Is Artificial Intelligence a New Gateway to Anticompetitive Collusion?

Also by CDT Intern Hannah Babinski

Roughly 85 percent of adults in the United States interact with the Internet on a daily basis.[1] Commerce over the Internet has in many ways made the lives of Americans easier, more convenient, and streamlined. But has it also opened the door for companies to utilize new and innovative technology to take advantage of their customers, suppliers, and workers by engaging in collusive price fixing? And if so, what can be done about it?

Antitrust scholars have been raising this question for several years,[2] but the new innovations in artificial intelligence are bringing renewed attention to it.

Under the U.S. antitrust laws, unlawful collusion – specifically, price fixing, the form of collusion we focus on here – encompasses any agreement among competing companies to set prices at inflated levels.[3] The law condemns collusion because it subverts the free market and denies consumers the benefits of prices determined by competition, where companies honestly compete against each other to win customers by offering more attractive products and services at more affordable prices.[4] The antitrust laws have traditionally drawn a distinction, for a mix of policy and practicality reasons, between price-fixing agreements and what is referred to as “conscious parallelism.” The distinction lies in that the latter can actually constitute honest competition, with companies separately and independently monitoring each other’s prices in order to look for opportunities to gain an advantage over their competitors and attract new customers.

The former, in contrast, is an agreement to avoid honest competition. Antitrust enforcers, and courts, have recognized that conscious parallelism is not without its problems. Companies can monitor each other’s prices in order to see how high they can inflate their own prices, and this can result in prices that are higher than if vigorous competition were taking place. Yet, enforcers and courts have concluded that it is impossible, as a practical matter, to identify and stop conscious parallelism that inflates prices without risking interfering with honest competitive responses to normal price and value fluctuations of goods and services.[5] So they look for indicators that the price monitoring and adjustments are not independent but rather are mutual, intended coordination.

This grey area between independent price monitoring in the interest of honest competition and orchestrated price coordination is also referred to as “tacit collusion” – recognizing that it has the same adverse effect on competition as intentional, “express” collusion, even though it is not treated as unlawful under the antitrust laws.[6]

The use of computer algorithms – and increasingly, their use in more sophisticated artificial intelligence – to manage companies’ determination of optimum pricing has re-opened the questions around tacit collusion. Is it technologically feasible for an algorithm to engineer inflated prices by tacit collusion? Does tacit collusion become easier and more likely with the aid of an algorithm? Where might it occur? How would it be detected by antitrust enforcers? Could current antitrust law be applied and adapted to better address the resulting harms to competition and consumers?

An algorithm can perform at light speed the component operations involved in determining optimum pricing – monitoring the prices of all competitors, and the purchases made at those prices, at various locations throughout a territory; calculating the effects of various changes in price; and adjusting accordingly. Because of this, the means of collusion are far more powerful, and the potential scale of harm is exponentially greater. Furthermore, the coordinated movements can be more subtle individually when they can occur multiple times every millisecond; this also makes them harder to detect.

In the past, collusive price coordination, whether express or tacit, has been shown to be easier to accomplish, and therefore more likely to occur, in markets where the following are true:

• the companies selling are relatively few, and the barriers to new entry by other companies are relatively high, due to high initial investment costs or other reasons, so consumers have few choices;
• the product or service is homogeneous, meaning the product or service offered by one company is essentially the same as the product or service offered by the other companies, so it is easier for the companies to converge on a price; and
• sales tend to be frequent and regular, and the price is transparent, so it is easier for the companies to monitor for changes.

These market characteristics make it easier for companies to coordinate their collusion, and easier for them to enforce the collusive agreement by facilitating the detection of any deviation from the agreed-upon inflated price by a company seeking to sneak an advantage over the others.[7] A classic example of a market susceptible to collusion is retail sales of gasoline. Tech markets that exhibit these characteristics include virtual private networks (VPNs), online ride-hailing, digital advertising, and cloud storage, among others. (And with the growth of e-commerce, the market for any product or service can be a tech market.)

An algorithm could be a powerful tool in aid of a price-fixing agreement, by making it easier to monitor the marketplace, to calculate the inflated price to which all companies will agree, to detect when a company is not adhering to the agreed price, and to determine and impose an effective “punishment” in response.[8]

Some of the market characteristics noted above may not be as necessary for algorithmic collusion, thanks to the light-speed monitoring and adjustments that algorithms are capable of. For example, the products and services may not have to be as homogeneous, or priced exactly the same, as long as they are similar enough that consumers see them as reasonable substitutes for each other. An algorithm can more easily take into account variations and assign appropriate price differentials that still result in prices being inflated above their competitive market levels.

But if the use of an algorithm could facilitate coordinated pricing, it could also make it easier for enforcers to detect and prove it. Proving unlawful price fixing requires evidence of mutual anticompetitive intent – of a de facto agreement – a “meeting of the minds,” a mutually communicated understanding, a “conscious commitment to a common scheme.”[9] This evidence can be circumstantial, but if circumstantial evidence is relied upon, it cannot be equally consistent with conscious parallelism; it must suggest the existence of a de facto agreement.[10] This circumstantial evidence is likely to be present in a similar fashion whether or not an algorithm is used to facilitate the agreement.

But the use of an algorithm could make it easier for enforcers and courts to confidently ascribe anticompetitive intent to interactions that they previously had to give the benefit of the doubt and accept as procompetitive or benign – as mere conscious parallelism. An algorithm can provide a window into the mind of the programmer, almost like a diary entry or a “smoking gun” email communication – if enforcers know what to look for.

For example, an algorithm could be programmed to test where the sustainable maximum price is, by experimenting with incremental price increases to see if other companies follow. Or it could be programmed to promptly follow another company’s price increase, but to be slower in following another company’s price decrease. Or it could be programmed to retaliate against another company’s price decrease with an even greater price decrease of its own, targeted at the other company and the places where it has most of its sales – thereby not only erasing the other company’s opportunity for increased sales it hoped to achieve by its price-cutting but punishing it even further by taking away some of its existing customers.

These anticompetitive actions, if performed discreetly by humans, could be difficult for enforcers to detect, and even more difficult to ascribe intent to. However, an algorithm’s code can provide a roadmap into the mind of the human programmer. And if the programmer was acting as the agent of the company using the algorithm, or acting at its request, the intent revealed in the programming could be ascribed to the company. Evidence of an agreement is still needed to prove a case of explicit unlawful collusion. But if more than one company is using the same algorithm, or algorithms designed in conjunction with each other, or algorithms programmed to monitor each other, it may be easier to infer an agreement by showing that both companies share the same anticompetitive intent.

Along with examining the algorithms’ code and how companies are using them, enforcers can also look for traditional tell-tale indicators of possible collusion as cause for closer investigation. For example, prices that seem “stickier” in staying high despite changes in cost or consumer demand.[11] There might also be a pattern of price changes suggesting retaliation against a price-cutter, and subsequent harmonization could be circumstantial evidence of an agreement between the price-cutter and the retaliator that brings the penitent price-cutter back into the collusive fold.

But what if there is no indication that the company or the programmer had premeditated intent for the algorithm to facilitate collusion? What if the initial programming was ostensibly neutral, and the algorithm has “figured out” on its own (i.e., through “machine learning”) how to coordinate the company’s pricing with other companies in a way that leads to everyone’s prices settling at higher levels, and with higher profits for all participating companies, as a result of their not competing?

Can the current antitrust laws effectively address these new challenges? What adjustments to those laws might be useful?

As explained above, an important part of the reason tacit collusion has been accepted, or acquiesced in, is that it is so difficult to confidently judge the motive for what appears to be coordinated pricing. And having the algorithm available to examine could help clarify that motive, allowing enforcers to identify coding instructions that are inconsistent with pricing competitively. So if two or more companies selling similar products or services are using algorithms that are programmed to enable anticompetitive pricing, that can be evidence of, at minimum, a deliberate facilitating practice that foreseeably leads to inflated prices. That might be a rule-of-reason violation, or it might even give rise to a presumption of a per se price-fixing agreement. With machine learning, on the other hand – where the algorithm is given general instructions to optimize pricing and “learns” on its own to do so through coordination with other companies’ pricing – the companies could try to further distance themselves from the algorithm’s actions.

They could claim that they did not set out to program their algorithm to coordinate with competitors to keep prices inflated and that they are as surprised as anyone that their algorithm may have figured out on its own how to do so. However, even in this situation, the algorithm provides a useful window. Here it’s a window that also enables the company using it – or the programmer on the company’s request – to monitor and make follow-up assessments of how the algorithm is operating in practice. So enforcers and courts could make a similar presumption that holds companies legally accountable for setting their pricing algorithms loose on the marketplace with a “set it and forget it” blessing, and never following up on the algorithm’s functioning and capability.

In order for these situations to give rise to antitrust liability, enforcers and the courts would need to recognize that the traditional reasons that conscious parallelism had to be given the benefit of the doubt no longer apply. Today, it is possible to “read the mind” of an algorithm, so the company employing it can be held accountable when it fails to do so in order to keep the algorithm from engaging in coordination that leads to inflating prices above competitive levels. This higher enforcement sensitivity enabled by “investigating the algorithm” would still be entirely consistent with traditional antitrust principles – still deferring to companies acting independently in the free marketplace to decide how to competitively offer their products and services. Enforcers, and the courts, would need to accept that a commensurate adjustment in interpreting existing law is warranted, along with developing more technologically sophisticated analytical techniques.

If enforcers and the courts are unwilling to take these interpretive steps, or if these interpretive steps prove ineffective, Congress should consider enacting legislation to clarify the law to better enable effective antitrust enforcement against collusion by algorithm, while holding to traditional antitrust principles.

Increased transparency, through required reporting of how algorithms are designed and used, could help facilitate the detection of collusion by algorithm.[12] (Proceeding with caution, mindful that increased transparency can be a double-edged sword, also potentially facilitating anticompetitive coordination among companies.) Enforcers could also educate themselves by using algorithmic models to simulate conditions conducive to algorithmic tacit collusion and run tests to determine if, when, and how it occurs.[13]

Unless enforcers and the courts act, or Congress does, algorithms have the potential to supercharge price coordination and to lead to widespread price hikes, aggrandizing company profits at the expense of consumers forced to pay more than they should.

---

[1] Andrew Perrin & Sara Atske, About Three-in-Ten U.S. Adults Say They are ‘Almost Constantly’ Online, Pew Rsch. Ctr. (Mar. 26, 2021), https://www.pewresearch.org/short-reads/2021/03/26/about-three-in-ten-u-s-adults-say-they-are-almost-constantly-online/.

[2] E.g., Ariel Ezrachi & Maurice E. Stucke, Artificial Intelligence & Collusion: When Computers Inhibit Competition, 2017 U. Ill. L. Rev. 1775 (2017); Federal Trade Commission, Hearings on Competition and Consumer Protection in the 21st Century, Hearing 7, Session 1, Algorithmic Collusion (Nov. 14, 2018), https://www.ftc.gov/media/71288.

[3] Collusion can also take place in the other direction, with companies as buyers, agreeing to keep the prices they pay for goods and services they buy, and the wages and benefits they pay their workers, at depressed levels. Platforms for commercial transactions are considered to be sellers of services to users on both sides of the platform. And there are other forms of collusion besides agreements directly about price. For example, companies can “allocate markets” as noncompete zones, by agreeing to sell to different sets of customers. Or they might agree to slow innovations and improvements in product and service quality to a pace that is more profitable for all of the colluding companies. All of these forms of collusion cause similar harm to competition and the free market. And they could all potentially be impacted by computer algorithms and artificial intelligence. Here, we focus on price-related collusion by companies acting as sellers of goods and services, which we will refer to as “price fixing.”)

[4] Price Fixing, Bid Rigging, and Market Allocation Schemes: What They Are and What to Look For, Dep’t. of Just. Antitrust Div. (2021), https://www.justice.gov/atr/file/810261/download.

[5] See, e.g., In re Text Messaging Antitrust Litig., 782 F.3d 867 (7th Cir. 2015).

[6] See Ariel Ezrachi & Maurice E. Stucke, Sustainable and Unchallenged Algorithmic Tacit Collusion, 17 Nw. J. Tech. & Intell. Prop. 217, 224 (2020).

[7] See Marc Ivaldi et al., The Economics of Tacit Collusion, Eur. Comm’n (March 2003), https://ec.europa.eu/competition/mergers/studies_reports/the_economics_of_tacit_collusion_en.pdf.

[8] See Ariel Ezrachi & Maurice E. Stucke, Algorithmic Collusion: Problems and Counter-Measures, at 3-4, Roundtable on Algorithms and Collusion, OECD, June 2017, https://one.oecd.org/document/DAF/COMP/WD%282017%2925/en/pdf.

[9] Monsanto Co. v. Spray-Rite Serv. Corp., 465 U.S. 752, 764, 768 (1984).

[10] See Ezrachi & Stucke, supra note 5, at 222-23.

[11] See Francisco Beneke & Mark-Oliver Mackenrodt, Remedies for Algorithmic Tacit Collusion, 9 J. of Antitrust Enf’t., at 161-62 (2021).

[12] Ezrachi & Stucke, supra note 2, at 1806-07.

[13] Ezrachi & Stucke, supra note 5, at 258.