Unscarcity
Sign in for free: Preamble (PDF, ebook & audiobook) + Forum access + Direct purchases Sign In

Unscarcity Research

Algorithmic Collusion: When Rivals Share One Pricing Brain

A judge blocked New York's ban on rent-setting software. Antitrust law hunts for an agreement; a shared algorithm raises rents without one.

12 min read 2600 words Updated October 2026 /a/collusion-without-agreement-algorithmic-pricing

Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

Collusion Without Agreement: When Rivals Share One Pricing Brain

On September 29, 2026, a federal judge in Manhattan blocked New York from enforcing the first state law in the country to ban landlords from setting rents with pricing software. The company that sued is RealPage, a Texas firm whose algorithms recommend what to charge for each apartment and each renewal. Its argument was that a recommendation is speech, and that New York was punishing landlords for listening. Judge Valerie Caproni called the case “a close call,” found RealPage only “marginally” likely to win, and suspended the landlord half of the statute anyway. Her reason fits in one sentence: “Under the rubric of preventing price fixing, the statute prohibits normal commercial conduct just because it is facilitated by software.”

She has a point, and so do the tenants. A landlord who walks the block reading For Rent signs breaks no law. A market in which every large landlord asks the same machine what to charge does not behave like a block full of landlords reading signs. Antitrust law was built to find an agreement between people. The question underneath RealPage v. James is what the law does when cartel prices show up and there is no agreement to find.

The answer matters well beyond rent. This site argues that cash handed out in a scarce market flows upward to whoever owns the scarce thing, and the usual way to put it is a question: do you think the landlords wouldn’t notice? They no longer have to. The software notices for them.

The machine with too little empathy

In October 2022 ProPublica described how RealPage’s YieldStar software had spread through the American apartment business. Landlords feed it their leases, renewals, and vacancies, and it returns a price for every unit, every day. In a video promoting the product, one RealPage executive asked another what role the software had played in rents that had jumped as much as 14.5%. “I think it’s driving it, quite honestly,” Andrew Bowen answered. “As a property manager, very few of us would be willing to actually raise rents double digits within a single month by doing it manually.”

That sentence is the product. A human leasing agent has a conscience, a quota, and a vacant unit that costs money every week it sits empty, and all three push toward a deal. One of the algorithm’s developers told ProPublica that leasing agents had “too much empathy.” The software has none. RealPage discouraged bargaining with renters and sometimes recommended accepting a lower occupancy rate in exchange for higher rents, and former employees said as many as 90% of its recommendations were adopted. In one Seattle neighborhood, 70% of the apartments were run by ten property managers. All ten used RealPage’s pricing software.

The Justice Department, which sued in August 2024, said the engine also ran on something no sign-reading landlord has: rivals’ private books. Its complaint alleged that RealPage fed nonpublic, competitively sensitive data from competing landlords into the recommendations it gave each of them. That December the White House Council of Economic Advisers put numbers on it: at least 10% of American rental units priced with RealPage products, an average of $70 a month added to the rent in buildings that use pricing algorithms, more than $3.8 billion taken from renters in 2023.

RealPage has always denied that its software fixes prices. It has even argued the reverse: landlords who price by hand, the company told ProPublica, “typically” phone around to check competitors’ rents, which is the older and cruder way to coordinate. Dozens of its customers have paid anyway. RealPage and the property managers named in the tenants’ class action have agreed to settlements worth about $360 million.

Antitrust is looking for a handshake

Section 1 of the Sherman Act, written in 1890, outlaws every “contract, combination… or conspiracy” in restraint of trade. Each of those words needs at least two parties who agreed to something. Charging what your rival charges because you watched your rival is called conscious parallelism, and it is legal. The law punishes the handshake, and the high price is only evidence that one may have happened.

Lawyers have one tool for a cartel whose members never met: the hub-and-spoke conspiracy. In 1939 the Supreme Court found one among eight film distributors that had never exchanged a word. Each had received the same letter from a Texas theater chain, with all eight named as recipients, and each went along knowing the others had been asked. In the algorithm cases the theory is that the software vendor is the hub and the subscriptions are the letters.

Appeals courts have now ruled twice on the same product, a hotel pricing tool called Rainmaker, and landed on opposite sides. In August 2025 the Ninth Circuit threw out a case against Las Vegas Strip hotels: licensing the same pricing software as your competitors, it held, does not by itself restrain trade. In July 2026 the Third Circuit revived a near-identical case against Atlantic City casino-hotels, finding that the plaintiffs had plausibly described a price-fixing agreement with the vendor at its center. The Ninth Circuit never reached the hub-and-spoke theory, which the plaintiffs had dropped on appeal. The Third Circuit took it on directly, and the fact it leaned on was the sharing of nonpublic data.

The same line runs through the settlement RealPage reached with the Justice Department on November 24, 2025. Live recommendations may draw only on a landlord’s own data and on public information. Rivals’ private data may train the models only once it is at least a year old. “Auto-accept,” the feature that turns a recommendation into the rent with nobody deciding, must be something a landlord can configure and override, no feature may default toward an increase, and RealPage may not reward customers for taking its advice. There was no fine. Two days later RealPage sued New York.

Take away the hub and the prices still rise

If the story ended there, the fix would be simple: ban the pooled private data and go home. The research says it does not end there.

In 2020 four economists published an experiment in the American Economic Review. They put simple reinforcement-learning algorithms into a textbook pricing game and let them play. Nobody told the algorithms to cooperate, and they had no channel to talk through. They “consistently learn to charge supracompetitive prices, without communicating with one another,” the authors wrote, and they held those prices the way cartels do, by answering a rival’s price cut with a short price war and then drifting back up.

A second team found the same pattern outside the lab. Pricing software became widely available to German gas stations in 2017. Stations that adopted it and had a competitor nearby raised their margins by 9%. In towns with exactly two stations the result was sharper. When one station automated, nothing happened. When both did, margins rose 28%. A single algorithm cannot collude. It takes two.

Large language models behave the same way. In a 2024 study, pricing agents built on LLMs “autonomously collude in oligopoly settings to the detriment of consumers,” and “seemingly innocuous phrases” in the instructions were enough to increase the effect. No shared vendor, no private data, no instruction to coordinate. Two competing managers who each tell their system to protect long-run profit may have founded a cartel without knowing it, and no court will find a letter.

This is what cheap intelligence does to a market. Tacit coordination was always possible in theory and hard in practice, because humans are slow, distracted, and tempted to cheat when a unit sits empty. Software watches every rival at once, answers within the hour, and never loses its nerve. The friction that kept oligopolists honest was a shortage of attention, and attention is no longer short.

Three ways to write the law

Legislators have tried three designs, and September’s ruling sorts them.

Ban the tool. New York’s law, in force since December 2025, amended the state’s antitrust statute to treat a landlord’s knowing or reckless use of a covered pricing recommendation as an unlawful agreement, whatever data the software was fed. That is the provision Caproni suspended, because it reaches software that only reads public listings. The half of the law aimed at the software companies still stands, and the governor’s office says “we’re not backing down.”

Ban the input. Connecticut’s statute, effective January 2026, is the first state law limited expressly to nonpublic competitor data. San Francisco, the first city to act, drew the same line in 2024, and Minneapolis’s ordinance took effect in March 2026. About a dozen cities and counties now have ordinances, and in July and August 2026 tenants began suing landlords under them in San Francisco, San Diego, Seattle, and Philadelphia, with penalties that run from $1,000 to $7,500 per violation. California’s AB 325, also in force since January, takes a different cut. It covers every industry and any “common pricing algorithm” that uses competitor data, public or not, but it outlaws two specific acts: using one as part of an agreement to restrain trade, and using one to coerce someone into adopting the price it recommends. It also lowers what a plaintiff has to allege to get into court.

Make the machine announce itself. A separate New York law requires this sentence next to any price personalized by software: “THIS PRICE WAS SET BY AN ALGORITHM USING YOUR PERSONAL DATA.” Retailers challenged it on the same First Amendment grounds RealPage used, and in October 2025 a federal judge dismissed their suit.

Put the outcomes side by side. The rule that makes the machine identify itself survived. The rules that name a mechanism (pooled private data, the price that accepts itself, pressure to comply) are being signed by the companies they bind. The rule that told landlords not to listen to software at all is the one a court stopped. The broad ban lost its first test. The narrow ones are the ones being enforced.

What the framework adds

Two of the Five Laws apply directly.

Truth Must Be Seen says every decision affecting resources or rights should be observable, auditable, and traceable. A rent is such a decision. A tenant should be able to learn that software set the number, which inputs it used, and whether a person had the power to change it. New York’s disclosure rule is a first, thin version of that right.

Freedom Is Reciprocal marks where the software’s freedom to advise ends: at the point where the bill for the advice lands on people who never agreed to it. The Justice Department’s settlement already draws that boundary in engineering terms. It sits at the shared private books and at the price nobody chose.

Both are patches, and honesty requires saying so. A sign telling you an algorithm set your rent does not lower your rent, and you cannot shop your way around a price when every building on the street consults the same adviser. Pricing software earns its keep in a particular kind of market: the buyer cannot walk away, and supply cannot respond. Housing is the purest case, which is why the biggest fight over a cartel without a handshake is about apartments and not sneakers.

That is the argument for taking essentials out of the price system altogether. Where shelter is provided as infrastructure, the way the Foundation proposes, there is no rent to optimize and nothing for a pricing brain to coordinate. Where land rents are taxed back to the community that created them, the prize for coordinating shrinks. And any plan that sends people checks instead, basic or high, has to explain what happens when the check meets the algorithm standing between the tenant and the lease. The same logic is spreading to other toll booths: robots priced against the wage they replace, compute rented by a handful of landlords. Michels called the general tendency the iron law of oligarchy. It used to require meetings.

What to watch

  • The New York case. The injunction is preliminary and the state can appeal. If the legislature narrows the statute to nonpublic data, Connecticut’s version becomes the national template.
  • The two appeals courts. Opposite outcomes on the same software in the Ninth and Third Circuits are the kind of tension the Supreme Court eventually gets asked to settle.
  • The city lawsuits. The follow-on cases filed in the summer of 2026 are the first test of whether a local ordinance can do what federal antitrust law has struggled to.
  • Agents that set prices. Every study above used software simpler than what a mid-sized landlord can rent today. The referee we ask machines to be has to be able to tell when the players are machines too.

Sources

Share this article:

Share Share Share