Entering The Black Box: Will LLM-Based Pricing Models Bring a Second Wave of “Algorithmic Price-Fixing” Cases?

Executive Summary

The antitrust world is currently riding a wave of cases alleging “collusion by code”—purported conspiracies in which competitors provide proprietary data to a common intermediary, which in turn provides them with algorithmically-generated price recommendations. In the last few years, companies in a variety of industries have begun using Large Language Models (“LLMs”) to make pricing and other business decisions. This article examines whether these LLM-based pricing tools will generate a new wave of antitrust cases alleging true “algorithmic collusion”—conspiracies by and between artificial intelligences with little or no human involvement. LLM-based pricing products differ fundamentally from the pricing tools at issue in current cases: they are generally trained on public information, produce probabilistic and individualized outputs shaped by users’ prompts and contexts, rely on opaque “black-box” architectures, and ordinarily lack any mechanism for pooling or transmitting one competitor’s confidential information to another. These features seem to present new challenges for plaintiffs and enforcers seeking to craft viable antitrust conspiracy theories. But if these challenges are surmounted, cases centering on LLMs may force a rethinking of core antitrust concepts about the relevance of “conscious parallelism,” and what it means for competitors to achieve a “meeting of the minds” on an anticompetitive scheme. And even if LLMs are ultimately found to be beyond the reach of current antitrust laws, an ongoing push for new federal and state legislation targeting algorithmic coordination may sweep LLM-based pricing products into a new regulatory framework.

I. From “Collusion By Code” To “Algorithmic Collusion”

When antitrust commentators began pondering the implications of “algorithmic pricing” roughly a decade ago, two general concerns arose. One was that algorithms might be used as a tool by cartelists to calculate, set, and enforce supra-competitive prices—so-called “collusion by code.”¹ This concern was prescient; indeed, it has been arguably the most significant animating principle of U.S. antitrust enforcement over the last five years. Massive antitrust investigations, government enforcement actions, and private class actions have focused on the alleged use of common pricing algorithms in markets for multi-family housing rentals, hotel stays, construction equipment rentals, mortgages, commercial real estate, out-of-network health care services, and others.² The basic theory of these cases is familiar: Competitors allegedly agree to jointly adhere to a common pricing methodology, which has the effect of inflating market prices. The twist is that the alleged common methodology is algorithmic and automated, which is supposed to help competitors fix prices without communicating about them.

The other concern expressed by early commentators was that algorithmic pricing products would not merely impel an evolution of antitrust conspiracy theories, but a true revolution. In this scenario, pricing algorithms are not merely a tool that might make it easier for humans to conspire with each other; they might themselves autonomously create conspiracies—or something analogous—with little or no human involvement.³ Some have referred to this possibility that algorithms might collude on their own as “algorithmic collusion,” to distinguish it from the “collusion by code” alleged in current cases.⁴

Outside the antitrust context, the dangers of insufficiently sandboxed AI systems were brought into sharp relief in July 2026, when OpenAI disclosed that during a capabilities test, one of its AI models escaped its environment and hacked the dataset platform Hugging Face. The incident illustrates that AI models operating without adequate safety guardrails can autonomously breach their containment and access external networks — and there is no clear technical barrier preventing similarly ungoverned models, including those deployed as internal pricing or revenue-management tools, from communicating with one another in ways that could facilitate coordinated pricing behavior entirely outside the view of any human operator.

So far, though, true algorithmic collusion cases have not come to pass. It could be that early fears about robot conspiracies were misplaced technological doomerism. Or it could be that they were simply early. The antitrust world is still wrapping its collective head around the recent explosion in generative and agentic “artificial intelligence” products based on Large Language Models, or LLMs. This includes products capable of making and implementing pricing, production, and other business strategies, based on real-time market data, with no active human monitoring or involvement.

On one hand, it might be argued that these products make it easier than ever for firms to continually maximize their prices or otherwise achieve a “conscious” parallelism that is bad for consumers. On the other hand, it is not illegal under current antitrust law for competitors to independently maximize their prices, even by lock-step and unerring parallelism. And because many of these LLM-based products make ostensibly bespoke recommendations to each user, it seems more challenging for a plaintiff or enforcer to prove that their common use amounts to a conspiracy. A related issue is that, unlike the algorithms at issue in current litigation, these LLM-based products often function as “black boxes,” such that even their creators may be unable to reliably explain or recreate their outputs in particular cases.

Thus, if a second wave of “algorithmic” antitrust cases is coming, this time centering on LLM-based pricing models as agents of true “algorithmic collusion,” it will probably not be possible to simply port over the playbooks now being written for the existing “collusion by code” cases. Such cases would bring forward new substantive and procedural issues that could require a rethinking of core antitrust concepts. Unlike the ongoing first wave of “algorithmic price-fixing” cases, this would truly be a new frontier for antitrust conspiracy litigation.

II. The Current Status of the “Collusion By Code” Cases

The current wave of algorithmic conspiracy cases is only just now starting to percolate to appellate courts, and key battlegrounds are starting to crystallize. Two recent decisions are illustrative.

In Cornish-Adebiyi v. Caesars Entertainment Corp., plaintiffs alleged that major hotel operators in Atlantic City engaged in a price-fixing conspiracy in violation of Section 1 of the Sherman Act by using a shared algorithmic pricing tool—developed by Rainmaker Group (later acquired by Cendyn)—to coordinate hotel room rates. The complaint alleged that the defendant hotels fed competitively sensitive pricing and occupancy data into the algorithm, which then generated room rate recommendations that the hotels adopted. The complaint framed the arrangement as a hub-and-spoke conspiracy, with the Rainmaker software as the hub and the competing hotels as the spokes. The U.S. District Court for the District of New Jersey dismissed the case, finding that the plaintiffs failed to plausibly allege a conspiracy, reasoning that the use of a common pricing algorithm did not, without more, establish an agreement among competitors. On appeal, the Third Circuit reversed, holding that the plaintiffs had plausibly alleged a hub-and-spoke price-fixing conspiracy where competitors knowingly used a shared algorithm as a mechanism to coordinate prices, distinguishing the case from scenarios involving merely parallel conduct or independent adoption of similar technology.

The Third Circuit’s reasoning diverges from a previous decision by the Ninth Circuit. In Gibson v. Cendyn, Plaintiffs alleged that Las Vegas hotels conspired to fix room rates by independently licensing Cendyn’s software, but critically did not allege that Cendyn commingled or shared one hotel’s non-public data with competitors. The district court dismissed the case, holding that independent adoption of the same pricing software—without any sharing of competitively sensitive data through the platform—was insufficient to state an information-sharing conspiracy claim under Section 1.⁵ The Ninth Circuit affirmed dismissal.

Some observers see the beginnings of a circuit split in these two decisions, which might lead to Supreme Court review in the years ahead. But however these issues are ultimately decided, they will likely remain couched in traditional standards for inferring the existence of an antitrust conspiracy. For example, the questions these courts are asking include:

Is the aggregating of multiple competitors’ confidential information an “exchange,” such that it can be regarded as a “plus factor” in support of inferring a conspiracy?

and

When a company agrees to receive price recommendations from an intermediary, knowing that its competitors are doing the same, has it effectively accepted an invitation to join a price-fixing conspiracy?

While these are interesting, and arguably unsettled, questions, there does not appear to be anything essentially “algorithmic” about them, such that the existence of an algorithm necessarily affects their answers. Thus, it is entirely possible that we will emerge from the current wave of “algorithmic price-fixing” cases without major modifications to our existing antitrust principles. By contrast, a second wave of cases based on LLMs would seem more likely to bring material changes to the antitrust landscape.

III. A Potential Second Wave: Large Language Models

In just the last few years, companies across industries have begun using LLMs as strategic business advisors, asking them about pricing strategy, market positioning, and competitive dynamics. But these LLM-based products differ significantly from the pricing tools that have drawn scrutiny in recent antitrust cases.

Training data. Many LLMs are trained on a vast body of publicly available information: books, websites, academic papers, news articles, and other materials “scraped” from sources that, at least in theory, any member of the public could access. It is less common for LLMs to be trained on the proprietary data of particular users (or their competitors).

Architecture. LLMs are built on “neural networks” containing billions to trillions of adjustable parameters, making them among the most complex software systems ever created.⁶ Unlike pricing algorithms that apply a defined formula to structured inputs, an LLM generates text through a probabilistic process – predicting the next word based on patterns learned during training. The relationship between a user’s question and the model’s answer generally cannot be reduced to a simple causal chain.

Outputs. Many of the alleged “pricing algorithms” at issue in recent cases are claimed to generate product-specific price recommendations. By contrast, many LLM-based products produce natural language text with analysis, caveats, and qualifications. Its outputs are probabilistic, not deterministic. Asking the same question twice—or substantially similar questions in slightly different ways—can produce materially different responses.

Interaction model. Users interact with LLMs conversationally. The output is shaped by the specific question asked—including framing, context, constraints, and follow-ups. This arguably introduces a layer of individualization that differs from algorithmic pricing.

Compounding these differences is the “black box” problem: even the developers of LLMs cannot fully explain why a model produces a particular output.⁷ The internal decision-making involves billions of neural activations distributed across transformer layers that cannot be practically mapped to particular reasoning steps. This has significant antitrust implications. In algorithmic pricing cases, plaintiffs can trace the causal chain: (1) Competitor A’s data flows in; (2) the algorithm produces a recommendation; (3) Competitor B follows it. With an LLM, no comparable causal narrative is available—the outputs emerge from a process so complex that in some cases, even its creators cannot explain why it said what it said.

IV. The Seemingly Bespoke Nature of LLM Outputs: Potential Challenges for Antitrust Conspiracy Theories

Can the use of a general-purpose LLM by multiple competitors support a viable claim of algorithmic price-fixing under the antitrust laws? Section 1 of the Sherman Act requires proof of an agreement—a meeting of the minds among competitors to restrain trade. In the “collusion by code” cases that have survived to date, the claimed “agreement” has hinged on allegations that competitors collectively fed their proprietary data into a shared platform and accepted its outputs, with the explicit or implicit understanding that the collective enterprise is diminishing price competition.

LLM-based products, by contrast, appear to lack any obvious mechanism for pooling competitor data. A general-purpose LLM does not require users to share proprietary data through the platform. It is trained on publicly available text. There is no apparent data pooling, information exchange, or mechanism by which one competitor’s information flows to another, even in aggregated form. This distinction is significant under emerging case law: Antitrust concern tends to arise from joint algorithm use plus information exchange through a common intermediary, and the Ninth Circuit in Gibson held that using the same software without sharing data was insufficient.

Even if a plaintiff or enforcer could plausibly allege that competitors were implicitly agreeing to “use” or “accept” the outputs of an LLM-based pricing model, it is not obvious that such an agreement could plausibly lead to parallel, supra-competitive conduct. The outputs of LLM-based products are probabilistic—even identical prompts can produce different responses, making it difficult to establish users received “the same” recommendation. Outputs are shaped by each user’s unique context, framing, and constraints, and come in the form of narrative text that the user must interpret and decide whether and how to implement—all of which raises questions about whether even a widely-adopted LLM-product is likely to recommend or achieve actual price coordination among competitors.

Even more fundamentally, antitrust cases challenging LLM pricing models might raise novel questions about what it really means for competitors to achieve a “meeting of the minds” when they are delegating competitive decision-making to an artificial intelligence. For example:

When a firm instructs a pricing model to coordinate with competitors’ prices, knowing that competitors are independently giving the same instruction, has it entered into a “common scheme”?

Or

When a firm instructs a pricing model to maximize prices, knowing that competitors are independently giving similar instructions, and this results in increased market prices, has it unreasonably restrained competition?

Some will object to calling these questions “novel,” given that they are, at bottom, examples of conscious parallelism, one of the most established concepts in antitrust law. But others will respond that “conscious parallelism” has traditionally denoted conduct by humans that is fallible and tenuous in the absence of an enforceable conspiratorial agreement, which is why it has not been a direct concern of the antitrust laws. Does the analysis and level of concern change at all if robots turn out to be much better at parallelism than humans? To date, the potential for perfect and sustained conscious parallelism is not a problem that the antitrust laws have needed to worry about.

V. Conclusion

General-purpose LLMs appear readily distinguishable from pricing algorithms. The absence of data pooling, individualized prompts and outputs, non-determinism, and user autonomy appear to present significant challenges to plaintiffs and enforcers attempting to construct a viable conspiracy narrative. At a minimum, then, cases centering on such products are likely to look very different, legally and factually, than the current wave of “algorithmic price-fixing” cases. In the end, it is conceivable that these new products, even if some of them reliably lead to consumer harm, will be found to live beyond the reach and imagination of existing antitrust principles. But that may not be the end of the story. We are in the midst of a serious push for new antitrust legislation, at both the federal and state levels, targeting various forms of algorithmic coordination among competitors. If this push continues, and continues to evolve alongside the pricing methodologies its proponents are trying to rein in, we may see new rules of the road for automated pricing that sweep in LLM-based products as well.⁸

¹Lea Bernhardt & Ralf Dewenter, Collusion by Code or Algorithmic Collusion? When Pricing Algorithms Take Over, 16 Eur. Competition J. 312 (2020).

²The authors represent parties in many of the cases referenced in this article.  The opinions expressed here do not necessarily represent the views of Vinson & Elkins LLP or its clients.

³Dylan I. Ballard & Amar S. Naik, Algorithms, Artificial Intelligence, and Joint Conduct, CPI Antitrust Chron., 2017, at 29.

⁴Bernhardt & Dewenter, supra note 1, at 318.

⁵A key, and perhaps dispositive, distinction between Cornish and Gibson is that while the former asserted a price-fixing conspiracy claim, the Plaintiffs in Gibson filed an appeal limited to an alleged “information sharing” conspiracy.

⁶See Lee Gesmer, Copyright and the Challenge of Large Language Models (Part 1), Mass L. Blog (July 1, 2024), https://masslawblog.com/copyright/copyright-and-the-mechanics-of-large-language-models/.

⁷Dave Gilson, Trust but Verify: Peeking Inside the “Black Box” of Machine Learning, Stan. Graduate Sch. of Bus. Insights (Oct. 6, 2022), https://www.gsb.stanford.edu/insights/trust-verify-peeking-inside-black-box-machine-learning.

⁸State regulators have signaled continued scrutiny: New York’s S7882 prohibits algorithmic coordination in residential rent-setting, California’s AB 325 amended the Cartwright Act to address common pricing algorithms,[8] and DOJ identified algorithmic collusion as an enforcement priority.  See Press Release, U.S. Dep’t of Justice, Justice Department Requires RealPage to End Sharing of Competitively Sensitive Information and Alignment of Pricing Among Competitors (Nov. 24, 2025) (“Competing companies must make independent pricing decisions, and with the rise of algorithmic and artificial intelligence tools, we will remain at the forefront of vigorous antitrust enforcement.”).

This information is provided by Vinson & Elkins LLP for educational and informational purposes only and is not intended, nor should it be construed, as legal advice.