AI AND ANTITRUST: "THE ALGORITHM MADE ME DO IT"

By Robin Feldman¹, Caroline A.Yuen

As the dawn of artificial intelligence ("AI") rises rapidly, competition authorities should contemplate the potential for hazy days ahead. Undoubtedly, AI’s already ubiquitous presence² offers exciting possibilities, from enhancing efficiency,³ to leveling the playing field for non-native speakers,⁴ to enabling scientific discovery.⁵ Despite these breathtaking advancements, however, recent data from the Pew Research Center reveal

[Page 1]

that only 15% of adults surveyed were "more excited than concerned about the increasing use of AI in daily life," with 46% expressing "an equal mix of concern and excitement."⁶

Policymakers also manifest concerns about AI, exemplified by the extent to which government actors are racing to pass laws, benchmarks, and guidelines to regulate the development and use of AI technology.⁷ And in the intellectual property realm, courts have seen a wave of copyright cases brought by individual and business content creators over the use of copyrighted work in the training of generative AI tools.⁸

[Page 2]

Generative AI and AI more broadly are not only making waves in the courtroom;⁹ these waves have reached legal academia as well.¹⁰ Although there is an abundance of literature focused on AI and its effects on intellectual property ("IP") law,¹¹ IP’s antagonistic cousin, competition law, is gently coming to the fore.¹² Emerging competition

[Page 3]

law literature has begun to explore effects of using algorithms on competition.¹³ Moreover, recent cases increasingly assert antitrust claims against algorithmic collusion.¹⁴

This paper seeks to add to this emerging antitrust and AI scholarship. As AI becomes a more accurate and skillful tool, it could conceivably lead to more anticompetitive hub-and-spoke arrangements that current competition laws may not be fully equipped to evaluate. Building on Feldman’s previous work concerning the pharmaceutical supply chain and competition- related issues raised by algorithms,¹⁵ this paper examines the pharmaceutical supply chain as an example of an industry with concentrated intermediaries. We argue that the widespread adoption of increasingly powerful algorithms will exacerbate the susceptibility of such industries to anticompetitive effects. Specifically, as intermediaries become equipped with more accurate and skillful AI-driven algorithms, the intermediaries will become better facilitators of tacit collusion. Moreover, a concentrated intermediary level enhances that power.

Finally, we outline some thoughts on mechanisms that have the potential to curtail such collusion. Inspired by anti-money laundering compliance schemes adopted around the world, we propose designing a model for anti-collusion compliance, and we outline the crucial characteristics that such a model should possess.

[Page 4]

I. UNDERSTANDING THE RELEVANT DOMAINS: AI, ANTITRUST, AND THE PHARMACEUTICAL SUPPLY CHAIN.

Each of the domains—AI, antitrust, and the pharmaceutical supply chain—constitutes a complex arena. Understanding the intersection of the three requires a brief tour of the relevant aspects of each, with a deeper dive to follow.

A. ARTIFICIAL INTELLIGENCE

It is difficult to remember when hearing the term "artificial intelligence" would not bring to mind ChatGPT. Before November 2022, however, the notion of artificial intelligence conjured up images of WALL-E,¹⁶ Marvin the Paranoid Android,¹⁷ and Ava.¹⁸ Today, the term "artificial intelligence" or "AI" feels familiar; but ask almost anyone what AI is, and the term becomes elusive.

In its broadest sense, artificial intelligence describes technology that can mimic and perhaps surpass human intelligence.¹⁹ Although the idea of AI in its broadest sense could be said to have been present as early as in ancient Greece,²⁰ efforts to create artificial intelligence did not begin in earnest until the 1950s, marked by the publication of Alan Turing’s seminal paper, Computing Machinery and Intelligence.²¹ In his paper, Turing posed the simple question "Can machines think?" While recognizing the challenge of defining "machine" and "think," he answered his own question in the form of a thought experiment: the imitation game,²² now widely known as the Turing Test. The test is simple and consists of three participants: a human interrogator, a human participant, and a machine participant. If after posing several questions to the two participants, the human

[Page 5]

interrogator cannot distinguish between the responses of the human and those of the machine, then that machine can think.²³

This simple conception of "thinking" may partially explain why models such as ChatGPT gained widespread popularity overnight.²⁴ Unlike old-fashioned computer chatbots, which used rule-based programming to recognize certain keywords in user prompts and provide pre-loaded responses,²⁵ ChatGPT can respond to any prompt, and respond in a human-like way.

Consider ELIZA, one of the earliest examples of a computer chatbot, developed at the MIT Artificial Intelligence Lab by Joseph Weizenbaum²⁶ to act as a therapist. Weizenbaum described ELIZA’s inner workings this way:

The gross procedure of the program is quite simple; the text is read and inspected for the presence of a keyword. If such a word is found, the sentence is transformed according to a rule associated with the keyword, if not a content-free remark or, under certain conditions, an earlier transformation is retrieved. The text so computed or retrieved is then printed out.²⁷

Though the actual details of that process are more complicated, ELIZA’s performance was ultimately limited by its preprogrammed responses and the user’s input.²⁸ If the user withholds information, there is only so much that ELIZA can do, and very soon the conversation feels inauthentic:

Eliza: Hello, I am Eliza. I’ll be your therapist today.
User: How should I deal with sadness
Eliza: Are such questions on your mind often?
User: Yes

[Page 6]

Eliza: I see. Let’s explore that a bit more.
User: I feel worried about the earth’s future
Eliza: Do you often feel worried about the earth’s future?²⁹

When ELIZA does not have a prepared response, it simply reverts to repeating what the user said in the form of a question.³⁰ In short, ELIZA was programmed by a human who gave ELIZA a specific set of instructions to follow—an algorithm. Like a parrot repeating phrases it hears without understanding the content of what is parroted, when a user provided an input, ELIZA went through its algorithm to search for keywords, and if found, then inserted them in preprogrammed sentence structures. This is now commonly known as traditional or classical computer programming.³¹

In contrast to the likes of ELIZA, ChatGPT can respond in a human-like way to any prompt. Using the same initial prompt "How should I deal with sadness," ChatGPT responded with a 7-point answer and ended with the sympathetic encouragement to "[r]emember, everyone experiences sadness at times, and it’s a normal part of life. You’re not alone, and there are people and resources available to support you through this."³² Unlike ELIZA, who recycles through its preprogrammed responses, ChatGPT’s response to the

[Page 7]

same prompt is different every time;³³ it can even revise its initial response. For example, when asked to "please summarize [your response] in three sentences[,]" ChatGPT’s new response became: "To deal with sadness effectively, acknowledge and accept your feelings without judgment. Talk to trusted individuals or a therapist for support and comfort. Engage in self-care, practice mindfulness, and seek professional help, if necessary, while giving yourself patience and time to heal."³⁴

ChatGPT’s astounding performance, in comparison to ELIZA, can be explained by the difference in how the model functions. ChatGPT was trained through reinforcement and deep learning,³⁵ each a subfield of machine learning. Although delving into the technological details of machine learning is beyond the scope of this paper, one key detail is important to understand: unlike traditional programming, which relies on rules and is deterministic,³⁶ machine learning relies on data and through adjusting its internal parameters, adapts and learns.³⁷ As Arthur Samuel explained, machine learning is the field of study that gives computers the ability to learn without explicitly being programmed.³⁸ In very simple terms, machine learning takes a vast amount of data, feeds it through a machine-learning model, which tries to identify patterns in data. The accuracy of the model is evaluated and the model’s internal parameters—knobs and dials—are adjusted to improve the accuracy of its predictions. This process is repeated until a certain level of accuracy is reached. Remarkably, given that a machine-learning system is not simply following a set of rules, the system can recognize patterns and relationships in datapoints that humans might miss.³⁹ The trained system can then output descriptions of that new data, make predictions based on that data, or prescribe what to do given that data.⁴⁰

This technology extends far beyond chatbots. The use of machines to recognize patterns and generate outputs has been deployed in advanced pricing algorithms in many industries. As will be explained in more detail, pricing algorithms—a computer program

[Page 8]

specifically designed for the purpose of outputting the optimal price—similarly evolved from traditional programming to machine learning systems.

Thus, we consider AI not in the broad context of machines capable of human-like intelligence, but in the narrower context of self-learning algorithms that process large swathes of data to recognize patterns and solve complex problems, such as pricing.⁴¹

B. OVERVIEW OF COLLUSION—FROM SMOKE-FILLED ROOMS TO TACIT COLLUSION TO ALGORITHMIC COLLUSION

Political cartoonists of the early 1900s captured the notion of conspiracy among competitors with drawings of smoky rooms full of well-fed men in suspenders chomping on cigars and plotting to keep prices high. Artificial intelligence, however, could lead to a completely different kind of collusion. Yet, despite the differences, the algorithmic version merely lies a little further down the evolutionary pathway. The following section will trace algorithmic collusion and its implications back to its roots in early U.S. antitrust law.

Among many other provisions, the Sherman Act and the Clayton Act together prohibit agreements among competitors to unreasonably restrain trade.⁴² Section 1 of the Sherman Act of 1890 declares as illegal "[e]very contract, combination in the form of trust or otherwise, or conspiracy, in restraint of trade. . . ."⁴³ To prove an illegal conspiracy in restraint of trade, one must show (1) an agreement among separate competitors and (2) that the agreement constitutes an unreasonable restraint on trade.⁴⁴ Thus, section 1 concerns concerted action between two or more actors. In contrast, section 2 of the Sherman Act concerns monopolistic (that is, unilateral) action.

In addition to the Clayton and Sherman Acts, Section 5 of the Federal Trade Commission Act ("FTC Act") also makes unlawful any "[u]nfair methods of competition . . . and unfair or deceptive acts or practices in or affecting commerce. . . ."⁴⁵ This broadly drafted provision has been the subject of fierce debate, particularly with respect to its ability to be invoked as a stand-alone provision separate from the Sherman and Clayton Acts.⁴⁶ Indeed, the FTC recently filed an administrative complaint against three of the largest pharmacy benefit managers alleging that they have violated Section 5

[Page 9]

of the FTC Act.⁴⁷ Given the intended scope of this paper, however, the authors focus on Section 1 of the Sherman Act, setting the FTC Act debate to the side.

Whether an agreement constitutes an unreasonable restraint on trade under Sherman Act Section 1 depends on the type of restraint and the nature of the relationship involved. Within these, a small category of restraints is thought to be so egregious in their anticompetitive effects that they are deemed, per se, unreasonable, regardless of any procompetitive effects or the intentions of the parties involved.⁴⁸ In particular, an agreement between competitors—that is, a horizontal agreement to fix prices or divide up and allocate the market—is per se illegal.⁴⁹ Once the agreement is proven, the court generally makes no further inquiry into the parties’ motives or the actual impact of the restraints, even if the agreement might have procompetitive effects.⁵⁰ Thus, in per se cases, the most important determination for the court to make is simply the existence of an agreement.

In contrast, most other alleged violations of Section 1 are analyzed through the rule of reason standard by balancing the procompetitive and anticompetitive effects of the restraint to see whether the net effect of the challenged activity is anticompetitive.⁵¹ In such cases, the court might look to the parties’ intention and motives as relevant factors in determining whether there has been a Section 1 violation. Under Supreme Court precedent, vertical agreements are analyzed under the rule of reason standard.⁵²

1. THE RELATIONSHIP OF THE PARTIES

The stringency in the court’s treatment of agreements among competitors makes intuitive sense. In a well-functioning, competitive market, we expect direct competitors to compete tooth-and-nail to attract customers. We would not expect them to compare

[Page 10]

notes, let alone enter into agreements with one another. As a result, communications between competitors generally rouse suspicion.

To illustrate this intuitive outlook on markets, consider the Lysine cartel, which operated between 1991 and 1995 on a global level to fix the price of the common food additive lysine.⁵³ Archer-Daniels-Midland Company ("ADM") and other lysine producers met and agreed over several meetings how to raise the price of lysine via coordinated price hikes, as well as how to allocate sales volumes among themselves.⁵⁴ The FBI was able to obtain direct evidence of the conspiracy through an ADM insider,⁵⁵ who recorded hundreds of hours of conversations and videos of the secret meetings, proving, beyond a shadow of a doubt, that ADM and the other lysine producers were working in concert to fix the price of lysine. Unsurprisingly, the executives representing ADM at the meetings were all convicted of violating Section 1 of the Sherman Act and sentenced to lengthy terms of imprisonment.⁵⁶ ADM also agreed to a fine of $100 million, which was the largest criminal recovery in the history of the Sherman Act, at the time.⁵⁷

Of course, cases backed up by direct evidence of actual agreements to restrain trade are rare.⁵⁸ As the Supreme Court noted in Falstaff Brewing, "circumstantial evidence is the lifeblood of antitrust law."⁵⁹

2. TACIT COLLUSION

Under current antitrust law, not every instance of collusion will lead to a finding of actionable behavior. Rather, the behavior may fall within the more subtle activity known as "tacit collusion." Tacit collusion, or conscious parallelism, refers to coordinated activities between firms without there being an explicit agreement to coordinate.⁶⁰

Take for example In re Text Messaging Antitrust Litigation.⁶¹ There, the defendants were a trade association and four wireless network providers that offered text messaging

[Page 11]

services to consumers.⁶² The plaintiffs alleged that the defendants agreed to fix the price of texting. Defendants sought to prove price-fixing through a series of parallel behaviors, as well as aspects of industry structure and industry practices, which they claimed facilitated collusion on text messaging prices.⁶³ In its opinion, the 7th Circuit took the opportunity to stress the difference between express and tacit collusion, noting that: "Express collusion violates antitrust law; tacit collusion does not."⁶⁴ The court further elaborated:

It is true that if a small number of competitors dominates a market, they will find it safer and easier to fix prices than if there are many competitors of more or less equal size. For the fewer the conspirators, the lower the cost of negotiation and the likelihood of defection; and provided that the fringe of competitive firms is unable to expand output sufficiently to drive the price back down to the competitive level, the leading firms can fix prices without worrying about competition from the fringe. But the other side of this coin is that the fewer the firms, the easier it is for them to engage in "follow the leader" pricing ("conscious parallelism," as lawyers call it, "tacit collusion" as economists prefer to call it)—which means coordinating their pricing without an actual agreement to do so.⁶⁵

In other words, where a few firms dominate one industry, it is very easy for the resulting behavior of the firms to look collusive even if the competing firms did not actually agree to coordinate their behaviors. Indeed, in oligopolistic markets, tacit collusion is highly likely.⁶⁶

This is the crux of the challenge in such cases: Is there an actual agreement, or is this a case of mere tacit collusion? Where the parallel conduct of competitors can be explained by legitimate business rationale, something beyond those parallel actions will be needed to prove an agreement exists.

This point was further clarified in the 2007 Supreme Court case, Twombly. The Court held that in a tacit collusion case "plus factors" are needed to trigger antitrust scrutiny.⁶⁷ The Supreme Court explained:

[Page 12]

[A]n allegation of parallel conduct and a bare assertion of conspiracy will not suffice. Without more, parallel conduct does not suggest conspiracy, and a conclusory allegation of agreement at some unidentified point does not supply facts adequate to show illegality. Hence, when allegations of parallel conduct are set out in order to make a § 1 claim, they must be placed in a context that raises a suggestion of a preceding agreement, not merely parallel conduct that could just as well be independent action.⁶⁸

With the right configuration of conspirators, however, competitors may not need to communicate directly at all.

3. HUB-AND-SPOKE CONSPIRACY

Imagine an old-fashioned wagon wheel. It has a hub at the center, with spokes radiating out from the hub until they reach the outside rim. All the spokes touch the hub and the rim, but they never touch each other. This is the structure of what is known as a "hub-and- spoke" conspiracy. Competitors (the spokes) are able to collude without ever interacting with each other. Instead, an entity that interacts with all the competitors, such as a supplier, coordinates a conspiracy on behalf of the competitor spokes.⁶⁹

The hub-and-spoke structure involves both vertical and horizontal agreements.⁷⁰ The vertical agreements run between each spoke and the hub; the horizontal agreement exists among all of the spokes and reflects their agreement to participate in the hub’s scheme.⁷¹ Thus, this horizontal agreement is often thought of as the rim of the hub-and-spoke wheel.⁷² Given the combination of vertical and horizontal components with hub-and-spokes cases, the law generally abandons its more lenient view regarding vertical agreements.⁷³

The potential for hub-and-spoke agreements is drawing modern attention with the explosion of Al-powered pricing tools, but these types of conspiracies are not new.⁷⁴ In what is considered the first hub-and-spoke case, the Supreme Court in 1939 accepted the allegation of a conspiracy among theater distributors, orchestrated through a theater

[Page 13]

operator, without any evidence of communications—let alone agreement—between the direct competitors.⁷⁵

The case concerned an alleged conspiracy between eight film distributors who, together, distributed seventy-five percent of all first-class feature films in the United States.⁷⁶ These were the spokes, the ones who should have been competing in the market.

The hub of the scheme was Interstate Circuit, a movie theatre exhibitor who operated forty-three, first-run and second-run theatres throughout Texas.⁷⁷ Interstate had a complete monopoly of first-run theatres in almost all the cities in which they operated.⁷⁸

At the heart of the case was a letter that Interstate Circuit (the hub) sent to each of the eight film distributors in which Interstate demanded compliance with two conditions if the distributors wanted Interstate’s theaters to continue showing the distributors’ films.⁷⁹ The conditions included a minimum price for evening admissions and a double features policy with a requirement that a feature film could not be shown in conjunction with any other feature film.⁸⁰

There was no direct evidence of an agreement between the eight distributors. However, the court focused on their knowledge and motive, noting that they were each copied on the letters, and that they were each aware of the other distributors having been sent the same letter.⁸¹ The Court emphasized:

The O’Donnell letter named on its face as addressees the eight local representatives of the distributors, and so from the beginning each of the distributors knew that the proposals were under consideration by the others. Each was aware that all were in active competition and that without substantially unanimous action with respect to the restrictions for any given territory there was risk of a substantial loss of the business and good will of the subsequent-run and independent exhibitors, but that with it there was the prospect of increased profits. There was, therefore, strong motive for concerted action, full advantage of which was taken by Interstate and Consolidated in presenting their demands to all in a single document.⁸²

While the Court in Interstate Circuit did not coin the term "hub-and-spoke," the name emerged across time as a description for the behavior exemplified in the case.

[Page 14]

As the Interstate Circuit decision established, knowledge and motive of the parties are highly relevant to determining the existence of agreement. This is because, in this configuration, communication and/or agreement between the competitors is made unnecessary by the common third party. Competitors coordinate their conduct through the third party precisely so that they can circumvent the need to directly communicate with each other. This can make collusion harder to detect given that the only agreements that do exist are vertical between each spoke and the hub. Moreover, as discussed earlier in the article, vertical agreements and communication usually are considered a legitimate and common occurrence in a competitive market.⁸³

Hub-and-spoke conspiracies can serve as highly effective collusive mechanisms: As an OECD publication noted, "the hub creates collusive efficiency by reducing the need for horizontal coordination . . . [and] vertical agreements entered to effectuate the horizontal agreement may be harder for authorities to detect."⁸⁴ Given that actual agreements among the spokes are unlikely to exist, and would be difficult to prove in any event, the existence of an agreement turns on other plus factors. Although no prescriptive list of such factors exist, examples include the spokes’ knowledge of the agreements made between the hub and other spokes; a sudden and abrupt change to the business practice of the spokes; and spokes acting against their own self-interest.⁸⁵ The existence of any of the plus factors is not dispositive, and the courts take a totality of the circumstances approach towards finding an agreement based on circumstantial evidence.⁸⁶

Hub-and-spoke conspiracies blur the lines between express collusion (forbidden) and tacit collusion (normally not actionable), as well as horizontal actions (forbidden) and vertical actions (normally not actionable). This blurring of lines, along with the difficulty of developing methods for establishing proof, can sometimes create an uncomfortable result. In Interstate Circuit, for example, each distributor arguably had an independent good commercial reason for agreeing to the exhibitor’s requests after seeing that the competitors had been sent the same request. Although the court found an agreement among the competitors to collude, one cannot help but wonder whether the distributors actually wanted to collude. Perhaps not. As the dissent in the case explained:

[Page 15]

The Government stresses the fact that each of the distributors must have acted with knowledge that some or all of the others would grant or had granted Interstate’s demand. But such knowledge was merely notice to each of them that, if it was successfully to compete for the first-run business in important Texas cities, it must meet the terms of competing distributors or lose the business of Interstate.⁸⁷

The choice for the distributors was to participate or lose significant business. Regardless of any discomfort, Interstate Circuit launched the notion of the hub-and-spoke conspiracy, albeit without coining the name.

C. BRINGING IN THE LAST PIECE—THE PHARMACEUTICAL SUPPLY CHAIN

As pricing and other AI algorithms advance, competition officials have begun to examine the potential implications. For example, recognizing the increasing popularity and capability of algorithmic decision making, the intergovernmental Organization for Economic Co- operation and Development ("OECD") held roundtables in June 2017 and again in June 2023 for experts to discuss and explore the potentially anticompetitive effects of widespread use of algorithms.⁸⁸ In the U.S., legislation has been proposed to regulate the use of algorithms to facilitate collusion,⁸⁹ the Federal Trade Commission ("FTC") has published multiple statements setting out its position with respect to algorithms,⁹⁰ and the Department of Justice ("DOJ") is stepping up its antitrust enforcement efforts.⁹¹

Potential concerns over algorithmic collusion are heightened in markets where competition is concentrated within a small number of firms. One such market is the U.S. pharmaceutical drug market. As will be described in more detail later,⁹² the pharmaceutical drug market is mindbogglingly complex, and much ink has been spilled over the rising

[Page 16]

prices of prescription drugs.⁹³ Important for now, however, is that at the center of more visible players—such as the drug makers and the health plans—lies the pharmacy benefit managers. Pharmacy benefit managers, or PBMs, stand in the middle of the prescription drug supply network offering services to the other players, the manufacturers, the health plans, and the pharmacies.⁹⁴ The concern is that if these middlemen, armed with copious nonpublic data from the players they service in the drug market, turn increasingly to algorithmic decision making, they could become hubs through which they and other players in the market can collude.

II. THE BASIC CONCEPTS IN A MODERN SETTING

The basic competition doctrines have been on the books for almost a century. In addition, the structure of the modern pharmaceutical supply chain developed long before generative AI systems like ChatGPT burst onto the scene. Nevertheless, the modern pathways of AI—which travel well beyond the contours of generative AI systems—are prompting new concerns for competition law.

A. THE AGE OF AI ALGORITHMS

AI has breathed new life into the topic of pricing algorithms and hub-and-spoke conspiracies, with new cases placing the focus on how algorithm providers and/or developers facilitate collusion.⁹⁵ As outlined above pricing algorithms have undergone an evolution brought about by warp-speed advances in the field of AI, similar to the evolution brought about by chat-bots. In particular, older pricing algorithms were rule-based programs which took manually inputted pieces of data—such as market demand, supply levels, and competitor pricing—and applied a preprogrammed set of rules to the inputs to compute and return an optimal price.⁹⁶ Consider the way one might punch numbers and functions into a calculator to obtain a solution. The calculator takes the user’s input, runs it through its preprogrammed rules and calculates the correct answer. These traditional pricing algorithms relied heavily on human input.⁹⁷ Just as with the

[Page 17]

case of chatbots, however, pricing algorithms are now increasingly powered by machine learning.⁹⁸

Although there are many different ways an AI algorithm⁹⁹ could work, the crucial leap forward is in the modern AI algorithm’s relative independence from humans and its ability to adapt and learn.¹⁰⁰ Pricing and other AI algorithms today are trained to learn and adapt by ingesting swathes of data and identifying patterns and correlations in new data that even humans may not have recognized.¹⁰¹ These AI algorithms are likely to become more accurate over time, as both the science improves and the algorithms’ input data and learning improve.¹⁰² As Calvano, Calzolari, Denicolo and Pastorello succinctly put it:

[A]lgorithmic pricing is not new, but the software has recently evolved from rule based to reinforcement learning programs. These latter programs are much more "autonomous" than their precursors. Powered by [AI], they develop their pricing strategies from scratch, engaging in active experimentation and adapting to the evolving environment. In this learning process, they require little or no external guidance.¹⁰³

Pricing algorithms can have procompetitive effects.¹⁰⁴ Nevertheless, particularly in concentrated industries, where there is a high risk of using the same algorithm or similar algorithms trained on the same datasets, the use of AI algorithms might lead to uniform excessive prices.¹⁰⁵ These pricing algorithms have, thus, unsurprisingly formed the basis of many hub-and-spoke antitrust cases.

For example, in Meyer v. Kalanick,¹⁰⁶ an Uber rider filed an action against Uber co-founder Travis Kalanick alleging that Kalanick "conspired with Uber drivers to use Uber’s pricing algorithm to set the prices charged to Uber riders, thereby restricting

[Page 18]

competition among drivers to the detriment of Uber riders."¹⁰⁷ The plaintiff alleged that by using the Uber app, drivers agreed not to compete on price and to set their fee as determined by the Uber algorithm.¹⁰⁸ Moreover, Uber’s "surge pricing" model, a feature of the Uber app that increases standard ride fares during times of high demand, allegedly led to pricing above what would be charged in a normal competitive market.¹⁰⁹ The plaintiffs relied on evidence Uber organized events for drivers, and provided drivers with information about upcoming popular events that would likely boost demand for Uber drivers.¹¹⁰ Remarkably, Uber allegedly occupied eighty percent of the market for "mobile app-generated ride-share" services, with its chief competition, Lyft, occupying a mere twenty percent of the market.¹¹¹ Kalanick was eventually successful in forcing the case into arbitration,¹¹² although not before the District Court had a chance to consider Kalanick’s motion to dismiss.

At the motion to dismiss stage, Kalanick relied heavily on the lack of any horizontal agreement among the drivers to argue that there was no evidence of a conspiracy.¹¹³ Kalanick argued that each driver independently decided that it was in his or her best interest to enter into contract with Uber, which facilitates payment and rider-matching.¹¹⁴ Kalanick argued that while it was a condition of the agreement with Uber that drivers use Uber’s pricing algorithm, this did not diminish the independence of the drivers’ decisions.¹¹⁵ In other words, Kalanick sought to classify the case as one of pure tacit collusion (one without any plus factors), rather than express collusion. The District Court denied Kalanick’s motion to dismiss and, in the process, recognized the capacity of technological advancements to facilitate this type of collusion:

Defendant argues, however, that plaintiff’s alleged conspiracy is "wildly implausible" and "physically impossible," since it involves agreement "among hundreds of thousands of independent transportation providers all across the United States." [] Yet as plaintiff’s counsel pointed out at oral argument, the capacity to orchestrate such an agreement is the "genius" of Mr. Kalanick and his company, which, through the magic of smartphone technology, can invite hundreds of thousands of drivers in far-flung locations to agree to Uber’s terms. [] The advancement of

[Page 19]

technological means for the orchestration of large-scale price-fixing conspiracies need not leave antitrust law behind.¹¹⁶

Uber is the hub, and the spokes are the uber drivers. Indeed, technology has advanced even further in the eight years since the Uber judgment. In In re RealPage,¹¹⁷ a company that provided rental-property pricing software, took pricing and supply data from property owners and managers and used its software to produce price recommendations for rental units.¹¹⁸ The company is said to have acted as a hub by serving as an intermediary between the horizontal property owner and manager competitors. The plaintiffs, private lessees, alleged that RealPage promised its lessor clients that they will outperform the market with software that uses a database of rental prices in each client’s area (including competitors’ prices) and provides the optimal price to charge prospective tenants.¹¹⁹ Users of RealPage’s software must allow the company to use their pricing and supply data in its algorithms to help the client set their own rent prices but also to help set rent prices of its horizontal competitors.¹²⁰ By 2022, RealPage’s software was being used to price over four million multifamily housing units across the United States.¹²¹

The RealPage case involved consolidated class actions in multidistrict litigation. Both classes alleged a hub-and-spoke conspiracy, although the evidence differed for each, and the case also involved other allegations. The district court ruled on the motion to dismiss in December of 2023, granting dismissal in part and denying dismissal in part. On the hub- and-spoke allegations, the court held that although the evidence alleged was not sufficiently strong to fall within the purview of a per se finding for either class,¹²² one of the classes alleged sufficient factors to constitute a plausible allegation under the rule of reason.¹²³

RealPage had urged the court to fully dismiss the hub-and-spoke allegations based on Gibson, in which the district court dismissed allegations that a group of Las Vegas hotels used a revenue management system to collude in setting room rates. The court was not convinced by the Gibson comparison, however, pointing out that unlike Gibson where it was unclear whether the revenue management system recommendations were based on confidential inputs, the plaintiffs here alleged that RealPage’s software produces recommendations based on confidential competitor inputs.

In rejecting the Gibson comparison, the court emphasized the undisputed evidence that each client provided RealPage its proprietary commercial data knowing that it would require the same from its horizontal competitors in order to recommend rental prices.¹²⁴

[Page 20]

In other words, sharing one’s proprietary information generally is not in one’s own interests. Thus, absent knowing that competitors would follow suit, the "spokes" chose an action that would have been against their interests and in this kind of case, knowing that competitors would follow suit can thus provide plausible evidence of engaging in a conspiracy.

Again, in line with Interstate Circuit, this case turned on the knowledge of the horizontal competitors. As the court put it:

It would clearly not be in any individual Defendant’s economic self-interest to contribute its data to RealPage without knowing that it would benefit from its horizontal competitors doing the same. Put another way, the contribution of sensitive pricing and supply data for use by RealPage to recommend prices for competitor units is in Defendants’ economic self- interest if and only if Defendants know they are receiving in return the benefit of their competitors’ data in pricing their own units.¹²⁵

In a similar vein, a series of cases relating to real property have arisen, each alleging that a pricing algorithm enabled competitors to collude.¹²⁶ In Duffy v. Yardi,¹²⁷ the plaintiff alleged that a group of landlords used Yardi’s pricing algorithms to allegedly inflate rental prices.¹²⁸ Specifically, the plaintiffs alleged that Yardi and its clients exchanged confidential information with each other through Yardi’s automated pricing software, as well as that the defendants engaged in direct conversations with their competitors about pricing through market surveys.¹²⁹ Along the same lines, in Cornish-Adebiyi v. Caesars,¹³⁰ the plaintiffs alleged that a group of hotels used a revenue management company called Cendyn, which provides an algorithmic pricing software known as "Rainmaker" that gathers real-time pricing and occupancy data from the hotels, to generate an optimal room rate for the hotels at supra-competitive rates.¹³¹

As outlined above, this phase of antitrust action against algorithmic collusion was shaped initially by aggrieved private litigants. Although the DOJ and the FTC had been increasingly vocal over the use of algorithms and the attendant anticompetition

[Page 21]

risk factors,¹³² until August 2024, this was largely relegated to filing joint statements of interest in pending algorithmic collusion cases.¹³³ The trajectory of antitrust enforcement concerning pricing algorithms is now, however, witnessing a seismic shift. What began as a series of lawsuits spearheaded by aggrieved private parties has now burgeoned into a broader regulatory offensive marked by the DOJ’s lawsuit against RealPage, Inc., and their landlord clients.¹³⁴ In its recent lawsuit the DOJ, joined by 8 states, alleges various violations of antitrust law, underscoring two crucial facts that the District Court had emphasized in In Re RealPage.¹³⁵ First, the DOJ asserted that each landlord used RealPage’s software "knowing or learning that RealPage will use this data to train its models and provide . . . pricing not only for the landlord, but for the landlord’s competitors (and vice versa)."¹³⁶ Second, the DOJ stressed that each landlord had agreed to provide to each other "nonpublic, competitively sensitive transactional data" through providing such data to RealPage for use in its algorithmic models.¹³⁷

The DOJ further argued that in certain industries, such as the rental multifamily housing industry, widespread use of algorithms can facilitate coordination among competitors making such use more likely to "restrain, rather than promote, competition."¹³⁸ Specifically, the DOJ identified inelastic demand for rental housing, and concentration among landlords in local markets as "industry characteristics [that] exacerbate the harm to the competitive process[.]"¹³⁹Although the connection between rental properties and pharmaceutical drugs may not be immediately apparent, the inelastic demand and market concentration that the DOJ identified could, as will be shown below, just as well be used to describe the pharmaceutical drug industry.¹⁴⁰

[Page 22]

Although the cases described above concern the use of AI algorithms in the context of real property, AI algorithms are ubiquitous.¹⁴¹ For example, in May 2024, a class action complaint was filed on behalf of healthcare providers against a company that provided a pricing algorithm to health plans.¹⁴² The complaint alleges that MultiPlan, Inc., a company that helps health plans "identify and negotiate fair reimbursements for out-of-network claims[,]"¹⁴³ suppressed and fixed the rates at which its health plan clients reimbursed providers for out-of-network services.¹⁴⁴ According to the plaintiffs, plans including the powerhouses Cigna, Aetna, and UnitedHealth¹⁴⁵ agreed to allow MultiPlan to use their "competitively sensitive reimbursement data to help drive [MultiPlan’s] algorithm[,]" which "repriced" out-of-service providers’ reimbursement rates to "near-uniform suppressed rates[.]"¹⁴⁶ The plaintiffs argue that this collusion is a per se horizontal price- fixing conspiracy and that, in the alternative, the conduct constitutes an unlawful hub-and-spoke conspiracy.¹⁴⁷ As will be discussed,¹⁴⁸ under the right market conditions, algorithmic decision-making tools make collusion amongst competitors not just a possibility, but a near certainty.

B. EXISTING LITERATURE ON AI AND COMPETITION

As previously discussed, the use of algorithms is not new. However, academic literature discussing the impact of AI on competition law is a more recent development. In a series of publications, Professors Ezrachi and Stucke make significant strides in

[Page 23]

helping frame the way one might think about algorithmic collusion.¹⁴⁹ They identify four scenarios of collusion that might occur using computer algorithms. They call these scenarios: Messenger, Hub-and-spoke, Predictable Agent, and Autonomous Machine.¹⁵⁰

In the Messenger scenario, humans orchestrate the collusion but are aided by computers.¹⁵¹ Professors Ezrachi and Stucke explained that in this "basic" scenario, the enforcement is straightforward.¹⁵² An agreement between the humans to collude and restrict competition does not become legal just because they used a machine to facilitate it.¹⁵³

Their second category, Hub-and-spoke, envisions a single algorithm dictating market prices that will be charged by multiple users of the algorithm.¹⁵⁴ The authors argue that in this scenario, although a single agreement between the algorithm provider and the user may not in itself produce any anticompetitive effect, the result could become anticompetitive if a cluster of such agreements exists.¹⁵⁵ They emphasize the importance of intent evidence here, in what they view as a "classic hub-and-spoke conspiracy," and give the example of companies agreeing to use the same algorithm to raise prices.¹⁵⁶

Predictable Agent, their third category, involves humans each separately designing a machine to achieve a specific goal by reacting to definite changes in the market.¹⁵⁷ Assuming no intentional agreement or coordination in the market exists, but that an industry adopts similar algorithms, Ezrachi and Stucke argue that such a market "may exhibit the conditions for tacit collusion."¹⁵⁸

The final scenario, the Autonomous Machine, hypothesizes a market in which competitors separately and independently create a computer algorithm to achieve a definite goal, and the machines autonomously, through self-learning, decide how to achieve the goal.¹⁵⁹ Ezrachi and Stucke regard this category of possible collusion to be most tricky.¹⁶⁰

Ezrachi and Stucke’s framework provides a useful tool for conceptualizing the ways in which AI might impact competition law. In this paper, however, we consider AI and

[Page 24]

competition in its industry context. Indeed, Ezrachi and Stucke recognized that some markets will be more susceptible to algorithmic collusion than others: "Undoubtedly, concerns within the legal and economic communities about the phenomenon increased with the improvement in computing power that enabled more robust market data collection. In concentrated markets, algorithmic tacit collusion may form a business strategy rather than merely reflect a market outcome."¹⁶¹

The next sections, therefore, build on Feldman’s earlier work regarding both the pharmaceutical supply chain and competition concerns posed by algorithms.¹⁶² We will take a deeper dive into how hub-and-spoke collusions may arise specifically in the pharmaceutical drug market and how AI algorithms could exacerbate already significant antitrust concerns.

C. THE PHARMACEUTICAL DRUG INDUSTRY: ALL ROADS LEAD TO PBMS

The pharmaceutical drug-delivery system, which brings a drug from the manufacturer across multiple steps to the consumer, is staggeringly complex.¹⁶³ For example, pharmaceutical companies invest in research and develop a brand drug; physicians write prescriptions for the drug; pharmacies fill the prescriptions; and health insurance plans, or employer-sponsored plans, foot the bill (the payor). Finally, there is the patient. These players will no doubt be familiar. However, weaved in among these players are intermediaries who may be foreign. These are Group Purchasing Organizations, Wholesalers, Pharmacy Services Administrative Organizations, and arguably the most central intermediary of them all, Pharmacy Benefit Managers (PBMs) and their consultants. To try to follow the pharmaceutical drug distribution pipeline, one would eventually come across a PBM.¹⁶⁴ In theory, PBMs work for the payor, helping to ensure that claims are processed properly and to keep drug prices low. However, as will become clear, the interests of the PBMs and the payors often diverge.

PBMs offer various services to the payor. First, PBMs help payors process patient and pharmacy claims.¹⁶⁵ After you see the doctor, your doctor’s office submits a bill to the insurer. Although the bill is submitted to the payor, it is the PBM that processes the claim to figure out the coverage that you—the patient—has, whether any copayments are needed, whether you have met your deductible, and ultimately what the insurance needs to pay. The next stage arises when you go to the pharmacy to obtain medicine prescribed to you. The pharmacy will likely charge you an out-of-pocket fee for the

[Page 25]

prescription, which may reflect your deductible and/or copayments or coinsurance under your policy. After the pharmacy charges you, however, it submits a claim to the PBM for reimbursement of the insurer’s share of the prescription dispensed. This is because pharmacies front much of the cost of the drugs when first obtaining the drugs from wholesalers, so the drugs will be sitting on the shelves when the patient walks in. Thus, in addition to processing your doctor’s claim, the PBMs also process these claims from pharmacies based on the rate that the PBMs previously negotiated with the pharmacy.

Second, PBMs negotiate drug prices with drug manufacturers on behalf of payors.¹⁶⁶ How does this work? A drug company sets an initial price for its drug; this is the drug’s list price. A PBM representing a payor says to the drug company, "I have one million lives under this payor’s plan; I can offer you access to these patients, in return for a discount on the drug." This is a request for a simple volume discount. Put simply, if the pharmaceutical company can boost its sales volume of the drug that they manufacture, the company’s overall revenue may increase even if that means tightening the profit margin per unit. Thus, pharmaceutical companies agree to a give a discount, which in this industry, comes in the form of a rebate delivered to the plan later on, rather than as a discount on the price at the front end.

As the industry has evolved, however, the PBMs’ remuneration from their payor clients is based on the rebate that the PBM promised to achieve.¹⁶⁷ Thus, the payor might say to its PBM, "negotiate the biggest discount you can get on the drug prices, and I’ll give you half the discount." Of course, the PBM could try to negotiate tooth and nail with the pharmaceutical company to bring the list price of a drug down. However, should the pharmaceutical company simply decide to increase the list price and offer a bigger rebate, a PBM can hardly be expected to resist. It’s a win-win: The PBM gets paid significant sums of money, and the pharmaceutical company can expect high sales volume from the happy PBM. In this way, the original lofty goal of helping its payor client keep drug prices low is subverted by the PBM’s own financial interests.

As a side note, on top of payment from the payors for their services, PBMs also get paid by the pharmaceutical companies—the very same companies with whom they are supposed to negotiate. A recent study funded by the Pharmaceutical Research and Manufacturers of America (PhRMA) identified at least three additional sources of income for PBMs paid by the manufacturer: administrative fees, data and data portal fees, and the group purchasing organization vendor fees.¹⁶⁸ These fees generally rise when drug prices rise, again aligning the interests of the intermediary and the drug companies—

[Page 26]

given that both benefit when prices go up—against whom they are supposed to be negotiating vigorously.

Apart from volume, PBMs hold another key, one that opens the gate to the formulary.¹⁶⁹ Formularies are yet another confusing aspect of the U.S. healthcare system. The formulary is a list of drugs covered by a health plan. If a drug is not on a health plan’s formulary, patients will not be reimbursed by the plan if they purchase the drug. In a formulary, drugs are divided into tiers, in which each tier represents a level of patient cost-sharing.¹⁷⁰ In addition to cost-sharing, the drug might also be subject to what is known as "utilization management"—a set of restrictions on that drug to reduce costs for the payor—which, unless satisfied, could result in the drug price not being covered by the plan. These utilization management restrictions could require patients to obtain prior authorization from the plan before they fill their prescription. Another common utilization management tool is to restrict the quantity of a drug that will be covered by the plan. The formulary, therefore, not only lists all the drugs that the plan will cover, but also determines the extent of the patients’ out of pocket costs and the ease with which a patient can obtain (and a physician can prescribe) the drug through utilization management.¹⁷¹ Thus, PBMs are, quite literally, the gatekeepers to whether and how a drug will be covered by payors.

Drug manufacturers want their drugs to be on the formulary, and they also want their drugs in favorable positions on the formulary,¹⁷² ideally with as few utilization controls as possible. If a competing drug in the same class could be excluded from the formulary, then that would be great for the manufacturer, too.¹⁷³ Who holds the key to the formulary? The PBMs, a fact that is not immediately obvious to most. Often, on health plan websites, formulary decisions are said to be made by a pharmacy and therapeutics (P&T) committee (or equivalent entity) of doctors and pharmacists,¹⁷⁴ who consider the drug’s effectiveness,

[Page 27]

safety, and cost.¹⁷⁵ However, when one looks closely, our omnipresent friend rears its head again. PBMs offer formulary design services. For many plans, rather than having its own P&T committee, the plan chooses to contract with a PBM for its formulary design services. The PBM, using its own P&T committee or equivalent, then performs the role of developing formularies for the plan, saving the plan significant amounts of administrative labor.¹⁷⁶ There is no doubt that the PBMs hold great power through their control over formulary design, and formulary design can have a significant effect on patients.¹⁷⁷

The PBM’s salient position in the pharmaceutical drug supply network gives it unparalleled access and control over nonpublic claims data. As processor of patient and even pharmacy claims, the PBM possesses a wealth of information that provides insight into consumer buying patterns: the impact of different drug formulary positions on consumer behavior, the effect of pricing and utilization management decisions, etc. The data flows from more than a single plan. The PBM has that data with respect to patients covered by each of the payors it works for. No single payor has access to this wealth of nonpublic data. Even within a single payor, the contract that a payor signs with the PBM usually allows only limited audit rights if any at all. Furthermore, the PBM does not divulge any information with respect to its negotiations with the pharmaceutical companies, not even the rebated price per drug.¹⁷⁸ This secrecy that the PBMs perpetrate keeps the extent of rebates obtained, the net price per drug negotiated, and the PBMs’ remuneration structures hidden. In short, this creates an information asymmetry. Although this has led to recent calls for legislation that would force PBMs to disclose their remuneration

[Page 28]

structure and deals with pharmaceutical companies,¹⁷⁹ these attempts have been met with rigorous pushback from the PBM industry.

In addition, the PBM industry is extremely concentrated. Just three PBMs manage roughly 80% of the U.S. prescription drug market.¹⁸⁰ Given such dominance, and the fact that pharmaceutical companies are also feeding a PBM information in the form of sensitive pricing data, the PBM will have effectively funneled multiple data sources into one, including multiple levels of data that other vertical and horizontal players cannot access. Moreover, in recent years, the PBM industry, and pharmaceutical industry more broadly, have seen aggressive consolidation through mergers and acquisitions.¹⁸¹ PBMs own and are owned by insurers. They also own their own pharmacies and increasingly have physician practices. Through these tactical corporate maneuvers over the years, the lines between PBM, pharmacy, provider, and insurer are much fuzzier than one would expect.¹⁸² Undoubtedly, the convergence of roles makes the PBM even more powerful. Nevertheless, these tactical maneuvers have largely escaped antitrust scrutiny.¹⁸³

One might wonder why health insurance companies tolerate a system that drives up the drug prices that they ultimately pay for. After all, many insurance companies are well-funded and sophisticated market players. Why would they continue to participate in a market transaction that appears to place them at a disadvantage?

A number of market factors may explain the failure of health insurers to hold the line on rising drug prices. First, an insurer’s business is health care. The nitty gritty of patient and provider claims administration, adjudication, and processing is not the insurer’s core business, and, therefore, may be of less interest from a business perspective. Delegating that task, however, has inadvertently led to relinquishing the data associated with all these claims—data that, as discussed, PBMs guard fiercely.

[Page 29]

With the increasing importance of data in the modern context, that relinquishment now creates outsized disadvantages for the insurer. Thus, some insurers—particularly those who wield strength in the market through their size and patient flow—have begun to request access to their PBM’s files for claims, rebates, and other data at least with respect to its own plans. The "big three" PBMs, however, tend to move in lockstep, offering the same client terms to all comers. Conversations Feldman has had with large players suggest that they are unable to negotiate access to their own data in their contracts with the major PBMs during the complexity of contract negotiations.¹⁸⁴ Even if access is granted, an insurer would need to dedicate a significant amount of time and money to analyze—not to mention to develop the expertise—to understand and interpret the masses of complicated claims and pharmaceutical data. Smaller PBMs might be willing to deviate from the standard terms offered, but they may not have the capacity to handle a major health insurer’s flow, or they may not be able to offer the experience that a profit-making large health insurer would need to justify transferring its business. Most importantly, this discussion assumes the payor is a large and sophisticated insurance company, rather than a smaller health insurer.

Moreover, when the health insurer is a large player in the market, PBMs have begun to offer "price-protection" plans, if the insurer presses for access to its own data. Why would an insurer want to engage in the cost and difficulty of interpreting masses of complex data, when the PBM guarantees that the plan’s bottom-line costs for drugs will not increase more than 2%, 4%, or something alike? Drug prices are the only bottom-line issue clamoring for an insurer’s attention. Of course, that deal allows the current system to continue—one that locks in current high prices as a floor and can operate to disadvantage newer and less-expensive competitors, given the intricacies of the current pharmaceutical supply chain system.¹⁸⁵

In a similar vein, insurance companies, in theory, could also negotiate a pass-through of rebates. This model requires a PBM to pass on a greater percentage or all of the negotiated rebates to the insurance company as the payor.¹⁸⁶ Yet, for a PBM to agree to forego such a lucrative revenue stream, the PBM will demand certainty in the form of

[Page 30]

charging its insurer client higher administrative fees.¹⁸⁷ Some insurers might opt for pass-through pricing models but some might reasonably decide that all that hassle with data, plus higher administrative fees, is simply not worth it. Even in cases where PBMs do pass rebates through to the health insurer, that does not necessarily mean those savings are passed to the patients.¹⁸⁸ Critically, PBMs in the last few years have begun to outsource certain activities to what are being called "rebate GPOs," often located off-shore—a move that insulates rebate activity from falling within the definition of money flows that must be passed through to the payor.¹⁸⁹

In addition, today the largest payors are the PBMs.¹⁹⁰ With recent integration, the three largest PBMs, CVS Caremark, Express Scripts, and OptumRx are vertically integrated with the three largest health insurance companies, Aetna, Cigna, and UnitedHealthcare.¹⁹¹ Indeed, according to a study by the American Medical Association, 69% of commercial drug insurance coverage in 2023 was provided by health insurers that are vertically integrated with a PBM.¹⁹² Some suggest this integration might explain why even where PBMs received significant pharmaceutical company rebates, and why, in the case where a sizeable portion of rebates were passed to the payors, less than 1% of the rebate was passed on to the patients.¹⁹³ Vertical integration is not necessarily problematic on its own. Given the intricacies of the health insurance system and the pharmaceutical supply chain industry, however, the configuration may raise concerns in certain circumstances.

The consolidation of PBMs and insurers also leaves smaller PBMs and smaller insurers in a tough position. A PBM’s value depends heavily on a network effect. The higher the number of big insurance clients it has, the more lives it can offer pharmaceutical companies, and the more it can leverage volume in negotiations against pharmaceutical companies. Thus, insurance companies will be more inclined to entrust their plans to the biggest PBMs reinforcing the industry concentration. Smaller PBMs are also hard-pressed not to partner with larger PBMs,¹⁹⁴ which again intensifies concentration in the PBM market. This occurs in a market where many large insurers have little motivation to challenge their PBMs, and smaller insurers have little bargaining power against their PBMs.

Finally, it is important to realize that in certain circumstances, the incentives can stack up in such a way that insurers might actually prefer higher drug prices, given unintended consequences from: 1) the way in which the Affordable Care Act of 2010 limits a plan’s

[Page 31]

ability to profit from insurance dollars, unless input costs rise and 2) the fact that with Medicare plans, higher prices push patients sooner into the territory in which the government pays for more of the cost of the drug through reinsurance.¹⁹⁵ In short, as with any profit-making entity, most insurance companies are ultimately driven by the bottom line. In that context, lack of effective action from insurance companies is unsurprising.¹⁹⁶

D. PBMS AS THE HUB

In a less technologically advanced era, and before PBMs developed their current level of concentration, the idea of a PBM collecting, processing, and analyzing even nonpublic claims data might have sounded innocuous. After all, in a less technologically advanced era, processing, analyzing, and making predictions based on—often voluminous—data might seem more like a burden than an opportunity. With current technological capabilities, however, PBMs can process swathes of claims data using AI algorithms that are designed to optimize pricing and decision-making. As is the nature with AI, these models will learn from more and more data, getting better at recognizing patterns and making predictions.¹⁹⁷ From the facts above, the idea that the PBM might play a critical role in collusion starts to take shape, and it does not take a giant leap to see why the pharmaceutical industry presents such a perfect opportunity.

Consider this hypothetical: Manufacturer A produces a drug, Life Elixir, which has three therapeutically equivalent alternatives produced by Manufacturers B, C, and D. Ideally, each company could maximize its market share by ensuring that patients use, and physicians prescribe, their version of Life Elixir, and maximize profit by putting the price of Life Elixir and its alternatives up as high as possible. PBM-X has all the historic data of patient claims made under Plan A, Plan B, and Plan C, all of which are clients of PBM-X. Being one of only three PBMs in the market, PBM-X also has relationships with Manufacturers B, C, and D. Therefore, PBM-X also has pricing information relating to the Life Elixir alternatives.

Here, PBM-X could, on its own, develop the optimal package of pricing and formulary factors that would increase all the Life Elixir companies’ income, avoid competition, and

[Page 32]

maximize the PBM’s own income, which is also based on higher prices. The circumstance would be a win-win for everyone—except for competition and the consumer.

Any of the companies—or more likely all of them—could ask PBM-X "what can we do to keep our market share up and profits as high as possible?" PBM-X could then run the masses of data it has through an algorithm whose objective is to produce the optimal price and formulary positions of each of the versions of Life Elixir to allow each company to maximize profits. PBM-X then can suggest each of these strategies to Manufacturers A, B, C, and D.

Whether the manufacturers ask for the strategy, or PBM-X came up with the strategy of its own volition, the hypothetical could play out in a few different ways. In one scenario, the pharmaceutical companies intended to collude. In this scenario, a few pharmaceutical companies purposefully approached the PBM knowing that each of the other pharmaceutical companies was doing the same. In this scenario, a case could be made that the pharmaceutical companies agreed to collude since they used an algorithm supplied by the hub knowing that their competitors would do the same thing and that such convergence would likely lead to uniformity in the output. Indeed, this is the scenario that Ezrachi and Stucke envisaged; it is also in line with the FTC and DOJ’s position.¹⁹⁸

An alternative scenario is that none of the pharmaceutical companies intended to collude. They are simply operating as they would within this market. The contrast between this scenario and the one above is explained in the United States’ Note to the OECD. Where competitors entered into agreements with one company to use their pricing algorithm, and there is evidence to show that they did so "with the common understanding that all of the other competitors would use the identical algorithm, that evidence could be used to prove an agreement among the competitors."¹⁹⁹ This is in contrast to situations where a company enters into an agreement with the hub to use a pricing algorithm while completely oblivious to the fact that its competitors are doing the same.²⁰⁰ In these cases, even if the use of the same or similar algorithms has the effect of aligning prices, no antitrust liability attaches.²⁰¹ Thus, a prosecutor might try to argue that each pharmaceutical company knowingly used the same PBM, who clearly offered the same algorithm, but the argument is not without its flaws. The pharmaceutical companies would simply reply that the market is structured this way, and it is not their fault that there are effectively only three PBMs in the market to choose from. To access the health plan

[Page 33]

clients, the pharmaceutical companies need to negotiate with the PBMs, and of course they knew their competitors are likely negotiating with PBM-X, given the lack of choice of PBMs.

Moreover, even if the pharmaceutical company, on some level, suspects that its competitors are using the same or a similar algorithm, the pharmaceutical company still has a rejoinder. There are seemingly no actions they could possibly take to demonstrate good-faith efforts to comply with the law. What safe-harbors or behavior could they follow?

To be perfectly fair, is there any way that pharmaceutical companies could benefit from a PBM’s algorithm without violating antitrust laws? Where there are no real market alternatives, competitors cannot help but converge on a single or set of similar algorithms. If it is not a hub-and-spoke conspiracy with PBM-X, it would be a hub-and-spoke conspiracy with PBM-Y.

This is a challenging aspect of the pharmaceutical supply chain: The PBM industry itself is oligopolistic.²⁰² As with In re Text Messaging, many of the actions by the PBMs and the pharmaceutical companies do not in fact require any agreement or intent to collude. Anticompetitive effects may arise irrespective of agreement or intent due to the oligopolistic markets at play.

Moreover, this oligopolistic characteristic can be seen not only at the PBM level but also at the pharmaceutical company level. An APSE study²⁰³ found that about 300 simple molecule drugs (16% in 2022) had two to three manufacturers, and about 150 drugs (8% in 2022) had four to five manufacturers. Similarly, thirty-four biologic products (13%) had two to three manufacturers, and twelve products (4%) had four to five manufacturers. Only six biologic products had more than five manufacturers.

Although many commentators describe the pharmaceutical industry as monopolistic,²⁰⁴ it is the oligopolistic structures that are of interest to our analysis here. Specifically, this article has discussed the potential of AI to facilitate collusion, which can only arise where

[Page 34]

there are at least two independent firms competing in a market. Thus, our focus is on the pharmaceutical industry’s oligopolies.

Looking further into the future, there is yet another alternative and perhaps more complicated scenario. Humans are less than perfect creatures, and the human brain puts its individualistic stamp on each, less-than-perfect decision. As algorithms continually improve, each one may increasingly be able to approach the perfect answer to the problem of maximizing profits. In other words, if there is a perfect answer, and algorithms continually improve, we may reach a point when our near perfect algorithms are all capable of reaching the same close-to-perfect answer. In that case, you have three PBMs, each occupying a large part of the market, each operating its own algorithms whose purpose is simply to decide on the best strategies for maximizing profit. Over time, these three algorithms would converge on outcomes, driving prices up and eliminating pharmaceutical company competition. In such a case, does it matter whether any of the PBMs, or the pharmaceutical companies, intended to collude?

III. THE OUTLOOK

Pressure on pharmacy benefit managers is boiling over.²⁰⁵ Yet, how PBMs operate is still a matter of extreme secrecy. One cannot determine the rebate-per-drug in a PBM’s contract with a pharmaceutical company. Nor can one determine the base on which the PBM has contracted to receive fees from pharmaceutical companies. Nor can we determine how many patients need what drug and at what level of coverage. PBMs not only possess this data, but also, thanks to technological advances, they are now capable of using the data to train AI models that will make increasingly accurate predictions relating to the pharmaceutical industry. This is especially problematic given their already-dominant position in the supply chain. The addition of PBM consultants, using AI algorithms to advise PBMs, may further muddy the waters by adding yet another layer of insulation between the pharmaceutical companies and the algorithm.²⁰⁶

[Page 35]

These concerns over collusion within the pharmaceutical industry are hard to overstate.²⁰⁷ It is no secret that the United States has some of the highest drug prices in the world. A study published in February 2024 found that U.S. prices for brand drugs were 422% of prices in comparison to OECD countries, or 322% when adjusted for estimated rebates in the United States.²⁰⁸ Concerns over exorbitant drug prices are widely expressed by academics, policymakers, patients, and health providers, with even drug companies themselves acknowledging the problem.²⁰⁹

U.S. competition authorities lack the requisite tools to promote a competitive marketplace while treating all players in the market fairly and reasonably. This requires a much clearer hub-and-spoke framework that works for the pharmaceutical industry and enables better detection of collusive agreements among players such as pharmaceutical companies and PBMs, especially in light of today’s and tomorrow’s advancements in AI.

Although significant further research and development is required, the authors hypothesize a compliance mechanism modelled on Know Your Client (KYC) and Anti-Money Laundering compliance schemes. Broadly speaking, these mechanisms place banks, and other facilitators of transactions, under a legal obligation to conduct certain checks with respect to taking on new clients and monitoring transactions to identify and effectively resolve suspicions of money laundering. The actual requirements of these compliance

[Page 36]

regimes vary around the world.²¹⁰ However, they are often guided by the Financial Action Task Force (FATF), an intergovernmental money laundering watchdog.²¹¹ The FATF makes recommendations for a comprehensive anti-money laundering framework that form the basis on which countries can build their own regulations.²¹² The FATF also identifies jurisdictions vulnerable to money laundering and urges them into action.

Anti-money laundering regimes usually require, at first, a risk assessment of a transaction based off a checklist of potential red flags.²¹³ The obligation to perform these risk assessments is usually placed on a facilitator of the transaction—such as banks and law firms. These red flags can be raised over the purpose of a transaction, the country from which the money originates, the political exposure of the person making the transaction, and so on. After the level of risk is determined, a correspondingly comprehensive investigation must take place, taking care not to tip off the person making the transaction. If the legitimacy of the transaction cannot be verified, the matter gets referred to an internal appointed compliance officer. Finally, if still unsatisfied with the legitimacy of the transaction, the officer is empowered to make a suspicious activity report to the relevant regulatory authority.

This kind of model for compliance has three key strengths that could help in the detection of potentially collusive agreements:

(1) The mandatory risk assessment.²¹⁴ An antitrust compliance checklist is not a new idea. Indeed, law firms around the world help their clients establish internal antitrust policies to curb the risk of being found in violation of national antitrust

[Page 37]

laws.²¹⁵ These checklists are voluntary; there is no obligation on the business to have an anticompetition policy in place. While some big companies do set out anticompetition policies,²¹⁶ there is no uniform framework set up for market participants to follow.
(2) An obligatory system of compliance forces the firm to keep a paper trail of checks performed and decisions made. Of course, anti-money laundering checks do not always prevent money laundering. Nevertheless, banks have a clear set of obligations that they need to satisfy in order to discharge their duty. Moreover, when the regulator comes knocking, parties are sure to have a clear paper trail to show that they complied. The paper trail might assist regulators in monitoring activities within industries that could have an anticompetitive effect. It could also assist companies with developing a framework for compliance, rather than being buffeted by the vagaries of the moment.
(3) The run-up-the-ladder reporting mechanism. Under normal circumstances, the bank has many privacy obligations owed to its clients. However, this privacy concern is superseded by its obligation to report suspicious activity, provided that the bank has good reasons for its suspicions. In the pharmaceutical industry, this is particularly important given that many aspects of the industry are shielded in secrecy. Calls for more public transparency regarding, for example, a PBM’s remuneration terms have been met with the argument that such transparency would actually make collusion more likely.²¹⁷ This more limited form of transparency could enable regulators to, in appropriate situations, take a peek behind the curtain so that they can in fact assess what has happened.

A decision would of course need to be made as to the level at which to impose such an obligation. Further research is also needed to develop a uniform checklist that would in fact serve the purpose of preventing collusion, rather than becoming an easy way for a market player to claim it has done its duty.

While we, as a society, wait to see just how far and wide AI will go, the path ahead will become increasingly clear. For now, what is clear is that the convergence of AI,

[Page 38]

antitrust law, and the pharmaceutical industry necessitates proactive regulatory measures to safeguard competition and consumer welfare. By proactively addressing hub-and-spoke collusion facilitated by AI algorithms, policymakers can foster a more competitive and transparent pharmaceutical market, ultimately benefiting patients and society at large.

[Page 39]

——–

Notes:

^1. Robin Feldman is Arthur J. Goldberg Distinguished Professor of Law, Albert Abramson ’54 Distinguished Professor of Law Chair, Director of the AI Law & Innovation Institute, University of California Law. Caroline A. Yuen is Adjunct Professor of Law, University of California Law and Senior Researcher at the AI Law & Innovation Institute, University of California Law. The authors wish to thank Alexander Whisnant for outstanding research assistance. We are also deeply grateful to Zachary Henderson, Merav Magen, and Gideon Schor for their insights and comments on various drafts of the work.

^2. See McKinsey & Company, The state of AI in 2023: Generative AI’s breakout year (Aug. 1, 2023), https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-in-2023-generative-AIs-breakout-year (survey finding one-third of all respondents reporting that their organizations regularly used generative AI in at least one function); Joe McKendrick, AI Adoption Skyrocketed Over the Last 18 Months, Harv. Bus. Rev. (Sept. 27, 2021), https://hbr.org/2021/09/ai-adoption-skyrocketed-over-the-last-18-months; Michelle Faverio & Alec Tyson, What the Data Says About Americans’ Views of Artificial Intelligence, Pew Rsch. Ctr. (Nov. 21, 2023), https://www.pewresearch.org/short-reads/2023/11/21/what-the-data-says-about-americans-views-of-artificial-intelligence/ (survey showing ninety percent of Americans have heard "at least a little" about artificial intelligence).

^3. See, e.g., Trevor Laurence Jockims, How Spotify AI Plans to Know before You Do What Your Brain Wants to Listen to, CNBC (Apr. 14, 2024, 11:35 AM), https://www.cnbc.com/2024/04/14/how-spotify-ai-plans-to-know-whats-going-on-inside-your-head.html (discussing Spotify’s AI disc jockey); Barbara Krasnoff, Google Gemini, Explained, Verge (June 12, 2024, 7:39 AM), https://www.theverge.com/24176860/google-gemini-explained-ai-assistant (explaining the various uses of Google’s Gemini AI including its function in summarizing search results).

^4. See, e.g., Salvador Ordorica, Comparing and Contrasting AI and Human Translation, Forbes (June 5, 2023, 10:15 AM), https://www.forbes.com/councils/forbesbusinesscouncil/2023/06/05/comparing-and-contrasting-ai-and-human-translation/ (recognizing AI translation as a "blossoming subfield of natural language processing. . . ."); Petr Malyukov, How Global Businesses Can Use AI Translation To Improve Productivity, Forbes (last updated Mar. 11, 2022), https://www.forbes.com/councils/forbesbusinesscouncil/2022/03/10/how-global-businesses-can-use-ai-translation-to-improve-productivity/ ("Language barrier difficulties are already being successfully tackled by AI translators. Recent breakthroughs in the field of machine learning have helped translation evolve by leaps and bounds").

^5. See, e.g., Adam Zewe, Scientists Use Generative AI to Answer Complex Questions in Physics, MIT News (May 16, 2024), https://news.mit.edu/2024/scientists-use-generative-ai-complex-questions-physics-0516 (discussing the use of AI to detect phase changes ); Shira Polan, AI Uncovers Hidden Differences in Male and Female Brain Structures, Neuroscience News (May 14, 2024), https://neurosciencenews.com/ai-brain-sex-differences-26101/ (noting the use of AI to distinguish between biological male and female brains by identifying patterns in structure and complexity); Sara Frueh, How AI is Shaping Scientific Discovery, Nat’l Acads. (Nov. 6, 2023), https://www.nationalacademies.org/news/2023/11/how-ai-is-shaping-scientific-discovery (highlighting how AI can be used to identify trends in large datasets, predict outcomes, and simulate complex scenarios).

^6. Brian Kennedy et al., Public Awareness of Artificial Intelligence in Everyday Activities, Pew Rsch. Ctr. (Feb. 15, 2023), https://www.pewresearch.org/science/2023/02/15/public-awareness-of-artificial-intelligence-in-everyday-activities/. See also Simon Fondrie-Teitler & Amritha Jayanti, Consumers Are Voicing Concerns About AI, Fed. Trade Comm’n (Oct. 3, 2023), https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2023/10/consumers-are-voicing-concerns-about-ai.

^7. See, e.g., Maneesha Mithal et al., Colorado Passes First-in-Nation Artificial Intelligence Act, Wilson Sonsini (May 21, 2024), https://www.wsgr.com/en/insights/colorado-passes-first-in-nation-artificial-intelligence-act.html (Colorado made history as the first US State to pass its own comprehensive AI Act.); Joshua M. Goodman et al., AI and Algorithmic Pricing: Current Issues and Compliance Considerations, Morgan Lewis (Apr. 29, 2024), https://www.morganlewis.com/pubs/2024/04/ai-and-algorithmic-pricing-current-issues-and-compliance-considerations (citing to the flurry of policy and legislative activity surrounding AI issues since February 2024).

^8. See, e.g., In re OpenAI ChatGPT Litigation, No. 23-cv-03223-AMO (N.D. Cal. Mar. 1, 2024) (consolidating three lawsuits brought by authors alleging copyright infringement related to OpenAI’s ChatGPT tool); Getty Images (US), Inc. v. Stability AI, Inc., No. 23-cv-00135-JLH (D. Del. filed Feb. 3, 2023) (alleging that Stability AI infringed the copyright of millions of photographs in training their AI product, Stable Diffusion); Andersen et al. v. Stability AI Ltd. et al., No. 23-cv-00201-WHO (N.D. Cal. Aug. 12, 2024) (class action brought by artists against Stability AI, Midjourney Inc., and DeviantArt alleging that the companies used copyrighted materials to create each company’s respective generative AI tools).

^9. See, e.g., Intercept Media Inc. v. OpenAI, Inc., No. 1:24-cv-01515 (S.D.N.Y. filed Feb. 28, 2024) (alleging DMCA violations for inclusion of journalists’ work in training data for AI product); N.Y. Times Co. v. Microsoft Corp., No. 1:23-cv-11195-SHS (S.D.N.Y. filed Dec. 27, 2023) (alleging defendant infringed copyrights by using journalistic reporting as training data for generative AI product); Authors Guild v. OpenAI, Inc., No. 1:23-cv-08292 (S.D.N.Y. filed Feb. 2, 2024) (alleging copyright infringement when defendant used authors’ voices, characters, and stories to train AI product); Silverman v. OpenAI, Inc., No. 3:23-cv-03416 (N.D. Cal. filed July 7, 2023) (alleging defendant violated state unfair competition law by using plaintiff’s copyrighted works to train AI product). See also Pamela Samuelson, Generative AI Meets Copyright, 381 Sci. 158 (2023); Matthew Sag, Copyright Safety for Generative AI, 61 Hous. L. Rev. 295 (2023); Daryl Lim, Generative AI and Copyright: Principles, Priorities and Practicalities, 18 J. Intell. Prop. L. & Prac. 841 (2023). See also supra note 6.

^10. See, e.g., Rowena Rodrigues, Legal and Human Rights Issues of AI: Gaps, Challenges and Vulnerabilities, 4 J. Responsible Tech. 100005 (2020) (discussing the human rights challenges arising out of AI use); Simon Chesterman, Artificial Intelligence and the Limits of Legal Personality, 69 Int’l & Compar. L. Q. 819 (2020) (discussing the legal status of artificial intelligence); Sara Gerke et al., Ethical and Legal Challenges of Artificial Intelligence-Driven Healthcare, in Artificial Intelligence in Healthcare 295 (Adam Bohr & Kaveh Memarzadeh eds., 2020) (highlighting the areas within healthcare that will likely be affected by artificial intelligence and the attendant ethical and legal challenges). In addition to scholarly research, many institutions have introduced new institutes and centers focused on exploring AI and law. See, e.g., Stanford Artificial Intelligence & Law Society (SAILS), Stan. L. Sch., https://law.stanford.edu/stanford-artificial-intelligence-law-society-sails/ (last visited Sept. 12, 2024); AI, Platforms, and Society Center, Berkeley L., https://www.law.berkeley.edu/research/bclt/project-on-artificial- intelligence-platforms-and-society/ (last visited Sept. 12, 2024); AI Law & Innovation Institute, UC L. SF, https://www.uclawsf.edu/center-for-innovation/ai-law-innovation-institute/ (last visited Sept. 12, 2024); Artificial Intelligence and the Law, Berkman Klein Ctr. for Internet & Soc’y at Harvard U., https://cyber.harvard.edu/projects/artificial-intelligence-and-law (last visited Sept. 14, 2024).

^11. See, e.g., Daryl Lim, AI & IP: Innovation & Creativity in an Age of Accelerated Change, 52 Akron L. Rev. 813, 815 (2018) (arguing that although AI is already affecting the "fundamental notions underlying how and why we incentivize creation and innovation[,]" the current legal landscape can adapt to the "Fourth Industrial Revolution"); Mark A. Lemley, How Generative AI Turns Copyright Law Upside Down, 25 Sci. & Tech. L. Rev. 21 (2024) (examining the challenges and transformations in copyright law brought about by generative AI technologies); Michael Grynberg, AI and the "Death of Trademark", 108 Ky. L. J. 199 (2019) (discussing the potential implications of AI on trademark law); Simon Chesterman, Good Models Borrow, Great Models Steal: Intellectual Property Rights and Generative AI, Pol’y & Soc’y 1 (2024) (exploring the legal and policy implications of training generative AI models and owning AI output).

^12. See, e.g., Nicolas Petit, Antitrust and Artificial Intelligence: A Research Agenda, 8 J. Eur. Competition L. & Prac. 361 (2017) ("[t]he hype around ‘Artificial Intelligence’ [] has reached the antitrust community.")

^13. See, e.g., Ariel Ezrachi & Maurice E. Stucke, Artificial Intelligence & Collusion: When Computers Inhibit Competition, 2017 Univ. Ill. L. Rev. 1775 (2017) (discussing how sophisticated computer algorithms including artificial intelligence are changing the competitive landscape and considering four scenarios in which artificial intelligence can foster collusion and raise legal challenges); Emilio Calvano et al., Artificial Intelligence, Algorithmic Pricing, and Collusion, 110 Am. Econ. Rev. 3267 (2020) (studying experimentally the effects on pricing when applying artificial intelligence algorithms to oligopoly environments); Michal Gal & Daniel L. Rubinfeld, Algorithms, AI and Mergers (NYU L. & Econ. Rsch. Paper, Working Paper No. 23-36, 2023) (identifying how algorithms may affect market dynamics and demonstrating how such effects can exacerbate anticompetitive conduct); Maria José Schmidt-Kessen & Max Huffman, Antitrust Law and Coordination Through AI-Based Pricing Technologies, in Multidisiplinary Perspectives on Artificial Intelligence and the Law (2023) (providing a taxonomy of computerized pricing technologies to date and identifying legal effects of these technologies on U.S. and EU antitrust law); Daryl Lim, Antitrust’s AI Revolution, 89 Tenn. L. Rev. 679, 722 (2022) (discussing the uncertainty of antitrust rules and arguing for "employing AI as a positive forensic and predictive tool" in antitrust rulemaking and enforcement).

^14. See United States v. RealPage, Inc., No. 1:24-cv-00710 (M.D.N.C. Aug. 23, 2024); In re RealPage, Inc., No. 3:23-md-03071 (M.D. Tenn. Dec. 28, 2023); Duffy v. Yardi Sys., Inc., No. 2:23-cv-01391-RSL (W.D. Wash. Apr. 18, 2024); Karen Cornish-Adebiyi, et al. v. Caesars Ent., Inc., et al., No. 1:23-cv-02536-KMW-EAP (D.N.J. filed May 9, 2023); Gibson v. MGM Resorts Int’l, No. 2:23-cv-00140-MMD-DJA (D. Nev. Oct. 24, 2023). See also infra text accompanying notes 102-143.

^15. See, e.g., Robin Feldman, Drugs, Money, and Secret Handshakes (Cambridge Univ. Press 2019) [hereinafter Secret Handshakes]; Robin Feldman, May Your Drug Price Be Evergreen, Oxford J.L. & Biosci. 1 (2018); The Future of Pharmaceuticals: Examining the Analysis of Pharmaceutical Mergers—FTC-DOJ Workshop Summary, Fed. Trade Comm’n 5 (June 14, 2022), https://www.ftc.gov/system/files/ftc gov/pdf/Future of Pharma Workshop—Summary.pdf (summarizing the author’s presentation regarding increased consolidation in the pharmaceutical industry) [hereinafter FTC-DOJ Workshop]; Robin Feldman & Evan Frondorf, Drug Wars: How Big Pharma Raises Prices and Keeps Generics off the Market (Cambridge Univ. Press 2017); Robin C. Feldman & Mark A. Lemley, Atomistic Antitrust, 63 Wm. & Mary L. Rev. 1869 (2022).

^16. Wall-E (Pixar Animation Studios 2008) (depicting a lonely trash-picking robot left on earth who falls in love and helps reconnect humans to the earth they had deserted).

^17. Douglas Adams, The Hitchhiker’s Guide to the Galaxy (Pan Books 1979) (explaining that, after Earth’s destruction, Arthur Dent meets Marvin the Paranoid Android, who is a depressed robot with a "brain the size of a planet" but is stuck doing menial tasks).

^18. Ex-Machina (Universal Pictures 2014) (depicting a computer programmer who after winning a week-long stay at his company CEO’s home, is told by the CEO that he had built a humanoid robot named Ava. The film follows Ava’s attempts to escape into the outside world).

^19. What is artificial intelligence (AI)?, IBM, https://www.ibm.com/topics/artificial-intelligence (last visited Sept. 14, 2024) ("Artificial intelligence, or AI, is technology that enables computers and machines to simulate human intelligence and problem-solving capabilities."); Artificial Intelligence, Stan. Encyclopedia of Phil., https://plato.stanford.edu/entries/artificial-intelligence/#HistAI (last visited July 12, 2018) ("So far we have been proceeding as if we have a firm and precise grasp of the nature of AI. But what exactly is AI? Philosophers arguably know better than anyone that precisely defining a particular discipline to the satisfaction of all relevant parties (including those working in the discipline itself) can be acutely challenging.").

^20. Adrienne Mayor, An AI Wake-Up Call From Ancient Greece, Project Syndicate (Oct. 15, 2018), https://www.project-syndicate.org/commentary/artificial-intelligence-pandoras-box-by-adrienne-mayor-2018-10 (explaining that the original ancient Greek myth of Pandora described Pandora as a "cunning automaton" that was "made, not born" for the purpose of punishing mankind).

^21. Alan M. Turing, Computing Machinery and Intelligence, 49 Mind 433 (1950).

^22. Id.

^23. Id.

^24. Andrew R. Chow, How ChatGPT Managed to Grow Faster than TikTok or Instagram, TIME (Feb. 8, 2023), https://time.com/6253615/chatgpt-fastest-growing/ (describing technologists’ surprise over the popularity of ChatGPT and noting that "[w]hereas many chatbots only know how to respond to certain keywords or triggers, ChatGPT can respond to complex questions and spit out comprehensive, essay-length answers on virtually any topic.").

^25. Consider Eliza, one of the earliest examples of a computer chatbot developed at the MIT Artificial Intelligence Lab. See Joseph Weizenbaum, ELIZA—A Computer Program for the Study of Natural Language Communication Between Man and Machine, 9 Commc’ns. ACM 1 (1966).

^26. See Michael Wallace & George Dunlop, ELIZA: a Very Basic Rogerian Psychotherapist Chatbot.

^27. Weizenbaum, supra note 24, at 37 (emphasis in original).

^28. David Pierce, From Eliza to ChatGPT: why people spent 60 years building chatbots, Verge (Feb. 28, 2024, 7:00 AM), https://www.theverge.com/24054603/chatbot-chatgpt-eliza-history-ai-assistants-video ("The tech behind Eliza was incredibly primitive: users typed into a text field, and the bot selected from a bunch of predefined responses based on the keywords in your question. If it didn’t know what to say, it would just repeat your words back—you’d say ‘My father is the problem’ and it would respond ‘Your father is the problem.’ But it worked!").

^29. This was generated using the JavaScript version of ELIZA developed by Michael Wallace with further enhancements by George Dunlop.

^30. Id.

^31. See, e.g., What is Machine Learning, GeeksforGeeks, https://www.geeksforgeeks.org/ml-machine-learning/ (May 26, 2024) (describing traditional programming as "rule-based code [] written by the developers depending on the problem statements."); Sriram Parthasarathy, Difference between Traditional Programming Versus Machine Learning from a PM Perspective, Medium (Apr. 21, 2019), https://productcoalition.com/difference-between-traditional-programming-versus-machine-learning-from-a-pm-perspective-3802b02bc7f6 (describing traditional programming as "a manual process—meaning a person (programmer) creates the program. But without anyone programming the logic, one has to manually formulate or code rules."); Ali Alkan, Classical Programming vs Machine Learning: Mastering Different Tools for Problem-Solving, Medium, https://medium.com/low-code-for-advanced-data-science/classical-programming-vs-machine-learning-mastering-different-tools-for-problem-solving-84a33ede08c7 (June 26, 2024) ("Classical programming offers precise control, while machine learning allows for adaptation and learning.").

^32. This was generated using ChatGPT 3.5 on June 26, 2024. The full response read "Dealing with sadness can be challenging, but there are several strategies that can help: Acknowledge your feelings: It’s okay to feel sad. Allow yourself to recognize and accept your emotions without judgment. Talk about it: Reach out to someone you trust—a friend, family member, or therapist. Sharing your feelings can provide comfort and support. Take care of yourself: Engage in self-care activities that you enjoy, such as exercise, hobbies, or relaxation techniques like deep breathing or meditation. Focus on the present moment: Practice mindfulness to stay grounded and avoid dwelling on negative thoughts. Seek professional help if needed: If your sadness persists or affects your daily life, consider speaking with a mental health professional who can provide guidance and support. Connect with others: Social support is important during difficult times. Spending time with loved ones or participating in social activities can uplift your mood. Give yourself time: Healing takes time, so be patient with yourself as you navigate through your emotions. Remember, everyone experiences sadness at times, and it’s a normal part of life. You’re not alone, and there are people and resources available to support you through this."

^33. See Winnie Jones, Does ChatGPT Generate the Same Response? Medium (Mar. 21, 2023) ("ChatGPT uses a process called sampling to generate responses, which involves randomly selecting words from its vocabulary to create a response. This means that even if the text input is the same, ChatGPT can generate different responses each time because of the randomness involved in the sampling process.").

^34. Id.

^35. See GPT-4, OpenAI (Mar. 14, 2023), https://openai.com/index/gpt-4-research/ (explaining how GPT-4 has been built using reinforcement learning with human feedback and deep learning).

^36. Used here to describe the manual process of a programmer writing rule-based code that instruct the machine on how to perform a given task.

^37. See Sara Brown, Machine learning, explained, MIT Mgmt. Sloan Sch. (Apr. 21, 2021), https://mitsloan.mit.edu/ideas-made-to-matter/machine-learning-explained (summarizing how machine learning works and outlining the various strands of machine learning).

^38. Arthur L. Samuel, Some Studies in Machine Learning Using the Game of Checkers, 3 IBM J. Rsch & Dev. 210, 211 (1959) ("We have at our command computers with adequate data-handling ability and with sufficient computational speed to make use of machine-learning techniques, but our knowledge of the basic principles of these techniques is still rudimentary. Lacking such knowledge, it is necessary to specify methods of problem solution in minute and exact detail, a time-consuming and costly procedure. Programming computers to learn from experience should eventually eliminate the need for much of this detailed programming effort.").

^39. See Brown, supra note 6.

^40. Id. ("In some cases, machine learning can gain insight or automate decision-making in cases where humans would not be able to. . . .").

^41. See Robin Feldman, Artificial Intelligence and Cracks in the Foundation of Intellectual Property, (UC S.F. Rsch. Paper, Working Paper, 2024).

^42. 15 U.S.C. § 1 (2024).

^43. Id.

^44. Am. Needle v. NFL, 560 U.S. 183, 189-90 (2010).

^45. 15 U.S.C. § 45(a)(1).

^46. See, e.g., Lina M. Khan, Section 5 in Action: Reinvigorating the FTC Act and the Rule of Law, 11 J. Antitrust Enf’t 149 (2023) (discussing the evolution of Section 5 of the FTC Act over the years since its passing and emphasizing the importance of the section in FTC’s enforcement work); William E. Kovacic & Mark Winerman, Competition Policy and the Application of Section 5 of the Federal Trae Commission Act, 20 Minn. J. Int’l L. 274 (2010) (setting forth a framework for using Section 5 for anticompetition enforcement); James C. Cooper, The Perils of Excessive Discretion: The Elusive Meaning of Unfairness in Section 5 of the FTC Act, 3 J. Antitrust Enf’t 87 (2015) (discussing the lack of certainty with respect to the conception of unfairness under Section 5).

^47. Complaint, Caremark Rx, LLC, FTC Docket No. 9437 (Sept. 26, 2024), https://www.ftc.gov/system/files/ftc gov/pdf/d9437 caremark rx zinc health services et al part 3 complaint public redacted.pdf.

^48. N. Pac. Ry. Co. v. United States, 356 U.S. 1, 5 (1958) ("[T]here are certain agreements or practices which because of their pernicious effect on competition and lack of any redeeming virtue, are conclusively presumed to be unreasonable and therefore illegal without elaborate inquiry as to the precise harm they have caused or the business excuse for their use.").

^49. Broad. Music, Inc. v. Columbia Broad. Sys., Inc., 441 U.S. 1, 8 (1979); see also Herbert Hovenkamp, The Rule of Reason, 70 Fla. L. Rev. 81, 83 (2018) (explaining that in antitrust per se cases, "power generally need not be proven and anticompetitive effects are largely inferred from the conduct itself.").

^50. United States v. Socony-Vacuum Oil Co., 310 U.S. 150, 218 (1940) ("[N]o showing of so-called competitive abuses or evils which those agreements were designed to eliminate or alleviate may be interposed as a defense.").

^51. Standard Oil Co. v. U.S., 221 U.S. 1, 60 (1911) ("[I]t was intended that the standard of reason which had been applied at the common law and in this country in dealing with subjects of the character embraced by the statute, was intended to be the measure used for the purpose of determining whether in a given case a particular act had or had not brought about the wrong against which [the Sherman Act] provided."). For an in-depth discussion regarding the complexities of the rule of reason versus per se analysis approach, see generally Hovenkamp, supra note 47 (criticizing the court’s conceptualization of the rule of reason as being a balancing exercise and arguing that the rule of reason produces arbitrary results).

^52. Leegin Creative Leather Prods. v. PSKS, Inc., 551 U.S. 877, 899 (2007).

^53. Scott D. Hammond, Deputy Assistant Att’y Gen., Antitrust Div., U.S. Dep’t Just., Caught in the Act: Inside an International Cartel (Oct. 18, 2005), https://www.justice.gov/atr/speech/caught-act-inside-international-cartel. See also United States v. Andreas, 216 F.3d 645 (7th Cir. 2000).

^54. Andreas, 216 F.3d at 651-54.

^55. The story of the lysine cartel and its unravelling was adapted into a movie, The Informant!, starring Matt Damon as Mark Whitacre, who remarkably blew the whistle on the cartel and worked with the FBI to obtain an abundance of direct evidence of price fixing and collusion.

^56. Sharon Walsh, 3 Former Officials at ADM Get Jail Terms, Wash. Post (July 9, 1999, 8:00 PM), https://www.washingtonpost.com/archive/business/1999/07/10/3-former-officials-at-adm-get-jail-terms/68b63662-d02b-4db8-a234-fd859eb6f403/.

^57. Id.

^58. See, e.g., United States v. Snow, 462 F.3d 55, 68 (2d Cir. 2006) ("[C]onspiracy by its very nature is a secretive operation, and it is a rare case where all aspects of a conspiracy can be laid bare in court with [ ] precision . . ."); Interstate Cir., Inc. v. United States, 306 U.S. 208, 221 (1939) ("As is usual in cases of alleged unlawful agreements to restrain commerce, the Government is without the aid of direct testimony. . . .").

^59. United States v. Falstaff Brewing Corp., 410 U.S. 526, 534 n.13 (1973).

^60. See, e.g., infra text accompanying note 62.

^61. In re Text Messaging Antitrust Litig., 782 F.3d 867 (7th Cir. 2015).

^62. Id. at 870.

^63. Id.

^64. Id. at 872.

^65. Id. at 871 (emphasis added).

^66. See, e.g., Richard A. Posner, Oligopoly and the Antitrust Laws: A Suggested Approach, 21 Stan. L. Rev. 1562 (1969); Donald F. Turner, The Definition of Agreement under the Sherman Act: Parallelism and Refusals to Deal, 75 Harv. L. Rev. 655, 666 (1962) ("[T]here is fair ground for argument that oligopoly price behavior can be described as individual behavior—rational individual decision in light of relevant economic facts—as well as it can be described as ‘agreement.’ It can readily be said that each seller in this situation, in refraining from price competition, is not agreeing with his competitors but simply throwing their probable decisions into his price calculus as impersonal market facts.").

^67. Bell Atl. Corp. v. Twombly, 550 U.S. 544, 554 (2007) ("An antitrust conspiracy plaintiff with evidence showing nothing beyond parallel conduct is not entitled to a directed verdict . . . ; proof of a § 1 conspiracy must include evidence tending to exclude the possibility of independent action").

^68. Id. at 556-57.

^69. United States v. Apple, Inc., 791 F.3d 290, 314 (2d. Cir. 2015).

^70. See, e.g., Bradley C. Weber, Hub-and-Spoke Conspiracies: Can Big Data and Pricing Algorithms Form the Rim?, 26 SMU Sci. & Tech. L. Rev. 25, 27 (2023) ("The classic structure of such a conspiracy includes a firm at one level of a supply chain—such as a buyer or supplier—who acts like the ‘hub’ of a wheel. Vertical agreements up or down the supply chain act as the "spokes" and, most importantly, a horizontal agreement among the spokes acts as the "rim" of the wheel. The distinguishing feature of a hub-and-spoke conspiracy is the participation of the vertically-aligned hub in the middle of the horizontal agreement.").

^71. Id.

^72. Apple, 791 F.3d at 314 n.15.

^73. See supra note 49 and accompanying text.

^74. See, e.g., Roundtable on Hub-and-Spoke Arrangements—Background Note by the Secretariat, OECD, Directorate Fin. & Enter. Affs. Competition Comm. 5 (Nov. 25, 2019), https://one.oecd.org/document/DAF/COMP(2019)14/en/pdf ("E-commerce phenomena such as pricing algorithms, price monitoring software or online platforms can be instrumental in supporting hub-and-spoke arrangements, or can lead to similar market outcomes.") [hereinafter OECD Background Note].

^75. Interstate Cir., Inc. v. United States, 306 U.S. 208, 226-27 (1939).

^76. Id. at 214.

^77. Id. at 215.

^78. Id.

^79. Id. at 215-17.

^80. Id.

^81. Id. at 221-22.

^82. Id. at 222 (emphasis added).

^83. See supra Part I.B. See also OECD Background Note, supra note 72, at 5 ("The topic [of hub-and-spoke conspiracies] raises interesting questions, as it essentially requires enforcement agencies to clearly define the point where a day-to-day, legitimate business occurrence, namely the exchange of often confidential information between suppliers and distributors, turns into an illegal horizontal agreement or concerted action that can be subject to harsh sanctions, and in some regimes criminal prosecution.").

^84. OECD, Hub-and-spoke arrangements—Note by the United States 2 (Nov. 28, 2019), https://www.ftc.gov/system/files/attachments/us-submissions-oecd-2010-present-other-international-competition-fora/oecd-hub and spoke arrangements us.pdf.

^85. See, e.g., Interstate Cir., Inc. v. United States, 306 U.S. 208, 221—22 (1939) (knowledge of arrangements between other competitors and abrupt changes to business practices); Apple, 791 F.3d at 316 (knowledge of arrangements between other competitors); Toys "R" Us v. FTC, 221 F.3d 928, 935 (2000) (abrupt shift from past business practices and competitors depriving themselves of a profitable sales outlet).

^86. See, e.g., 7-UP Bottling Co. v. Archer Daniels Mdland Co. (In re Citric Acid Litig.), 191 F.3d 1090, 1097 (9th Cir. 1999) ("[T]he crucial question is whether all the evidence considered as a whole can reasonably support the inference that [the defendant] conspired with the admitted conspirators to fix prices.").

^87. Interstate Cir., 306 U.S. at 240.

^88. See Algorithms and collusion, OECD, https://www.oecd.org/competition/algorithms-and-collusion.htm (last visited Sept, 20, 2024); Algorithms and collusion: Competition policy in the digital age, OECD (Sept. 14, 2017), https://www.oecd.org/competition/algorithms-collusion-competition-policy-in-the-digital-age.htm.

^89. See Press Release, Sen. Amy Klobuchar, Klobuchar, Colleagues Introduce Antitrust Legislation to Prevent Algorithmic Price Fixing (Feb. 2, 2024), https://www.klobuchar.senate.gov/public/index.cfm/2024/2/klobuchar-colleagues-introduce-antitrust-legislation-to-prevent-algorithmic-price-fixing.

^90. See, e.g., Hannah Garden-Monheit & Ken Berber, Price fixing by algorithm is still price fixing, Fed. Trade Comm’n Bus. Blog (Mar. 1, 2024), https://www.ftc.gov/business-guidance/blog/2024/03/price-fixing-algorithm-still-price-fixing; Andrew Smith, Using Artificial Intelligence and Algorithms, Fed. Trade Comm’n Bus. Blog (Apr. 8, 2020), https://www.ftc.gov/business-guidance/blog/2020/04/using-artificial-intelligence-algorithms; see also Press Release, infra note 130; FTC-DOJ Workshop, supra note 14.

^91. See text accompanying infra notes 131-138.

^92. See infra Section I I.C.

^93. See, e.g., William V. Padula, State and Federal Policy Solutions to Rising Prescriptions Drug Prices in the U.S., 22 J. Health Care L. & Pol’y 15 (2019); Arielle Bosworth et al., Off. of Health Pol’y, Dep’t of Health and Hum. Serv., HP-2023-25, Changes in the List Prices of Prescription Drugs, 2017-2023 (2023); Sydney Lupkin, Senators ask CEOs why their drugs cost so much more in the U.S., NPR (Feb. 8, 2024, 5:35 PM), https://www.npr.org/sections/health-shots/2024/02/08/1230174586/high-us-drug-prices.

^94. See generally Rebecca Robbins & Reed Abelson, The Opaque Industry Secretly Inflating Prices for Prescription Drugs, N.Y. Times (June 21, 2024), https://www.nytimes.com/2024/06/21/business/prescription-drug-costs-pbm.html?smid=url-share; Secret Handshakes, supra note 14, at 11-17 (Cambridge Univ. Press 2019).

^95. See supra note 11.

^96. See, e.g., Zach Y. Brown & Alexander MacKay, Competition in Pricing Algorithms, 15 Am. Econ. J.: Microeconomics 109, 115-16 (2023) (explaining how a pricing algorithm determines prices using a formula.)

^97. See, e.g., AI Revenue Management: Hands-On Experience Building a System for Airline Dynamic Pricing, AltexSoft (Apr. 4, 2024), https://www.altexsoft.com/blog/ai-revenue-management-dynamic-pricing/ (describing the early approach by airlines to forecast demand for flights and adjust prices based "mainly on the gut feeling of industry experts" which was used to build pricing models).

^98. See generally Qiaochu Wang et al., Algorithms, Artificial Intelligence and Simple Rule Based Pricing (Apr. 24, 2023) (unpublished manuscript), https://papers.ssrn.com/sol3/papers.cfm?abstract id=4144905 (describing different types of pricing algorithms and studying their respective effectiveness); Marco Bertini & Oded Koenigsberg, The Pitfalls of Pricing Algorithms, Harv. Bus. Rev. (Sept.-Oct. 2021), https://hbr.org/2021/09/the-pitfalls-of-pricing-algorithms (describing AI pricing algorithms as those algorithms that are "intended to help firms determine optimal prices on a near realtime basis.").

^99. AI algorithm is not a term of art, but we use it throughout this paper to distinguish it from its more primitive ancestor, the traditionally programmed algorithm.

^100. Emilio Calvano et al., Artificial Intelligence, Algorithmic Pricing and Collusion, 110 Am. Econ. Rev. 3267, 3268 (2020) (studying experimentally the effects on pricing when applying artificial intelligence algorithms to oligopoly environments).

^101. See, e.g., id.

^102. Id.

^103. Id.

^104. See, e.g., Calvano et al., supra note 12; Bruno Salcedo, Pricing Algorithms and Tacit Collusion (Nov. 1, 2015), https://brunosalcedo.com/docs/collusion.pdf (unpublished manuscript); John Asker et al., Artificial Intelligence and Pricing: The Impact of Algorithm Design (Nat’l Bureau of Econ. Rsch., Working Paper No. 28535, 2021).

^105. See, e.g., Brown & Mackay, supra note 92; infra Section II.B.

^106. Meyer v. Kalanick, 174 F. Supp. 3d 817 (S.D.N.Y. 2016).

^107. Id. at 820.

^108. Id. at 822-23.

^109. Id. at 821.

^110. Id.

^111. Recall that this market concentration at the hub level—as opposed to the spoke level—also existed in Interstate Circuit. See id.

^112. Kalanick argued that as a rider, Meyer was bound by the compulsory arbitration clause in the terms of service. In 2016, the District Court held that Meyer was not bound by the terms of service, but that order was subsequently vacated by the Second Circuit and remanded back to the District Court in 2017. In 2018, the case was sent to arbitration. The arbitrator found for Kalanick and Uber.

^113. Meyer v. Kalanick, 174 F. Supp. 3d 817, 823 (S.D.N.Y. 2016).

^114. Id.

^115. Id.

^116. Id. at 825-26 (emphasis added).

^117. In re RealPage, Inc., No. 3:23-md-3071 (M.D. Tenn. Dec. 28, 2023).

^118. Id.

^119. Id.

^120. Id.

^121. Id.

^122. Id. at 23, 24.

^123. Id. at 30.

^124. Id.

^125. Id.

^126. See, e.g., Gibson v. MGM Resorts Int’l, No. 2:23-cv-00140-MMD-DJA (D. Nev. Oct. 24, 2023) (alleging unlawful restraint of trade by multiple hotels agreeing to use a common third-party pricing software).

^127. Class Action Complaint, Duffy v. Yardi, No. 2:23-cv-1391 (W.D. Wash. Sept. 8, 2023), ECF No. 1.

^128. Id. ¶¶ 4-5.

^129. Id. ¶¶ 4-5, 18.

^130. Class Action Complaint, Cornish-Adebiyi v. Caesars, No. 1:23-cv-2536 (D.N.J. May 9, 2023), ECF No. 1.

^131. Id. ¶¶ 5-6, 12-13.

^132. Alden Abbott, Why Antitrust Regulators Are Focused On Problematic AI Algorithms, Forbes (Mar. 13, 2024, 12:14 PM), https://www.forbes.com/sites/aldenabbott/2024/03/13/why-antitrust-regulators-are-focused-on-problematic-ai-algorithms/: Press Release, Fed. Trade Comm’n, FTC Restores Rigorous Enforcement of Law Banning Unfair Methods of Competition (Nov. 10, 2022) (on FTC website, https://www.ftc.gov/news-events/news/press-releases/2022/11/ftc-restores-rigorous-enforcement-law-banning-unfair-methods-competition): Statement of Interest of the United States, Cornish-Adebiyi, 1:23-cv-2536 (D.N.J. Mar. 28, 2024), ECF No. 96.

^133. See Statement of Interest of the United States, In re: RealPage, No. 3:23-md-3071 (M.D. Tenn. Nov. 15, 2023), ECF Nos. 627, 628; Statement of Interest of the United States, Duffy, No. 2:23-cv-1391 (W.D. Wash. Mar. 1, 2024), ECF No. 149: Statement of Interest of the United States, Cornish-Adebiyi, 1:23-cv-2536 (D.N.J. Mar. 28, 2024), ECF No. 96.

^134. United States v. RealPage Inc., No. 1:24-cv-00710 (M.D.N.C., Aug. 23, 2024).

^135. See supra text accompanying notes 113-121.

^136. Complaint at ¶¶ 225, 237-40, No. 1:24-cv-00710 (M.D.N.C., Aug. 23, 2024) (emphasis added).

^137. Id at ¶¶ 226-27, 236.

^138. Id. at ¶ 76.

^139. Id. at ¶ 73.

^140. See infra Section II.C.

^141. Zach Brown & Alexander MacKay, Are Online Prices Higher Because of Pricing Algorithms, Brookings, https://www.brookings.edu/articles/are-online-prices-higher-because-of-pricing-algorithms/ (July 7, 2022) (discussing the use of pricing algorithms by online retailers); Holli Sargeant, Algorithmic Decision-Making in Financial Services: Economic and Normative Outcomes in Consumer Credit, 3 AI & Ethics 1295 (2023) (discussing the use of algorithms in consumer credit markets); Bernard Marr, The True Value Of Data and AI in Human Resources, Forbes (Dec. 5, 2023, 2:55 AM), https://www.forbes.com/sites/bernardmarr/2023/12/05/the-true-value-of-data-and-ai-in-human-resources/ (discussing the use of algorithms in Human Resources departments); Jeremy Bowman, How AI is Being Used in the Travel Industry, Motley Fool (Aug. 5, 2024, 11:52 AM), https://www.fool.com/investing/stock-market/market-sectors/information-technology/ai-stocks/ai-in-travel/ (discussing the use of algorithms in the travel industry); Neil Sahota, Streaming into the Future: How AI Is Reshaping Entertainment, Forbes (Mar. 18, 2024, 10:00 AM), https://www.forbes.com/sites/neilsahota/2024/03/18/streaming-into-the-future-how-ai-is-reshaping-entertainment/# (d iscussing the use of algorithms in the streaming industry).

^142. Live Well Chiropractic PLLC v. Multiplan, Inc. et al, No. 1:24-cv-03680 (N.D. III. filed May 6, 2024).

^143. It is difficult to describe in one sentence what Multiplan does. On its website under "About Multiplan" the company simply states that "MultiPlan’s team of experts helps healthcare payors manage the cost of care, improve their competitiveness and inspire positive change." MultiPlan, https://www.multiplan.us/ (last updated July 19, 2024).

^144. Class Action Complaint at ¶ 2, Live Well Chiropractic PLLC.

^145. Id.

^146. Id. at ¶¶ 6, 9.

^147. Id. at ¶ 11

^148. See infra Section II.B.

^149. Ezrachi & Stucke, supra note 12; Ariel Ezrachi & Maurice E. Stucke, Sustainable and Unchallenged Algorithmic Tacit Collusion, 17 Nw. J. Tech. & Intell. Prop. 217 (2020); Ariel Ezrachi & Maurice E. Stucke, The Role of Secondary Algorithmic Tacit Collusion in Achieving Market Alignment, 26 Vand. J. Ent. & Tech. L. 461 (2024); Ariel Ezrachi & Maurice E. Stucke, Virtual Competition: The Promise and Perils of the Algorithm- Driven Economy (Harv. Univ. Press 2016).

^150. Ezrachi & Stucke, supra note 12, at 1782-84.

^151. Id. at 1782.

^152. Id.

^153. Id. at 1785.

^154. Id. at 1782.

^155. Id.

^156. Id.

^157. Id. at 1783.

^158. Id.

^159. Id.

^160. Id. at 1783-84.

^161. Ezrachi & Stucke, The Role of Secondary Algorithmic Tacit Collusion in Achieving Market Alignment, supra note 148, at 472.

^162. See supra note 14.

^163. See, e.g., Secret Handshakes, supra note 14, at 16 (a chart illustrating the complicated flow of money and drug products, and the numerous players involved in the system); Inmaculada Hernandez & Anna Hung, A Primer on Brand-Name Prescription Drug Reimbursement in the United States, 30 J. Managed Care & Specialty Pharm. 99, 99 (2024) (setting out the various players and their roles in the "complex enterprise" of prescription drug distribution).

^164. For a succinct but comprehensive walkthrough of the pharmaceutical drug pipeline, see Hernandez & Hung, supra note 163.

^165. See id.; Secret Handshakes, supra note 14.

^166. See Hernandez & Hung, supra note 163; Secret Handshakes, supra note 14.

^167. See Secret Handshakes, supra note 14, at 19-20 nn.4-8. It is worth noting that PBMs’ contracts with their payor clients are based on the rebate levels that the PBM thinks it can achieve. The actual rebate negotiated between PBM and drug company is never divulged to the payor. See Secret Handshakes, supra note 14, at 19.

^168. Pharmacy Group Purchasing Organizations (GPOs) are a relatively new invention in the PBM industry. It is unclear what benefits GPOs provide, if any, and at least some commentators consider GPOs to be yet another cloaking vehicle for PBMs to hide behind. Eric Percher, Trends in Profitability and Compensation of PBMs and PBM Contracting Entities, Nephron Rsch. (Sept. 18, 2023), https://nephronresearch.com/trends-in-profitability-and-compensation-of-pbms-and-pbm-contracting-entities/.

^169. For a template formulary, see Model Part D Comprehensive Formulary (Aug. 8, 2005), https://www.hhs.gov/guidance/sites/default/files/hhs-guidance-documents/comprehensive%20formularymodel_pdp_12_0.pdf. For a real example, see Blue Shield Cal., Blue Shield Rx Plus (PDP) 2024 Formulary (List of Covered Drugs) (Aug. 20, 2024), https://blueshieldca.adaptiverx.com/web/pdf?key=8F02B26A288102C27BAC82D14C006C6FC54D480F80409B68BEC6880E527BF1C2.

^170. See Cole Werble, Formularies, Health Affs. (Sept. 14, 2017), https://www.healthaffairs.org/content/briefs/formularies.

^171. T. Joseph Mattingly 2nd et al., Pharmacy Benefit Managers: History, Business Practices, Economics, and Policy, 4 Jama Health F. 1, 3 (2023); Secret Handshakes, supra note 14, at 13.

^172. A favorable position on the formulary is usually one where a branded product is placed on a lower cost-sharing tier or where it is subject to fewer utilization controls than a generic alternative. See Mariana P. Socal et al., Favorable Formulary Placement of Branded Drugs in Medicare Prescription Drug Plans When Generics Are Available, 179 JAMA Internal Med. 832 (2019).

^173. See Secret Handshakes, supra note 14, at 20 nn.10-13.

^174. See Formulary Management, AMCP (last visited Sept. 26, 2024), https://www.amcp.org/about/managed-care-pharmacy-101/concepts-managed-care-pharmacy/formulary-management#:~:text=The%20medications%20and%20related%20products,in%20the%20health%20care%20field. See also Robert B. Goldberg, Managing the Pharmacy Benefit: The Formulary System, 26 J. Managed Care & Specialty Pharmacy 341 (2020).

^175. Drug formulary FAQs, Blue Shield Cal. (last updated Oct. 1, 2023), https://www.blueshieldca.com/en/home/be-well/pharmacy/formulary-faqs (under "How are Blue Shield drug formularies developed?": "The Blue Shield Pharmacy and Therapeutics (P&T) Committee develops Blue Shield drug formularies. The P&T Committee reviews medical literature concerning safety, effectiveness, and current use in therapy to determine which drugs should be included on our formularies. The medical information reviewed is from a variety of nationally recognized sources such as Medline, other databases, pharmaceutical manufacturers, medical professional associations, and peer-reviewed journals."); Medicare Drug List Formulary, Cigna Healthcare https://www.cigna.com/medicare/member-resources/drug-list-formulary (last updated Mar. 8, 2024) ("Cigna Healthcare doctors and pharmacists pick these drugs for their effectiveness, safety, ease of use, and cost.").

^176. See Martha M. Rumore & F. Randy Vogenberg, PBM P&T Practices, 42 Health Care & L. 330, 330 (2017) ("PBMs serve as third-party administrators of prescription programs for commercial health plans, self-insured employer plans, Medicare Part D plans, the Federal Employees Health Benefits Program, and state government employee plans. They play a critical role in the prescription drug supply chain by performing a number of P&T-related activities, such as developing, maintaining, and enforcing the formulary. . . ."); Goldberg, supra note 174, at 342 ("Most PBMs have their own P&T committees that develop and approve a master formulary for the PBM. The P&T committee of the MCO can then take this formulary, along with the clinical and financial research provided by the PBM, and develop its own subset or custom formulary.").

^177. Gordon D. Schiff et al., A Prescription for Improving Drug Formulary Decision Making, 9 PLoS Med. 1, 1 (2012) ("formularies can unquestionably exert a powerful influence on prescribing decisions and medication utilization."); Rumore & Vogenberg, supra note 176, at 330 ("Because of the large percentage of Americans covered by PBMs, the commercial success of a drug in the U.S. depends largely on its inclusion on as many formularies as possible.").

^178. See Secret Handshakes, supra note 14, at 9 ("List prices, of course, are only the beginning of the story. Drug companies enter into a variety of contracts that provide for rebates from the list price. Although these price concessions are a closely guarded secret, and it is difficult to tease out the actual net price that different entities pay along the drug chain, the net price paid to the drug company is substantially less than the list price." (footnote omitted)).

^179. Congress is currently considering a number of bills that seek to address PBM-related issues, including the lack of disclosure. E.g., Pharmacy Benefit Manager Transparency Act of 2023, S. 127, 118th Cong. (2023); Pharmacy Benefit Manager Reform Act, S. 1339, 118th Cong. (2023); Protecting Patients Against PBM Abuses Act, H.R. 2880, 118th Cong. (2023). One bill has already been approved by the House of Representatives. Lower Costs, More Transparency Act, H.R. 5378, 118th Cong. (2023).

^180. U.S. Fed. Trade Comm’n, Off. Pol’y Planning, Pharmacy Benefit Managers: The Powerful Middlemen Inflating Drug Costs and Squeezing Main Street Pharmacies 1 (2024) ("[T]he three largest PBMs now manage nearly 80 percent of all prescriptions filled in the United States."); Adam J. Fein, The Top Pharmacy Benefit Managers of 2023: Market Share and Trends for the Biggest Companies—And What’s Ahead, Drug Channels (Apr. 9, 2024), https://www.drugchannels.net/2024/04/the-top-pharmacy-benefit-managers-of.html ("For 2023, nearly 80% of all equivalent prescription claims were processed by three companies: the Caremark business of CVS Health, the Express Scripts business of Cigna, and the Optum Rx business of UnitedHealth Group.").

^181. The Role of Pharmacy Benefit Managers in Prescription Drug Markets Part II: Not What the Doctor Ordered: Hearing Before the H. Comm. on Oversight & Accountability (2023) (statement of Hugh Chancy, RPh), https://oversight.house.gov/wp-content/uploads/2023/09/House-Oversight-testimony-Chancy.pdf; Adam J. Fein, Mapping the Vertical Integration of Insurers, PBMs, Specialty Pharmacies, and Providers: A May 2024 Update, Drug Channels (May 7, 2024), https://www.drugchannels.net/2024/05/mapping-vertical-integration-of.html.

^182. For an alarming illustration of this problem, see id.

^183. See, however, text accompanying supra note 46.

^184. See, e.g., Scott McEachern & Patrick Cambel, PBM Contracts: Understand then Optimize, Milliman White Paper (2020) (discussing the complexities of contracts with PBMs including the potential financial effects of small nuances in technical definitions); What to Know About Contracting with Pharmacy Benefit Manager, ProspHire (May 19, 2022), https://www.prosphire.com/blog/how-can-a-pharmacy-benefit-manager-help-protect-you/ ("Often these [PBM] contracts contain lengthy sections filled with nuances that dictate the type of arrangement the Plan and PBM are entering. Definitions are one key section that requires complete understanding and agreement."); Letter from PhRMA to DOJ Antitrust Div., HHS, & FTC, Request for Information on Consolidation in Health Care Markets (Docket No. ATR 102) (June 5, 2024), https://phrma.org/-/media/Project/PhRMA/PhRMA-Org/PhRMA-Refresh/Comment-Letters/PhRMA FTC-DOJ-HHS-RFI-on-Consolidation Final.pdf ("Lack of transparency and the complexity of PBM arrangements can make it difficult for plan sponsors to assess PBM performance on their behalf.") [hereinafter PhRMA Letter].

^185. See generally, Secret Handshakes supra note 14.

^186. PBMs with full pass-through pricing models do exist. Take, for example, Costco Health Solutions, a PBM with a pass- th rough model that makes money solely through an administrative fee. See Why Costco Health Solutions?, Costco, https://www.costcohealthsolutions.com/pages/TheCostcoDifference.aspx#:~:text=Why Costco Health Solutions%3F,any other sources of revenue (last visited Sept. 27, 2024).

^187. See Secret Handshakes supra note 14, at 33.

^188. See H. Comm. on Oversight & Accountability, The Role of Pharmacy Benefit Managers in Prescription Drug Markets 49 (2024) (reporting that in Texas, where state laws require PBMs to file annual reports, PBMs received $2.2 billion in manufacturer rebates, of which $91 million were retained by PBMs, and $2.07 billion were passed on to the payors, with only $15 million pass on to patients.)

^189. See id. at 33-35.

^190. See text accompanying supra notes 178-179.

^191. See id. See also PhRMA Letter supra note 183 at 5-6.

^192. José R. Guardado, Policy Research Perspectives: Competition in PBM Markets and Vertical Integration of Insurers with PBMs: 2024 Update 10.

^193. See supra note 184.

^194. See PhRMA Letter supra note 183 at 5.

^195. See, e.g., Health Insurance Rights & Protections, HealthCare.Gov, https://www.healthcare.gov/health-care-law-protections/rate-review/ (last visited Sept. 27, 2024); CMS.gov, Medical Loss Ratio Data and System Resources, https://www.cms.gov/marketplace/resources/data/medical-loss-ratio-data-systems-resources (last visited Aug. 13, 2024); Secret Handshakes supra note 14, at 35; John Weeks, The 80/20 Rule: Why Medical Insurers Are Not Interested in Cost Cutting (or Integrative Health)… Plus More, 15 Integrative Med. 18 (2016) (describing insurers as "cost-plus operators" and arguing that "[t]he more that medical services cost, the greater the dollars in their cut, long term. Insurers are not now, and will not ever be, major buddies of implementing designs for cost savings based on integrative health approaches."); Juliette Cubanski & Tricia Neuman, Changes to Medicare Part D in 2024 and 2025 Under the Inflation Reduction Act and How Enrollees will Benefit, KFF (Apr. 20, 2023), https://www.kff.org/medicare/issue-brief/changes-to-medicare-part-d-in-2024-and-2025-under-the-inflation-reduction-act-and-how-enrollees-will-benefit/; Secret Handshakes supra note 14, at 39. For a helpful hypothetical illustrating the Affordable Care Act’s profit limitation—known as the 80/20 rule—using a child’s request for a bowl of ice cream, see Secret Handshakes supra note 14, at 36.

^196. For a more in-depth discussion of the role insurance companies play in the pharmaceutical drug pipeline, see Secret Handshakes, supra note 14, at 32-43.

^197. See supra Section I.A.

^198. See supra notes 132-140. See also supra Section II.B.

^199. OECD, Algorithms and Collusion—Note by the United States 6 (May 26, 2017), https://one.oecd.org/document/DAF/COMP/WD(2017)41/en/pdf.

^200. Id. ("Absent concerted action, independent adoption of the same or similar pricing algorithms is unlikely to lead to antitrust liability even if it makes interdependent pricing more likely. For example, if multiple competing firms unknowingly purchase the same software to set prices, and that software uses identical algorithms, this may effectively align the pricing strategies of all the market participants, even though they have reached no agreement. Even when firms do not all adopt identical algorithms, the use of algorithms may increase price transparency and help to stabilize prices. However, enforcement agencies normally police the risk for interdependence through merger control (due, in part, to the difficulties in crafting an adequate remedy to interdependence) while prosecuting collusion directly. This distinction remains appropriate when evaluating the use of algorithms.").

^201. Id.

^202. See supra note 177 and accompanying text.

^203. Competition in Prescription Drug Markets, 2017-2022, Off. Assistant Sec’y for Planning & Evaluation (ASPE) (2023), https://aspe.hhs.gov/sites/default/files/documents/1aa9c46b849246ea53f2d69825a32ac8/competition-prescription-drug-markets.pdf [hereinafter ASPE Report].

^204. See, e.g., David Blumenthal, It’s the Monopolies, Stupid!, Commonwealth Fund (May 24, 2018), https://www.commonwealthfund.org/blog/2018/its-monopolies-stupid; Robert Pearl, Pharma Companies: A Conglomerate Of Monopolies, Forbes (Jan. 31, 2023, 4:30 AM), https://www.forbes.com/sites/robertpearl/2023/01/31/pharma-companies-a-conglomerate-of-monopolies/?sh=133819be1ce1. A recent study by the APSE, which examined prescription drug expenditures between 2017 and 2022, found that about 800 prescription small molecule drugs (43% of drugs in 2022) and 166 biologics (79% of biologic products) had only one manufacturer. ASPE Report. There is much literature on the various games that pharmaceutical companies play to unreasonably sustain their monopoly. See generally, Feldman, May Your Drug Price Be Evergreen, supra note 14; Secret Handshakes, supra note 14; Feldman & Frondorf, Drug Wars: How Big Pharma Raises Prices and Keeps Generics off the Market, supra note 14; Robin Feldman, The Price Tag of "Pay-for-Delay", 23 Colum. Sci. & Tech. L. Rev. 1 (2021); Patricia M. Danzon, Competition and Antitrust Issues in the Pharmaceutical Industry 26 (2014), https://faculty.wharton.upenn.edu/wp-content/uploads/2017/06/Competition-and-Antitrust-Issues-in-the-Pharmaceutical-IndustryFinal7.2.14.pdf. Equally, many of these games are becoming increasingly scrutinized by various legislative, regulatory, and rule-making bodies. See ASPE Report, supra note 202.

^205. See, e.g., Regulators Take Aim at Pharma Price Fixing, Arnold Ventures (Oct. 29, 2020), https://www.arnoldventures.org/stories/regulators-take-aim-at-pharma-price-fixing (discussing efforts by states and other policymakers to investigate alleged pharmaceutical price fixing and litigate the same); Robbins & Abelson, supra note 92. See also the FTC’s administrative complaint against the three biggest PBMs supra note 46.

^206. See, e.g., Amy Sayers & Chris Stebbins, PBM companies turn to AI to improve operations and deliver better customer experiences, EXL, https://www.exlservice.com/insights/white-paper/pbm-companies-turn-ai-improve-operations-and-deliver-better-customer-experiences (last visited Sept. 27, 2024). For examples of pharmacy benefit consultants; see, e.g., Who we serve, prescryptive, https://prescryptive.com/ (last visited Sept. 27, 2024); Data-Driven Decisions for Smarter Pharmacy Benefit Management, xevant, https://www.xevant.com/ (last visited Sept. 27, 2024).

^207. See, e.g., Class Action Complaint, Osterhaus Pharmacy Inc. et al. v. Express Scripts Inc., No. 2:24-cv-00039 (W.D. Wash. Jan. 9, 2024) (alleging that Express Scripts Inc. conspired with three competitors to fix the price of pharmacy reimbursement rates and fees); Complaint for Disgorgement, Injunctive Relief and Declaratory Judgment, State of Ohio v. Ascent Health Servs. LLC, No. 2:23-CV H 03 0179 (Ct. Com. Pl. Del. Cnty. Mar. 27, 2023). Several investigative pieces by journalist Bob Herman at Axios reveal details of problematic anticompetitive behaviors in the pharmaceutical industry. See, e.g., Bob Herman, Documents reveal the secrecy of America’s drug pricing matrix, Axios (Dec. 6, 2021), https://www.axios.com/2021/12/06/aon-express-scripts-contract-employers-drug-price-data (revealing that Aon, a benefits consulting firm, works exclusively with the three major PBMs, forming a "drug pricing coalition" and guarding data on drug costs from employers sponsoring the health benefit such that they cannot independently assess whether the prices are reasonable); Bob Herman, How consultants and pharmacy middlemen work the drug pricing system, Axios (Sept. 27, 2018), https://www.axios.com/2018/09/27/big-consultants-pbms-drug-prices-employers (describing the role of PBMs and consultants and how the big companies keep their market share by charging lower administrative fees and promising large rebates).

^208. International Prescription Drug Price Comparisons: Estimates Using 2022 Data, ASPE (2024), https://aspe.hhs.gov/sites/default/files/documents/277371265a705c356c968977e87446ae/international-price-comparisons.pdf.

^209. See, e.g., Aaron S. Kesselheim et al., The High Cost of Prescription Drugs in the United States: Origins and Prospects for Reform, 316 JAMA 858 (2016); Hagop Kantarjian et al., High Cancer Drug Prices in the United States: Reasons and Proposed Solutions, 10 J. Oncology Prac. e208 (2014); Changes in the List Price of Prescription Drugs, 2017-2023, ASPE (Oct. 6, 2023), https://aspe.hhs.gov/reports/changes-list-prices-prescription-drugs#:~:text=Over%20the%20period%20from%20January,to%20%24590%20per%20drug%20product; Our position on medicine pricing, novo nordisk, https://www.novonordisk.com/sustainable-business/access-and-affordability/pricing-position.html (last visited Sept. 27, 2024) ("Many Americans struggle to pay for our medicines and we are focused on working collaboratively toward sustainable solutions."); 2022 U.S. Pricing Transparency Report, Janssen Pharms., Inc. (2024) (accepting that "the reality for millions of patients is growing affordability and health equity gaps" though rejecting responsibility for the problem and arguing that the cause is "underinsurance and inadequate insurance benefit design driven by middlemen, including pharmacy benefit managers.").

^210. See generally Anti-Money Laundering (AML)-How it works and why it matters, fraud.com, https://www.fraud.com/post/anti-money-laundering-aml#AML KYC in UK Europe and rest of the world (last visited Sept. 27, 2024). In the United States, the anti-money laundering regulatory scheme started with the Bank Secrecy Act of 1970, which requires banks to report and identify certain transactions and clients and to maintain paper trails of financial transactions. Other federal laws have since been enacted that further strengthen the scheme. See History of Anti-Money Laundering Laws, FinCen, https://www.fincen.gov/history-anti-money-laundering-laws (last visited Sept. 27, 2024). In the U.K., the regulatory scheme is similarly made up of various pieces of legislation, including the Proceeds of Crime Act 2002, the Serious Organised Crime and Police Act 2005, and other laws. See UK Law And Guidance, Inst. Fin. Accts., https://www.ifa.org.uk/technical-resources/aml/uk-law-and-guidance (last visited Sept. 27, 2024). In the EU, a similar combination of laws is used. See Anti-money laundering and countering the financing of terrorism at EU level, Eur. Comm’n, https://finance.ec.europa.eu/financial-crime/anti-money-laundering-and-countering-financing-terrorism-eu-level en (last visited Sept. 27, 2024).

^211. See generally What we do, FATF, https://www.fatf-gafi.org/en/the-fatf/what-we-do.html (last visited Sept. 27, 2024).

^212. Id.

^213. See FATF, International Standards on Combating Money Laundering and the Financing of Terrorism & Proliferation: The FATF Recommendations 16-18 (2023), https://www.fatf-gafi.org/content/dam/fatf-gafi/recommendations/FATFRecommendations 2012.pdf.coredownload.inline.pdf [hereinafter The FATF Recommendations].

^214. Though broadly uniform, many countries’ risk assessment checks vary depending on the size of the business carrying out the check such that the obligations do not become unduly burdensome for smaller firms. See id. at 34-35 ("The nature and extent of any assessment of money laundering and terrorist financing risks should be appropriate to the nature and size of the business.").

^215. See, e.g., Jacqueline Arena & Andrew Foster, Antitrust Compliance Tools for In House Lawyers, ACC Docket, https://www.skadden.com/-/media/files/publications/2018/03/antitrust_compliance_tools_for_inhouse_lawyers.pdf (last visited June 10, 2024); Latham & Watkins, Competition Compliance Risk Checklist, https://www.lw.com/ admin/Upload/Documents/Competition-Law-Compliance-Checklist.pdf (last visited June 10, 2024); Baker McKenzie, Antitrust & Competition Hub: Distribution Antitrust Risk Tool (DART), https://www.bakermckenzie.com/en/insight/topics/antitrust-and-competition (last visited June 10, 2024).

^216. See, e.g., Compete Fairly, Walmart Ethics, https://www.walmartethics.com/content/walmartethics/en_us/code-of-conduct/build-trust-in-our-business/compete-fairly.html (last visited June 10, 2024).

^217. See The Role of Pharmacy Benefit Managers in Prescription Drug Markets, Part II: Not What the Doctor Ordered, Hearing Before the United States House of Representatives Committee on Oversight and Accountability, 118th Cong. 4, 8 (Sept. 19, 2023) (written testimony of Juan Carlos "JC" Scott, President and CEO of Pharmaceutical Care Management Association) ("In almost every industry and especially healthcare, the most effective way to lower costs is through increased competition. Mandating public disclosure of confidential information will only invite drug companies to collude and raise drug costs.").