When Pricing Algorithms Learn to Collude

WASHINGTON — Pricing algorithms are pushing the U.S. regulatory apparatus into a direct confrontation with the limits of century-old antitrust laws. From federal courts in Washington to municipal chambers in California, regulators and lawmakers are targeting systems that can inflate corporate profits without competitors exchanging a phone call, email, or handshake.

The legal battlefront took shape in November 2025, when the U.S. Department of Justice filed a proposed settlement in its landmark antitrust lawsuit against real estate software firm RealPage. The government accused the company of enabling a digital cartel among rival landlords by pooling nonpublic rental data to artificially inflate housing costs.

RealPage agreed to a court-appointed monitor and strict limits on data collection under the proposed settlement, which still awaits federal court approval. The restrictions specifically bar the software from using recent competitor data or highly localized geographic metrics that enabled neighborhood-level price coordination. RealPage paid no financial penalty and admitted no wrongdoing, while wider private litigation against the firm continues.

For more than 135 years, U.S. antitrust enforcement has relied on the Sherman Antitrust Act of 1890. Section 1 outlaws any “contract, combination, or conspiracy” in restraint of trade. Under established U.S. judicial precedent, “conscious parallelism”—companies independently watching each other and matching prices—is not illegal unless specific “plus factors” suggest actual collusion.

Economists and legal scholars divide algorithmic market coordination into three operational models. In the “Hub-and-Spoke” model, competitors feed proprietary, nonpublic data into a common third-party software provider. The shared algorithm produces pricing recommendations for participating rivals. RealPage represents this model, which is currently the only form of algorithmic coordination that regulators have successfully prosecuted.

A different dispute involves Amazon. The “Mirror” model is at the center of the Federal Trade Commission’s sweeping antitrust lawsuit against the company, filed in September 2023 under FTC Chair Lina Khan. A single dominant firm uses proprietary software to anticipate how rivals will respond to price changes and can raise prices unilaterally when it predicts that competitors will match the increase.

According to heavily redacted court documents that were later unredacted, the FTC alleges that Amazon developed and deployed a secret internal pricing tool codenamed “Project Nessie.” The algorithm allegedly identified products for which competitors including Target and Walmart were highly likely to automatically match Amazon’s price increases.

Project Nessie would raise Amazon’s price, wait for competitors to match the hike, and then hold the price at the elevated level. The FTC alleges that the strategy generated more than $1 billion in excess profits for the e-commerce giant. The agency also claims Amazon deliberately paused the algorithm during periods of intense public or regulatory scrutiny and turned it back on after the pressure subsided.

Amazon has vigorously disputed the allegations. The company says Project Nessie was a legacy, trial tool intended to prevent its automated price-matching algorithms from entering destructive “downward spirals” that could unsustainably depress profit margins. Amazon maintains that the tool was discontinued years ago and did not harm competition.

The third model, the “Ghost,” involves independent, self-learning pricing algorithms deployed by competing firms. Through reinforcement learning, the systems observe public market data and independently learn that they can earn more by holding prices high and refusing to undercut one another. No data is shared, no third-party vendor is used, and no human agreement is made.

Evidence of that model emerged in a 2024 study published in the *Journal of Political Economy*. Economists Stephanie Assad, Robert Clark, Daniel Ershov, and Jean-François Houde examined the retail gasoline market in Germany after automated pricing software became widely available to gas stations in 2017.

In local markets where only one gas station adopted the software, retail margins did not change. Where two competing stations both adopted automated pricing, average retail margins rose by approximately 38%. The algorithms achieved the increase without direct communication, messages, or agreements between the stations; the two independent AI agents learned on their own that they earned higher profits by backing off from price wars.

Laboratory simulations had raised similar concerns earlier. In a landmark study published in the *American Economic Review* in 2020, Emilio Calvano, Giacomo Calzolari, Vincenzo Denicolò, and Sergio Pastorello set independent reinforcement-learning algorithms against one another in a simulated retail market.

The algorithms received no instructions other than to maximize their own profits and were barred from communicating. They nevertheless consistently learned to charge prices well above competitive levels. When one algorithm lowered its price to capture market share, the others immediately responded with aggressive price cuts to “punish” the defector. After the defector returned to the higher price level, the competing algorithms restored the elevated pricing structure.

Writing later in 2020 in the journal *Science*, the researchers, joined by Wharton economist Joseph Harrington, warned that delegating pricing to autonomous software opens a backdoor to systemic collusion. Harrington argued that artificial intelligence can discover and enforce collusive rules without human awareness and that antitrust law must be fundamentally rewritten to target coordination arising entirely without human agreement.

The limits of existing law have also appeared in two conflicting federal appellate rulings involving hospitality companies. Both cases concerned pricing software supplied by Cendyn and used by major hotel and casino operators.

The U.S. Court of Appeals for the Ninth Circuit dismissed an antitrust lawsuit against Las Vegas hotel operators. Because Cendyn’s software did not pool confidential, nonpublic information to produce its rate recommendations, the court ruled that competing hotels’ independent adoption of the tool did not constitute an illegal conspiracy under the Sherman Act.

One year later, the U.S. Court of Appeals for the Third Circuit reached the opposite conclusion in a similar lawsuit against Atlantic City casino operators. The court revived the case after pointing to evidence that the casinos had entered highly sensitive, nonpublic room-occupancy data into the shared system and followed its pricing output approximately nine times out of ten. The pooling of proprietary data crossed the line into an illegal hub-and-spoke conspiracy, the court ruled.

While those federal boundaries remain contested, state and local lawmakers have pursued pre-emptive restrictions. San Francisco became the first city in the country to ban algorithmic rent-setting tools in the summer of 2024. Philadelphia, Minneapolis, and Seattle quickly followed.

New York enacted the nation’s first statewide ban on algorithmic residential rent-setting software in October 2025. California amended its state antitrust laws to lower the threshold for prosecuting algorithmic price coordination.

The legislative response prompted a constitutional counter-offensive. Just days after the DOJ announced its proposed settlement, RealPage filed a federal lawsuit against New York, challenging the statewide ban. The company argues that its software recommendations are lawful, data-driven speech protected by the First Amendment of the U.S. Constitution.

The RealPage settlement has been treated by regulators as a crucial step, but economic and legal experts warn that the company represents only the most easily targetable form of a wider problem. RealPage operated as a central “hub” coordinating competitors, while newer autonomous algorithms can stabilize prices and eliminate competition independently, leaving no paper trail for antitrust investigators.

For corporate boards and executive suites, autonomous pricing agents alter regulatory and reputational risk. Companies historically evaluated pricing software through financial dashboards: if margins improved and conversion rates remained stable, the software was considered successful. A quietly coordinated market and a highly successful competitive strategy can produce identical financial metrics, allowing traditional performance indicators to mask legal exposure.

When an algorithm learns to shadow rivals and avoid undercutting them, the deploying company remains legally and reputationally responsible for the outcome. Under modern compliance standards, corporations cannot outsource that liability to third-party software vendors.

Corporate governance experts and compliance advisers recommend that boards establish structured safeguards. Data-source auditing requires companies to map exactly what information their pricing systems can access. Systems that monitor and react to competitor prices in real time carry significantly higher accidental-coordination risk than systems relying strictly on internal variables such as inventory, cost, and historical demand.

Some firms are introducing diversification and noise insertion, including randomized response delays or distinct objective functions, to prevent independent algorithms from falling into parallel pricing patterns. Behavioral audits are also expanding beyond financial performance: audit committees are beginning to demand technical accounts of the signals pricing systems use, the parameters they optimize, and the occasions on which they altered pricing patterns in direct response to competitor movements.

Management teams can use counterfactual testing by running simulated market tests in which competitor price signals are delayed, randomized, or completely removed. If corporate margins collapse only when the pricing agent is barred from shadowing its rivals, it serves as a strong indicator that the software’s profitability relies on faded market competition rather than genuine operational efficiency.

Corporations are increasingly required to maintain auditable mandates for autonomous agents. By defining clear, human-set boundaries for when and how an algorithm may react to competitor prices, companies establish a defensible paper trail showing that they have not delegated compliance decisions to an unconstrained machine.