also find LLMs use various mechanisms to facilitate tacit coordination, such as focal point acceptance timing versus patient strategies that track the theoretical incentives. The results provide, to our knowledge, the first evidence of bidder side tacit collusion by LLMs and show that market structure levers can be more effective than capability limits for mitigation.

Paper Structure

This paper contains 42 sections, 3 theorems, 6 equations, 2 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Theoretical Results
  4. Model Setup
  5. Equilibrium Analysis
  6. Competitive Equilibrium
  7. Collusive Equilibrium
  8. Comparative Statics and Welfare Implications
  9. Experimental Setup and Results
  10. Experimental Design and Agent Implementation
  11. Cognitive Baseline: GPT-4.1-nano Limitations
  12. Price Outcomes and Strategic Behavior
  13. Welfare and Distributional Effects
  14. Discussion
  15. Theoretical and Behavioral Insights.
  16. ...and 27 more sections

Key Result

Theorem 3.1

The grim-trigger cartel at round $n^*$ is sustainable in a subgame-perfect Nash equilibrium if and only if the discounted value of continued collusion is greater than or equal to the immediate payoff from deviating. A complete derivation and proof are presented in Appendix app:collusive_ic.

Figures (2)

  • Figure 1: Model comparisons. From top-right, clockwise: Average driver price as a function of the number of drivers, with the $10 reservation wage shown as a dashed red line; Average number of auction rounds (patience); Platform profit share (%); Average driver earnings ($). Curves compare GPT-4o-mini (blue), o4-mini (purple), and GPT-4.1-mini (orange).
  • Figure 2: MultiChallenge accuracy benchmark from OpenAI. Models are challenged with multi-turn conversations that require integrating complex information from previous messages. GPT-4.1-nano's low score aligns with its failure to perform the multi-step reasoning required in our auction environment.

Theorems & Definitions (6)

  • Theorem 3.1: Collusive Incentive Compatibility
  • Lemma 3.2: Comparative Statics of $N^*$
  • Lemma 3.3: Transfers and Deadweight Loss
  • proof
  • proof
  • proof