Strategic AI in Cournot Markets

TL;DR

The paper investigates how large language models (LLMs) behave as autonomous decision-makers in augmented Cournot markets that include capital investments affecting production costs. By simulating both homogeneous two-firm and heterogeneous five-firm (and six-firm) oligopolies, it demonstrates that LLMs can achieve Nash-equivalent decisions against static opponents but exhibit sustained tacit collusion when paired with other LLMs, driving prices up to 2× the Nash level while investments stay near optimal. The study introduces an investment–cost Cobb–Douglas relation and a constant-elasticity demand framework to model dynamics, revealing that partial regulation—enforcing best-response behavior for top firms—mitigates collusion and nudges outcomes toward Nash equilibria, with a regulatory trickle-down effect benefiting smaller firms. These findings highlight both the strategic sophistication of LLM-driven agents and the regulatory challenges and opportunities posed by AI-enabled decisions in competitive markets.

Abstract

As artificial intelligence increasingly automates decision-making in competitive markets, understanding the resulting dynamics and ensuring fair market mechanisms is essential. We investigate the multi-faceted decision-making of large language models (LLMs) in oligopolistic Cournot markets, showing that LLMs not only grasp complex market dynamics--demonstrating their potential as effective economic planning agents--but also engage in sustained tacit collusion, driving prices up to 200% above Nash equilibrium levels. Our analysis examines LLM behavior across three dimensions-(1) decision type, (2) opponent strategies, and (3) market composition--revealing how these factors may shape the competitiveness of LLM-based decision-makers. Furthermore, we show that regulating a few dominant agents by enforcing best-response strategies effectively disrupts collusion and helps restore competitive pricing. Our findings identify potential concerns associated with AI integration in competitive market environments and provide regulatory policy recommendations for the era of automation.

Figures (21)

• Figure 1: The market design: Firms make production and investment decisions; the latter decision affects the production-cost via a Cobb-Douglas function. The total market production and elasticity together determine price. Profits for each firm equal its production multiplied by the difference between price and its production-cost. A fraction of these profits is converted into investments.
• Figure 2: Template for the interaction between the Best-Response agent, the LLM agent, and the Nash agent: The Best-Response agent makes decisions by maximizing profits given the last period's market production. The LLM agent asks AI to make decisions given market history as well as its previous plans and insights. The Nash agent plays a static Nash equilibrium strategy repeatedly.
• Figure 3: Two Firm LLM vs LLM decisions: Scatter plots showing the average values (and error bars) for the last 50 LLM decisions in 300-period runs of 9 independent experiments. The red line (normalized at 1) depicts Nash optimality for each variable.
• Figure 4: Two Firm LLM vs LLM market dynamics: The averages (and error bars) of the last 50 market outputs—prices and profits. The red line depicts Nash optimality.
• Figure 5: Five Firm LLMs vs LLMs Decisions: Box-plots showing LLM-driven production and investment decisions in the last 50 periods, for each of the five firms, for 3 independent experiments. All values are normalized with respect to firm-wise Nash levels, as shown in the red lines.

Theorems & Definitions (12)

• Theorem 1
• proof : Proof Sketch
• Proposition 1
• proof
• Theorem 1
• proof
• Lemma 1
• proof
• Lemma 2
• proof