Algorithmic pricing, collusion and the law in 2026

Algorithmic pricing and collusion meet in U.S. federal antitrust law at one requirement: an agreement. Section 1 of the Sherman Act declares illegal every "contract, combination ... or conspiracy, in restraint of trade" [1], so a claim against shared pricing software has to show that competitors agreed, directly or through the software. Since March 2024, federal enforcers and two federal appeals courts have drawn lines around that requirement [2, 8, 10]. Four states have statutes of their own [11, 15, 16, 18], and European and UK authorities have set out their views [23, 25, 27]. Statuses are as of 8 November 2026. This is research, not legal advice.

The texts at a glance

Algorithmic pricing and collusion under U.S. law

Both hotel cases below address a hub-and-spoke conspiracy. The Third Circuit describes the hub as "usually an agent, supplier, or purchaser" with whom all the competitors deal, and for a per se claim it requires a horizontal agreement among the spokes, the "rim" [10]. Without direct evidence, a plaintiff needs more than parallel prices, the plus factors: a motive to conspire, conduct against a defendant's own interests and evidence implying a traditional conspiracy [10].

What U.S. enforcers have said

On 28 March 2024, the FTC and the DOJ told the court in Cornish-Adebiyi, a New Jersey hotel case, that hotels "cannot use an algorithm to engage in practices that would be illegal if done by a real person". Plaintiffs need not identify direct communications between competitors, they wrote, and an agreement to use shared pricing recommendations or algorithms is unlawful even where participants keep some control over price [2].

RealPage: a consent decree read as a design specification

The United States and eight states sued RealPage on 23 August 2024 [6]. The amended complaint alleges that RealPage and six landlords agreed to share competitively sensitive information and to use RealPage's software to align pricing [4]. On 19 May 2026, the court in the Middle District of North Carolina entered a final judgment on the United States' claims [7]. Its terms, as proposed in November 2025, read like a specification [4].

• Within 180 days, no current or historical data from unaffiliated properties while the software runs and generates prices, with stated exceptions [4].
• Nonpublic competitor data in model training only if at least 12 months old and not from active leases [4].
• No models trained on such data that see geography finer than nationwide, or than statewide for three named models [4].
• No auto-accept without parameters the user sets, no guardrail without symmetrical upper and lower bounds, and nothing that impedes a user's override [4].
• A court-appointed monitor for three years [4].

Willow Bridge's proposed decree, published on 16 July 2026, goes further: it must obtain a certification from any other vendor whose revenue management product it uses, stating that the product meets the decree's feature rules or the RealPage judgment's [5]. On 30 September 2026, the court let the states' claims against RealPage, Pinnacle and Willow Bridge proceed, holding them not moot because the federal judgment can be terminated, does not reach all the challenged conduct and can be enforced only by the United States [7]. It decided motions to dismiss, so it took the allegations as pleaded [7].

Two appeals courts on shared hotel pricing software

On 15 August 2025, the Ninth Circuit affirmed dismissal with prejudice in Gibson v. Cendyn [8]. The hotels had licensed the same software, which was not alleged to pool or share one hotel's confidential information with another. "The choice of several competitors to contract with the same service-provider, when followed by higher prices" was not enough [8]. The plaintiffs had abandoned their hub-and-spoke claim on appeal, and a footnote says the analysis might change if confidential data had been shared among licensees [8]. The Supreme Court denied certiorari on 20 April 2026 [9].

On 29 July 2026, the Third Circuit revived the Atlantic City claims in Cornish-Adebiyi v. Caesars [10]. There, casino-hotels were alleged to send nonpublic pricing and occupancy data to Cendyn's Rainmaker software, which processed it with rivals' data and returned suggested rates that hotels charged 90 percent of the time [10]. The court held those allegations sufficient to support a finding that the hotels conspired to fix prices through the software, and said prices are fixed when agreed upon, whether or not conspirators always adhere to them [10].

State laws on common pricing algorithms

Four state statutes reviewed here target pricing software, three in rental housing [11, 15, 16, 18]:

• New York. Section 2 of General Business Law § 340-b bars knowingly or recklessly facilitating an agreement among landlords not to compete, including through software that performs a "coordinating function". Section 3 bars landlords from knowingly or recklessly setting rents on such software's recommendations [11]. On 29 September 2026, a federal court preliminarily enjoined enforcement of Section 3 as likely to fail intermediate scrutiny for commercial speech, and held that Section 2 survives [12]. A motion for reconsideration filed on 13 October 2026 had no ruling on the docket as viewed on 8 November 2026 [13, 14].
• California. AB 325 makes it unlawful to use or distribute a "common pricing algorithm" as part of a conspiracy, or to use or distribute one while coercing another person to adopt its recommended price or term [15].
• Connecticut. Using a revenue management device that runs on nonpublic competitor data to set rents or occupancy levels is an unlawful practice [16].
• New Jersey. The FAIR Act defines "parallel pricing coordination" to include tacit agreement, "mutual consent demonstrated without verbal or written communication". Under its Section 9 it takes effect on 1 July 2027, by this article's count [18].

A federal bill, the Preventing Algorithmic Collusion Act of 2025 (S. 232), was referred to the Senate Judiciary Committee on 23 January 2025, the last action in its status record [20, 21].

Europe and the UK

In Europe, France's and Germany's competition authorities wrote in 2019 that an algorithm that only observes and reacts to rivals' public behavior "might have to be categorised as intelligent adaptations to the market rather than coordination" [24].

The European Commission's 2023 guidelines on horizontal agreements say firms in illegal pricing practices cannot avoid liability on the ground that algorithms determined their prices. Like an employee, an algorithm remains under the firm's control [23]. A pricing rule in a shared tool, such as matching a particular competitor's price minus 5%, is likely caught "even in the absence of an explicit agreement to align future pricing" [23].

The UK Competition and Markets Authority (CMA) wrote in March 2026 that indirect information exchange through algorithm providers can be illegal and that it is screening for algorithmic collusion [27]. On 17 July 2026, France's Autorité de la concurrence reported stakeholders' concern that collusion could arise if agents take part in price negotiations [25], a point the article on what to test before an agent buys also covers.

What experiments add

In a 2020 study, Q-learning pricing algorithms consistently learned to charge supracompetitive prices without communicating [29]. In a 2024 study revised in August 2026, LLM pricing agents reached supracompetitive prices and profits autonomously, and small changes in prompt wording moved the result [30]. A March 2026 preprint found this collusion fragile: more competing agents and different algorithms broke it up [31]. The review of LLM agents as market participants covers the 2020 and 2024 studies in detail.

The agreement tests above turn on pooled nonpublic data, an intermediary or communication [8, 10, 23, 24]. Among the texts reviewed, New Jersey's act is the one whose definitions reach tacit agreement, and it applies to rental housing [18].

What a simulation can and cannot show

Parallel high prices alone do not show an agreement [10]. In a simulated market the competitive equilibrium is known before the first bid, so a run can show whether agents' prices rise above it without instruction, how much prompt wording moves them and how the number and mix of rivals changes the result [29, 30, 31]. The note on measuring LLM agents in simulated auctions sets out the induced-value methods behind that. Rillor's market simulation measures whether agents coordinate, tacitly or openly, to move prices, and whether what they tell other agents matches what they do, from the run log of bids, messages, prices and outcomes.

A run cannot show legal compliance. The rulings above turned on what data was pooled, what users agreed to and how often they followed recommendations [7, 8, 10], facts a laboratory run does not establish. The CMA's paper "Agentic AI and consumers", first published on 9 March 2026 and updated on 30 October 2026, says businesses remain responsible for pricing outcomes and must "understand, test and govern" the technologies they deploy [28]. A simulation result is behavioral evidence for counsel to weigh alongside those facts.

Read the full report

The full report adds the statutes' definitions, the RealPage judgment section by section, the landlord decrees, the European and UK statements and all 33 sources.