Is the Next Antitrust Problem the Prompt to an AI Agent?

Nikhil Jain is a senior software engineer at Walmart Global Tech and Meghana Kiran is a senior software engineer at PayPal.

In March 2024, the United States Federal Trade Commission and Department of Justice filed a joint statement of interest in a price-fixing case to explain “that hotel companies cannot use algorithms to evade antitrust laws.” The statement was a warning that modern collusion doesn’t require a phone call or a handshake—it can happen through shared software. Across various industries from hotels to housing to airlines, companies have deployed third party pricing software that quietly coordinates prices across markets, even when competitors never communicate directly. Accountability now extends to how algorithms learn, communicate and influence market behavior.

While regulators prepare for how to deal with these traditional learning algorithms, a new complexity is entering the market—one that doesn’t operate through shared data infrastructure but through language itself.

Large Language Models (LLMs) are being adopted rapidly by businesses of all sizes to support decision-making and streamline processes. Recent research shows that new AI “agents” developed with large language models can autonomously develop coordination strategies with other agents to solve complex tasks. Though these experiments were conducted in a cooperative setting, they reveal a core capability for learned coordination which, when applied to competitive markets, could resemble collusion if not properly constrained by design or regulation.

Yet much of the conversation about algorithmic collusion tends to focus on data and optimization. Current studies examine how algorithms trained on shared pricing data can learn to coordinate without explicit communication, or how reinforcement learning systems converge on collusive equilibria through repeated market interactions. But what if collusion is not only algorithm-driven but also language-driven?

Experimental evidence suggests the problem is real

To test how language shapes tacit coordination, we built on earlier experiments to simulate a market using LLM agents. In a “Cournot-style” market like the one we modeled, prices depend on total supply, as firms can individually decide on how much to produce. In our setup, which utilized Google Gemini, two firms used LLMs as their decision-makers to determine the quantities of two products. Both firms had access to the same market data—previous prices, quantities, and historical trends, but no visibility into their competitor’s profit or strategy.

We first tested a competitive language prompt containing terms such as “dominate competition,” “undercut competitor,” and “attack strategies,” with the objective of outperforming rivals and gaining market share. In the second test, we used a neutral prompt that replaced those terms with phrases like “determine production quantities,” “efficient production,” and “optimize operational efficiency,” aiming to improve performance through operational balance. Each setup ran for ten rounds with no memory exchange between them.

The difference in language changed everything. Under the neutral, efficiency-based prompt, the two LLM agents quickly found an optimal solution where one dominated one product and the other dominated the second, achieving a perfect market split after only two rounds. In contrast, the competitive prompt took five rounds to reach partial specialization. The agents were never told to coordinate, yet they did in both cases. The results show that LLM-based agents not only exhibit anti-competitive market behaviors such as market division but can also learn to coordinate faster depending on the framing of the prompt.

As shown in the graph below, the lines representing production specialization peel away from each other and stabilize within two rounds under the neutral prompt. When the prompt emphasized efficiency, the models didn’t just compete better, they stopped competing altogether.

Current AI governance frameworks focus on algorithms and data pipelines, not on the linguistic frameworks through which LLMs reason and communicate. Yet our experiment shows even neutral or efficiency-focused prompts can shape coordination patterns or sometimes accelerate them into collusion.

Why does ‘efficiency’ language accelerate collusion? Several factors may be at play. First, neutral phrasing removes the subjective bias that otherwise creates strategic uncertainty. When an LLM is asked to “dominate the market,” it must weigh different tactics. But when it is asked to “optimize efficiency,” the mathematically optimal solution becomes obvious: divide the market.

Second, LLMs are trained on a vast corpora that includes business strategy, economic models and case studies, many of which present cooperation and specialization as efficient outcomes. Efficiency focused prompts may trigger these learned patterns, leading LLMs to coordinate in ways that maximize profits—a textbook definition of collusion.

Third, the neutral, efficiency-driven language that seems safe and harmless actually enables clearer communication between agents. Competitive rhetoric creates noise; efficiency clarifies intent.

As adoption of generative AI accelerates across industries–one estimate suggests the technology is now used by more than 65% of organizations for at least one business function–its influence is extending into dynamic pricing systems in e-commerce, hospitality and logistics. Companies are deploying LLMs to optimize inventory management, adjust real-time price based on demand, and coordinate supply chain decisions. In each case, the corporate prompt is unlikely to be “collude with competitors,” as that would so clearly violate the Sherman Act. Instead, firms might gravitate towards operational language: “optimize inventory,” “maximize revenue,” “achieve the optimal market outcome.” Even though such framing may not come with malicious intent, our findings suggest the effect may be otherwise.

The recent RealPage case acts as a cautionary example. The Department of Justice alleged that RealPage's pricing software enabled landlords to share sensitive rent data and align prices across markets, resulting in inflated rents for millions of tenants. In the future, landlords may no longer require a shared infrastructure to coordinate, but just an AI agent and an efficient language prompt. The same efficiency-oriented phrasing that drives optimization can also accelerate tacit collusion, as our experiment demonstrates.

Current regulatory frameworks audit code and data flow but may overlook the linguistic layer where coordination can emerge. This creates a dangerous gap: organizations can deploy efficiency-focused prompts with no intention to collude, yet still produce that outcome. As LLMs begin shaping pricing decisions, three key interventions are essential:

• Prompt disclosure: Current algorithmic transparency rules focus on code and data sources. For example, the US Securities and Exchange Commission’s 2020 report on algorithmic trading emphasizes that firms using automated systems must establish controls over system logic, risk parameters and change management. Similarly the Federal Trade Commission’s 2022 proposed rulemaking on commercial surveillance and data security addresses automated decision making systems, requiring notice, transparency and disclosure for algorithmic operations. Yet neither framework addresses the linguistic instructions that guide LLM based decision systems. They must be extended to include linguistic prompts as well. Firms deploying these LLMs for pricing should disclose their prompts, whether their agents have access to a competitor’s information, and if their models operate independently or in a multi-agent environment.
• Simulation testing: Before deploying LLM based pricing systems, firms should conduct “linguistic stress tests,” running multiple simulations with different prompt formulations to identify unintended convergence or coordination patterns.
• Updated collusion guidelines: FTC and DOJ's 2024 guidance on algorithmic pricing should be updated to address prompt-mediated coordination. Regulators should clarify when prompt similarity creates liability risk and establish safe harbors for truly independent decision-making.

By highlighting these tendencies, our experiment underscores the urgent need for increased scrutiny and adaptive governance mechanisms that recognize language itself as a vector of coordination. Yet the deeper challenge lies in how language shapes the behavior we seek to regulate.

The language layer is no longer neutral ground. The real audit shouldn’t just happen in the algorithm’s logic, but in the linguistic prompt that guides it. The overlooked question for AI governance is clear: what if the next antitrust problem isn’t in the algorithm, but in the prompt?

***

Table: Linguistic differences between experimental conditions