Can Pricing Algorithms Collude? Three Mechanisms, Three Very Different Evidentiary Burdens

Published 2026-09-19. Last updated 2026-09-19. Sources are listed and linked at the end of this page.

Yes, but the word "collude" is carrying three separate meanings, and only two of them currently produce liability. Where an algorithm implements an agreement that humans reached, it is ordinary per se price fixing and the code is simply evidence: the US Department of Justice secured a guilty plea on that theory in 2015, and the UK Competition and Markets Authority issued an infringement decision on the same facts in 2016. Where competitors feed nonpublic data to a shared vendor whose software prices them all, the question is whether each participant joined a common scheme, and that theory is live but contested: the United States filed a proposed final judgment against a revenue management software supplier on 24 November 2025, entered without any admission of fact or law. Where self-interested learning agents arrive at supracompetitive prices with no communication at all, the economics is real and replicated, but there is no decided case anywhere, and in the United States the conduct most closely resembling it, conscious parallelism, is lawful. A joint study by the French and German competition authorities states plainly that on this third scenario "so far there seems to be no case practice, and the relevance of this scenario is yet to be confirmed."

Why the distinction is the whole case

The three mechanisms differ in what a plaintiff or an agency must prove, and they differ far more than the shared vocabulary suggests. The first requires proof of an agreement that already exists in emails and testimony. The second requires proof that each participant knowingly joined a scheme running through a common intermediary. The third requires proof of an agreement that, by construction, never occurred.

Mechanism one: the algorithm as the instrument of a cartel

This is settled law and it was settled early. In April 2015 the Antitrust Division announced a felony charge against an online poster seller who agreed to plead guilty to fixing prices on a marketplace. The department's account of the conduct is precise about the role of the software: the defendant and his co-conspirators "adopted specific pricing algorithms for the sale of certain posters with the goal of coordinating changes to their respective prices and wrote computer code that instructed algorithm-based software to set prices in conformity with this agreement." The defendant agreed to pay a $20,000 criminal fine. The division's statement at the time set the rule that has held since: "We will not tolerate anticompetitive conduct, whether it occurs in a smoke-filled room or over the Internet using complex pricing algorithms."

The UK reached the same place on overlapping facts. In August 2016 the Competition and Markets Authority issued a formal decision that two online sellers of posters and frames had broken competition law by agreeing not to undercut each other's prices on a marketplace, and imposed a fine of £163,371 on one of them, the other having received immunity for reporting the arrangement.

Note what the algorithm contributed in both matters. It made the agreement cheap to monitor and cheap to enforce, which is why cartels want one. It did not create the agreement and it did not change the legal test. For counsel, mechanism one is a documents case, not a data science case.

Mechanism two: hub and spoke through a shared vendor

This is where the contested litigation sits. In August 2024 the Justice Department and eight state attorneys general filed a civil antitrust suit against a revenue management software supplier, alleging violations of Sections 1 and 2 of the Sherman Act. The alleged structure is the classic hub with spokes: competing landlords agree to share nonpublic, competitively sensitive information about rental rates and lease terms with the vendor, whose software trains on that pooled data and returns pricing recommendations to each of them. The department's announcement framed the legal point directly: "Using software as the sharing mechanism does not immunize this scheme from Sherman Act liability," and, more bluntly, "Training a machine to break the law is still breaking the law." These are allegations in a filed complaint, not findings.

The most useful document for counsel is not the complaint but the Statement of Interest that the United States, through the Antitrust Division and the Federal Trade Commission, filed in the related private litigation. It addresses the argument defendants actually make, which is that retained pricing discretion defeats a price-fixing claim. The government's position is that "competitors may not agree to fix the starting point of pricing" even where the prices finally charged vary from that starting point. It also takes aim at the assumption that an algorithmic case must be built like a circumstantial oligopoly case, stating that concerted action "can be proven in various ways and does not require proof of parallel conduct and plus factors."

Europe reached a comparable structural result years earlier and through a different door. In Case C-74/14, decided in January 2016, the Court of Justice held that travel agencies using a common booking system could be presumed to have participated in a concerted practice when the system administrator sent them a message capping discounts and then made the technical change to enforce it, "unless they publicly distanced themselves from that practice, reported it to the administrative authorities or adduce other evidence to rebut that presumption." The Court also drew a limit that litigators should read carefully: "The presumption of innocence precludes the referring court from considering that the mere dispatch of that message constitutes sufficient evidence to establish that its addressees ought to have been aware of its content." Awareness has to be proved. Shared infrastructure alone is not enough.

What the remedy targets is as instructive as the theory. The proposed final judgment filed on 24 November 2025 was entered, in its own words, "without this Final Judgment constituting any evidence against or admission by any party relating to any issue of fact or law." Its substantive prohibitions are about the data pipe rather than the mathematics: the defendant must "cease using current or historical Unaffiliated Property Data in the Runtime Operation of any Revenue Management Product," and where historical competitor data may still be used for model training, it must be at least 12 months old and not drawn from active leases. Nobody was ordered to stop using an algorithm. The order governs which data may enter it and how stale that data must be.

Pending legislation points the same way. The Preventing Algorithmic Collusion Act of 2024, introduced in the Senate on 30 January 2024, would make it unlawful for a person "to use or distribute any pricing algorithm that uses, incorporates, or was trained with nonpublic competitor data," with a civil penalty framework and a written-report audit power exercisable by the Attorney General or the Commission. The bill has not been enacted, and it is drafted around the data input, not around learned behaviour.

Mechanism three: autonomous tacit collusion

Here the economics is much stronger than the law. The 2020 American Economic Review study of Q-learning agents in a repeated Bertrand setting reported that "the algorithms consistently learn to charge supracompetitive prices, without communicating with one another," sustained by strategies with a finite punishment phase followed by a gradual return to cooperation, and robust to asymmetries in cost or demand and to changes in the number of players. A 2021 RAND Journal of Economics study extended the result to sequential pricing, finding that "competing reinforcement learning algorithms can indeed learn to converge to collusive equilibria when the set of discrete prices is limited," with average profitability supracompetitive and roughly 67 percent of runs converging on a Nash equilibrium across 1,000 simulations at a learning horizon of 500,000 periods.

Those horizons matter. Five hundred thousand learning periods is not a quarter of trading in a real market, and the same author says so: "it is unlikely that pricing algorithms in use are completely and fully based on Q-learning," because of practical limitations. The simulations establish possibility, not prevalence.

Field evidence is thinner and points in both directions. The CMA, reviewing what it called the first empirical analysis of algorithmic pricing and competition in a real market, reported that "German retail petrol stations increased their margins by around 9 percent after adopting algorithmic pricing, but only where they faced local competition," with margins not rising until roughly a year after market-wide adoption. A study of marketplace repricing tools found the opposite first-order effect: sellers that adopt repricing drop their prices by 16.93 percent, and market prices fall by 9.67 percent. The collusive effect in that data is narrower and strategy-specific. Sellers developed "resetting" strategies that periodically raise prices hoping rivals follow, and the study found that "these strategies are effective at coaxing competitors to raise their prices," with competitor and market prices eventually rising by 11.4 percent in markets with fewer than six serious competitors.

The legal position on mechanism three is the gap. The leading law and economics treatment of the question concludes that "collusion by software programs which choose pricing rules without any human intervention is not a violation of Section 1 of the Sherman Act," and proposes, as a remedy for that gap, that "There is a per se prohibition on certain pricing algorithms" with defined properties. That is a proposal for new law, which tells you what the author thinks existing law does not reach. The four-scenario taxonomy in the University of Illinois Law Review reaches the same conclusion from the other direction, describing the autonomous case as one where coordination "is not the fruit of explicit human design but rather the outcome of evolution, self-learning, and independent machine execution."

The enforcement agencies agree, in writing. The United States told an OECD roundtable that implementation of pricing policies by one firm is unilateral conduct whether or not it factors in the prices of competitors, and is not actionable under Section 1 "without evidence establishing an agreement with another firm over the purpose or effect of a pricing algorithm." The CMA's own assessment is equally candid: "It is as yet unclear that competition authorities can object to hub and spoke and autonomous tacit collusion situations where, for example, there may not have been direct contact between two undertakings or a meeting of minds between them to restrict competition."

Method and limits: what this evidence does not establish

Three things, stated plainly.

First, none of the simulation literature establishes that autonomous algorithmic collusion is occurring at scale in real markets. It establishes that reinforcement learners can reach supracompetitive outcomes in stylised duopolies with small discrete price grids and very long learning horizons. The joint study by the French and German competition authorities says of the purely autonomous scenario that "so far there seems to be no case practice, and the relevance of this scenario is yet to be confirmed." Treating a simulation result as a market finding is the most common error in this area.

Second, a pricing dataset cannot by itself separate mechanism three from ordinary oligopoly. Supracompetitive margins, parallel movement, fast matching and punishment-and-recovery cycles are all consistent with independent profit-maximising firms responding to the same public information. This is inconvenient for anyone selling algorithmic collusion screening, and it should be said out loud: a statistical marker in price data is a reason to look at documents, data-sharing contracts and vendor architecture. It is not proof of an agreement, and in the United States it is not proof of a violation at all, because conscious parallelism without more is lawful. Screens narrow the search. They do not carry the burden.

Third, the widely cited prevalence figures describe monitoring, not coordination. The European Commission's e-commerce sector inquiry found that a majority of retailers track competitors' online prices and that "Two thirds of them use automatic software programmes that adjust their own prices based on the observed prices of competitors." That is a statement about tooling. It is not evidence of any agreement, and it applies equally to markets where algorithmic pricing made prices fall.

Where an audit does earn its cost is mechanism two. Whether nonpublic competitor data entered a model, at what age, in what aggregation, and whether the vendor returned outputs derived from rivals' data, are factual questions with documentary and technical answers. That is exactly the territory the November 2025 proposed judgment regulates, and exactly what the pending Senate bill would make reportable on request.

Frequently asked questions

Is tacit collusion illegal in the US?

No, not on its own. Parallel pricing without an agreement, usually called conscious parallelism, is not a Section 1 violation. The United States has told the OECD that pricing conduct by a single firm is unilateral whether or not it accounts for competitors' prices, and is not actionable "without evidence establishing an agreement with another firm over the purpose or effect of a pricing algorithm." Adding an algorithm to tacit conduct does not supply the missing agreement.

How do you prove algorithmic collusion in court?

It depends entirely on which mechanism is alleged. For an algorithm implementing a human agreement, you prove the agreement the ordinary way and use the code as corroboration. For a hub-and-spoke claim, the United States has argued that concerted action "can be proven in various ways and does not require proof of parallel conduct and plus factors," which points at the delegation and the data-sharing terms rather than at price correlations. For an autonomous claim there is no established route, which is why proposals exist to create one.

What is the difference between conscious parallelism and tacit collusion?

In practice the phrases are used for the same market outcome: rivals sustaining prices above the competitive level without any communication. Conscious parallelism is the legal term of art for the conduct that US law treats as lawful absent an agreement. Tacit collusion is the economic description of the outcome. The distinction that decides cases is not between those two labels but between both of them and concerted action.

What is a hub-and-spoke conspiracy in antitrust?

An arrangement where competitors do not deal with each other directly but each deals with a common intermediary that coordinates them, so the hub carries information or instructions between the spokes. In the algorithmic version the hub is a pricing or revenue management vendor and the information is nonpublic competitor data. Liability still requires that the spokes joined a common scheme. Case C-74/14 shows both halves of that: a presumption of participation can attach to users of a common system, and the presumption of innocence still bars treating mere dispatch of a message as proof of awareness.

Is algorithmic pricing illegal?

No. Setting your own prices with software, including software that reads public competitor prices, is lawful and widespread. What creates exposure is an agreement with competitors, or the use of rivals' nonpublic data through a shared vendor. The settlement filed in November 2025 illustrates the line: it restricts which data may enter a revenue management product and how old that data must be, and it does not prohibit algorithmic pricing.

Do pricing algorithms raise prices?

Sometimes, and not always. The best field evidence available cuts both ways. Sellers adopting marketplace repricing tools cut their own prices by 16.93 percent on adoption, while resetting strategies in concentrated markets raised prices by 11.4 percent, and German petrol station margins rose by about 9 percent after adoption, but only where local competition existed. Any claim that algorithmic pricing uniformly raises prices is stronger than the published evidence.

Sources

1. "Artificial Intelligence, Algorithmic Pricing, and Collusion," American Economic Review 110(10), 2020. https://www.aeaweb.org/articles?id=10.1257/aer.20190623
2. "Autonomous algorithmic collusion: Q-learning under sequential pricing," RAND Journal of Economics 52(3), 2021. https://research-portal.uu.nl/ws/files/243482156/The_RAND_J_of_Economics_-_2021_-_Klein_-_Autonomous_algorithmic_collusion_Q_learning_under_sequential_pricing.pdf
3. "Developing Competition Law for Collusion by Autonomous Artificial Agents," Journal of Competition Law and Economics, 2018. https://joeharrington5201922.github.io/pdf/Collusion%20and%20Autonomous%20Pricing_18.11.13.pdf
4. "Artificial Intelligence & Collusion: When Computers Inhibit Competition," University of Illinois Law Review 2017. https://illinoislawreview.org/wp-content/uploads/2017/10/Ezrachi-Stucke.pdf
5. "Algorithmic Pricing, Price Wars and Tacit Collusion: Evidence from E-Commerce," working paper, 2025. https://lmusolff.com/papers/Algorithmic_Pricing.pdf
6. "Former E-Commerce Executive Charged with Price Fixing in the Antitrust Division's First Online Marketplace Prosecution," US Department of Justice, 6 April 2015. https://www.justice.gov/opa/pr/former-e-commerce-executive-charged-price-fixing-antitrust-divisions-first-online-marketplace
7. "Justice Department Sues RealPage for Algorithmic Pricing Scheme that Harms Millions of American Renters," US Department of Justice, 23 August 2024. https://www.justice.gov/opa/pr/justice-department-sues-realpage-algorithmic-pricing-scheme-harms-millions-american-renters
8. "Statement of Interest of the United States of America," United States Department of Justice and Federal Trade Commission, filed 1 March 2024. https://www.ftc.gov/system/files/ftc_gov/pdf/YardiSOI-filed%28withattachments%29_0.pdf
9. "Proposed Final Judgment [RealPage, Inc.]," Antitrust Division, US Department of Justice, 24 November 2025. https://www.justice.gov/atr/case-document/proposed-final-judgment-298 and https://www.justice.gov/d9/2025-11/435436.pdf
10. "CMA issues final decision in online cartel case," Competition and Markets Authority, 12 August 2016. https://www.gov.uk/government/news/cma-issues-final-decision-in-online-cartel-case
11. "Algorithms: How they can reduce competition and harm consumers," Competition and Markets Authority, 2021. https://www.gov.uk/government/publications/algorithms-how-they-can-reduce-competition-and-harm-consumers/algorithms-how-they-can-reduce-competition-and-harm-consumers
12. "Algorithms and Collusion - Note by the United States," DAF/COMP/WD(2017)41, OECD Competition Committee, 26 May 2017. https://www.ftc.gov/system/files/attachments/us-submissions-oecd-other-international-competition-fora/algorithms.pdf
13. "Algorithms and Competition," joint study by the Autorité de la concurrence and the Bundeskartellamt, 2019. https://www.bundeskartellamt.de/SharedDocs/Publikation/EN/Berichte/Algorithms_and_Competition_Working-Paper.pdf?__blob=publicationFile&v=3
14. Judgment of the Court in Case C-74/14, Eturas and Others, 21 January 2016. https://eur-lex.europa.eu/legal-content/en/TXT/?uri=CELEX:62014CJ0074
15. "Report from the Commission to the Council and the European Parliament: Final report on the E-commerce Sector Inquiry," COM(2017) 229 final. https://competition-policy.ec.europa.eu/system/files/2021-10/2017_ecommerce_SI_final_report_en.pdf
16. S.3686, Preventing Algorithmic Collusion Act of 2024, introduced 30 January 2024. https://www.congress.gov/118/bills/s3686/BILLS-118s3686is.pdf

Read this page as Markdown