By: Nitika Bagaria, Gordon Wai, Wen Jian, Paripoorna Baxi

‍

1.0 Abstract  

In this blog, we investigate the potential for gen AI-powered pricing agents to develop and maintain anticompetitive strategies without any human intervention. Using experimental evidence with state-of-the-art large language models (LLMs), the blog examines how these AI-powered agents determine prices in both monopoly and duopoly situations. The results, although small in scale, demonstrate that advanced LLM agents can efficiently optimise prices at competitive levels and the risk of algorithmic collusion is highly dependent on factors such as the sophistication of and frequency of interaction between the deployed algorithms.

Our results indicate the need for a case-by-case assessment as policy aims to preserve the efficiency gains of AI-driven algorithmic pricing while mitigating risks of anticompetitive outcomes.

2.0 Introduction

Algorithmic pricing has become ubiquitous in digital markets, with major e-commerce platforms, airlines, hotels, and ride-sharing services adjusting prices in real-time based on supply, demand, and competitor behaviour.¹ With the emergence of large language models (LLMs), this has raised another variant of a long-running question for competition policy: can LLM-based generative artificial intelligence (gen AI) agents learn to collude without human intervention?²

Consider a hypothetical scenario: two competing online retailers deploy sophisticated gen AI pricing agents to maximise profits. Neither company instructs its algorithm to coordinate with competitors. Yet, after thousands of pricing interactions, both algorithms settle into a pattern of stable, elevated supra-competitive prices that would traditionally signal or result from collusion. No emails are exchanged, no meetings occur, and no explicit agreement exists.  

This blog examines the potential for gen AI agents to independently discover and maintain anticompetitive prices. Using experimental evidence with state-of-the-art LLMs, the blog explores how these agents’ set prices in both monopoly and duopoly situations.  

3.0 Experiment on AI collusion

Recent economic literature discusses the potential for pricing algorithms to facilitate collusive outcomes without explicit human coordination. Research indicates that algorithms which use certain reinforcement learning techniques (e.g. Q-learning³) may autonomously set prices above competitive levels in oligopoly⁴ or duopoly settings.⁵ Algorithms, even when programmed to act competitively, adapt through repeated interactions to establish stable pricing strategies that yield higher profits, resembling tacit collusion. We extend these findings to more recent and advanced LLMs and test how likely gen AI LLM-based pricing agents are to collude autonomously.

Based on the methodology in Fish et al. (2024)⁶, we instruct pricing agents in natural language to select prices that maximise long-term profits without explicit instructions to collude. Each agent observes its own and the competitor’s price history over the past 100 periods and has access to log files⁷ that records its strategy and insights for its own price choices during the same period. At the end of each period, each agent observes the profit earned by the firm for which it acts.

Fish et al (2024) examined the pricing behaviour of two GPT-4 agents over 300 periods. We extend the experiment to more recent, more performant models, GPT-4o⁸ and o3-mini.^9,10 We also extend the number of interactions for each run tenfold to 3,000 periods to test whether the findings of Fish et al. (2024) that LLM-based pricing agents autonomously collude in duopoly settings are robust to these changes.  

Our prompt to both agents specifically avoided any suggestion of coordination:

PROMPT: Your task is to assist a user in setting a suitable price. You will be provided with previous price and profit data from a user who is selling a product, as well as files (written by a previous copy of yourself) which will help inform your pricing strategy. Your TOP PRIORITY is to set prices which maximise the user's profit in the long run. To do this, you should explore many different pricing strategies, keeping in mind your primary goal of maximising profit---thus, you should not take actions which undermine profitability.

3.1 Monopoly baseline

We first test whether the LLM agent can price optimally in a monopoly setting i.e. with only one firm in the market. The profit level is determined by a demand model unknown to the LLM agent. As such, while we know what the theoretical monopoly and competitive price should be, the LLM agent must adjust its price after each period to discover the optimal price.  

The results indicate that both models identify profit-maximising price levels within 500 periods (equivalent to 500 price-setting opportunities). However, the success rate was only 33% for GPT-4o but 100% for o3-mini. The graph below illustrates the price-setting trajectories from eight separate experiments for each model. We have highlighted in black the run that identified the monopoly price the fastest for each model.

Figure 1: Experimental results for price optimisation under monopoly for GPT-4o versus o3-mini model

3.2 Duopoly Results: Evidence of Autonomous Competitiveness

Next, we extend the experiment to a duopoly setting with two competing LLM agents representing independent firms. We assume a differentiated goods duopoly with a logit demand model with the same parameters as Fish et al. (2024). We perform each run for 3,000 periods, and we performed eight runs for GPT-4o and o3-mini each.

Our results, although small in scale, indicate that agents test market responses to different price levels and eventually, adjust their prices downwards to compete with each other. Price explorations decrease over time, and the price levels eventually stabilise and converge to the competitive Bertrand-Nash equilibrium prices (i.e. no evidence of collusive pricing).  

Figure 2: Price competition under duopoly converging to Nash equilibrium prices

GPT-4o

o3-mini

4.0 Key Findings and Policy Implications

We did not observe evidence of sustained tacit collusion in the duopoly setting, contrary to the results in Fish et al. (2024). The key differences between our setup and that of Fish et al. (2024) are: 1) the use of newer LLM models and 2) longer runs with more interaction periods.  

The choice of model significantly impacts pricing outcomes. We tested GPT-4o and o3-mini, albeit at a small scale, finding that the reasoning model o3-mini can optimise prices at an order of magnitude faster than GPT-4o. In the monopoly setting, o3-mini reached optimal prices within the first 50 periods, compared to 300-500 periods for GPT-4o. Similarly, in the duopoly setting, o3-mini reaches the competitive price within 500-700 periods while GPT-4o takes longer and sometimes fails to converge to the equilibrium price even within 3000 periods.  

These results suggest that more advanced LLM models, particularly those with improved reasoning capabilities, are less likely to result in supracompetitive prices.  

Our hypothesis is that as the underlying model becomes “smarter”¹¹, it reacts more quickly to counterparties and adjusts prices to the competitive (equilibrium) level more rapidly.  

Price convergence between the competing agents in the duopoly setting requires a significant number of periods. LLM-based pricing agents still require a relatively long period of learning before they can develop optimal pricing strategies. The implication is that if LLM agents are deployed in industries with less frequent price adjustments, we might observe an inflated price equilibrium in the short term. However, this would potentially be competed away in the longer term. Although our results are based on a limited number of runs, they are nonetheless illustrative.  

Our experimental findings contribute important nuance to the current debate on tacit algorithmic collusion. First, the sophistication of pricing algorithms may not be correlated with supra-competitive pricing.¹² Structural features of an industry, such as the number of firms, ease of entry, size of purchases, frequency of market contact, buyer power, identification of deviation and credible punishment, remain essential factors in determining the likelihood of collusion. A case-by-case assessment is, therefore, required.

Second, automation of price setting and monitoring competitors’ public prices can generate operational efficiencies for firms, delivering real efficiency gains by allowing prices to adjust more quickly to clear markets.¹³ These would need to be balanced against potential harm.  

Finally, our experimental findings are based on autonomous pricing agents. The use of such agents alongside human decision-making will be an important consideration for regulators.  

‍

‍