When rents rise, blaming the algorithm is easier than building apartments. New Jersey has chosen the easier target.
On July 20, Gov. Mikie Sherrill signed the Forbidding the Algorithmic Inflation of Rent Act, or FAIR Act, declaring that landlords who use shared pricing tools are engaging in “collusion by algorithm.” The phrase is built for a press release. As an antitrust standard, it sweeps far too broadly.
Antitrust law already has a clear target. A software vendor can serve as the hub of a cartel by collecting competitively sensitive information, relaying rivals’ plans, pressuring users to accept common prices, restricting discounts, or helping participants detect and punish defections. If landlords use software to carry out an agreement that would be illegal around a conference table, the software offers no immunity. The U.S. Justice Department’s (DOJ) RealPage case and proposed settlements show the kind of conduct that warrants close scrutiny.
But shared pricing software, common data, and even some use of nonpublic information do not, by themselves, establish collusion. New Jersey has replaced a difficult, fact-intensive inquiry with a sweeping ban. The law may suppress tools that improve pricing accuracy, reduce costly errors, increase capacity use, and help firms respond to changing conditions.
The state has targeted a technology because it can facilitate unlawful coordination. Antitrust law should target the agreements and practices that suppress independent rivalry.
The rule is simple enough. Prosecute collusive agreements and the mechanisms that sustain them. Do not ban computation, common code, or nonpublic data merely because several firms use them.
When Similar Prices Become a Crime
The FAIR Act reaches far beyond a vendor telling two landlords to charge the same rent. It defines an “algorithmic device” broadly and bars rental owners from paying for or using the services of a “coordinator.”
A coordinating function includes collecting competitively sensitive information from multiple owners and using it to recommend rents, lease terms, or occupancy levels. It also covers setting terms based on another owner’s sensitive information. More vaguely, it reaches recommendations made to two or more owners through the same or a substantially similar algorithm when those recommendations facilitate “parallel pricing coordination.”
The law’s treatment of information is broader still. “Nonpublic” information generally means information unavailable to the public at no cost. A dataset that mixes public and nonpublic material counts as entirely nonpublic. The statute also reaches tacit coordination, which may be inferred from a pattern of parallel conduct.
Those provisions elevate two weak proxies into major grounds for liability. One is the use of data that costs money. The other is the appearance of similar conduct among competitors.
The FAIR Act does not ban every algorithm a landlord might use. A unilateral tool trained only on the owner’s own data, or a basic spreadsheet that requires human analysis, may fall outside its core. In practice, though, the law creates something close to a categorical ban on a large class of third-party revenue-management products.
A vendor that serves several landlords must now worry that individualized recommendations produced by similar code will be treated as parallel pricing coordination. That risk remains even when users can reject the recommendations and never see or learn their rivals’ data.
Other states have taken different approaches. New York’s 2025 law treats it as an unlawful agreement when a landlord knowingly or recklessly relies on a coordinating algorithm that collects data from multiple owners, processes that data, and recommends rental terms.
California chose a more disciplined rule in Business and Professions Code Section 16729. It makes use of a common pricing algorithm unlawful when the use forms part of a contract, combination, or conspiracy that restrains trade, or when one person coerces another to adopt the algorithm’s recommendation. California at least keeps agreement or coercion at the center of the offense.
That distinction is fundamental to antitrust law. Courts have long separated unlawful concerted action from lawful conscious parallelism. Competitors often respond in similar ways to common costs, demand shocks, regulation, interest rates, or publicly visible prices. Parallel conduct may support an inference of agreement, but it does not itself prove one.
New Jersey threatens to blur that line whenever software makes parallel behavior easier to detect.
The Code Is Not the Cartel
Former Federal Trade Commission (FTC) Acting Chair Maureen Ohlhausen made the central point in her 2017 speech, “Should We Fear the Things That Go Beep in the Night?” Algorithms can serve benign or malign ends, and familiar antitrust principles can usually tell the difference.
A cartel does not become lawful because its members communicate through code. Independent pricing does not become collusion merely because computers help firms process information faster.
Some skepticism remains warranted. Laboratory studies have found that reinforcement-learning agents can sometimes produce prices above competitive levels in repeated-game settings, even without explicit instructions to collude. Emilio Calvano and his coauthors, for example, reported in the American Economic Review that certain pricing algorithms learned strategies that sustained supracompetitive prices.
Other studies have reached different results depending on market structure, algorithm design, and the form of the recommendation. Algorithms are not a single species. Their competitive effects depend on how they work and how firms use them.
Antitrust analysis should focus on those details. Does the provider secure a common commitment from rivals? Does the system constrain or punish deviations? Can users monitor one another’s conduct? Does the vendor transmit firm-specific plans? Do firms communicate about pricing strategy? Is adoption so widespread in a concentrated market that the provider can discipline competition?
Those facts help distinguish software that facilitates an agreement from software that merely improves decision-making.
By contrast, little follows from the fact that several firms use the same vendor, receive recommendations from the same codebase, or respond similarly to market conditions. Businesses routinely rely on common accounting software, cloud infrastructure, payment processors, consultants, and data providers. Shared inputs do not ordinarily convert independent decisions into concerted action.
Pricing software should receive the same treatment unless its design or use supplies the missing agreement.
The Rent Is High for a Reason
The case for restraint goes beyond avoiding false positives. Algorithmic pricing can improve how markets work.
Friedrich Hayek’s classic essay, “The Use of Knowledge in Society,” explains that economic knowledge is scattered among millions of people and often concerns fleeting facts about time and place. Prices transmit that information without requiring any central planner to gather it all. A price change can reflect scarcity, abundance, shifting preferences, or higher costs, and it prompts people to adjust.
Rental housing depends on this kind of local knowledge. The economically relevant price of an apartment turns on the unit’s features, vacancy rates, lease length, expected turnover, seasonality, concessions, maintenance costs, nearby construction, neighborhood demand, and how much renters value moving now rather than later.
A human manager can process some of that information. An algorithm may process more of it, more consistently, and with less delay.
That does not turn the market into a centralized plan. A well-designed model can make prices more responsive. If demand weakens, it may recommend a lower effective rent or a larger concession before a unit remains vacant for months. If a local shortage emerges, it may reveal the value of adding units, renovating marginal properties, or directing investment toward the constrained area.
More accurate prices can reduce vacancies and shortages, improve matches between renters and units, and reveal opportunities for mutually beneficial exchange.
I recently made this point in Truth on the Market. Legal rules that make firms afraid to use pricing tools can reduce market efficiency at the expense of both producers and consumers. Economists generally define algorithmic pricing as the automated use of software to adjust prices in response to information. That automation can reduce decision costs, respond quickly to changes in inventory and demand, and correct the inertia and rough rules of thumb that often keep prices wrong for too long.
An efficient price is not always a low one. When housing is scarce, an accurate price may be high. Suppressing the price signal does not produce more apartments. It may instead lead to longer searches, arbitrary rationing, poorer maintenance, lower investment, or hidden nonprice terms.
New Jersey itself recognizes that its housing agenda must include more construction and land-use reform. The state’s builders are therefore right about the order of causes. Algorithms did not create zoning restrictions, construction costs, interest rates, or the shortage of developable land.
Blaming the messenger may create the appearance of action while leaving the shortage intact.
Private Data Is Not a Smoking Gun
The strongest economic critique of this emerging legal approach comes from economist Jay Ezrielev’s “Premature Antitrust Standards in Algorithmic Pricing.” Ezrielev focuses on common-data algorithms, which collect information from multiple firms and use a shared model to generate individualized recommendations.
Courts and lawmakers increasingly treat pooled nonpublic data as a near-conclusive warning sign. Ezrielev explains why that shortcut fails. The mere fact that information is nonpublic says little about how collusion would occur.
The traditional concern with sharing competitively sensitive information is straightforward. Rivals may use it to identify one another’s prices or output, monitor compliance with a cartel, and punish cheating. But a combined dataset visible only to an algorithm may give users no way to observe a rival’s conduct. The software may produce a recommendation without revealing any firm-specific data behind it.
Under the right conditions, aggregating nonpublic data can also improve competition. Sparse or noisy local markets are hard to forecast using public listings alone. Combined data on occupancy, renewals, cancellations, and transactions may help a model detect demand changes that no single owner could identify reliably.
Better forecasts can reduce vacancies, improve the timing of concessions, increase occupancy and output, and lower the cost of pricing errors. Those gains can benefit landlords and renters alike.
The same principle appears throughout the economy. Firms routinely entrust sensitive information to accountants, lawyers, investment bankers, insurers, cybersecurity vendors, and cloud providers. The quality of those services may improve as the intermediary gains broader experience. Antitrust law usually asks whether the arrangement weakens incentives to compete or enables coordination. It does not infer a cartel merely because several clients share private information with the same intermediary.
Ezrielev also identifies a paradox in the public-versus-nonpublic distinction. If legal risk depends on whether data are publicly available, firms may respond by publishing more of it. Yet public, firm-specific information can make tacit coordination easier by allowing rivals to detect price cuts and departures from a common pattern almost immediately.
A rule designed to prevent monitoring may therefore encourage the transparency that makes monitoring easier.
None of this makes nonpublic data irrelevant. Current, detailed, rival-specific information poses greater risks when a vendor reveals it to users, creates dashboards that allow reverse engineering, brings competitors together to discuss strategy, or uses the data to enforce a common objective.
Those are questions about design and conduct. They require examining access, aggregation, delay, anonymization, recommendation structure, and incentives. They do not justify New Jersey’s blunt rule that data unavailable to the public at no cost are presumptively suspect, much less its decision to treat any mixed dataset as wholly nonpublic.
The economically meaningful distinction is whether the system helps competitors coordinate and police a common plan or helps each user make a better independent decision.
A Ban Built for the Biggest Landlords
Broad bans also impose long-term costs that disappear in a debate focused on this month’s rent.
Pricing is a major commercial use of machine learning because it generates frequent feedback. Developers can test forecasts, learn from errors, improve data systems, and build better tools for inventory, logistics, capacity planning, and demand estimation. Rules that exclude useful data or expose a provider to liability because two customers receive recommendations from similar models reduce the payoff from that experimentation.
The burden will fall unevenly. Large landlords can build proprietary systems using their own portfolios, engineers, and legal departments. Smaller owners are more likely to rely on third-party tools that spread development costs across many customers.
A ban on shared platforms may therefore protect firms large enough to develop the technology in-house while denying smaller rivals comparable capabilities. A rule sold as a check on large corporate landlords could end up strengthening them.
A state-by-state patchwork makes matters worse. Software providers build products for national markets. Faced with conflicting definitions of “nonpublic data,” “coordination,” and “algorithmic device,” they may design every product to satisfy the strictest state or stop serving smaller markets altogether.
Startups will struggle more than incumbents with those fixed compliance costs. The likely consequences include less entry, slower model improvement, and fewer experiments with tools that might lower costs or expand output.
That result also conflicts with the federal push for leadership in artificial intelligence. The White House’s “America’s AI Action Plan” identifies faster innovation and adoption as central to economic competitiveness. One need not endorse every part of that plan to see the tension. The federal government wants firms to develop and deploy artificial intelligence, while states prohibit major classes of learning and decision tools without requiring proof of competitive harm.
Antitrust errors do more than raise compliance costs today. False positives can redirect research spending and determine which technologies reach the market tomorrow.
Police the Cartel, Not the Code
The answer is not to ignore algorithmic cartels. It is to adopt a federal framework that targets the mechanisms of coordination while leaving room for beneficial experimentation.
First, antitrust law should keep agreement and competitive effects at the center of the analysis. An explicit arrangement among competitors to accept a common price, restrict output, or use a vendor to enforce discipline should face per se condemnation. A novel software system that generates recommendations while preserving independent decision-making should ordinarily receive rule-of-reason analysis. Courts should examine market power, adoption, actual operation, and efficiencies.
Parallel outcomes and common software may support an inference of collusion when paired with plus factors, meaning evidence that makes independent conduct less plausible. They should not replace proof of concerted action.
Second, the DOJ and FTC should issue guidance on algorithmic pricing. The agencies have already opened a 2026 inquiry into updated competitor-collaboration guidance and identified algorithmic pricing and information sharing as subjects for review. Any resulting guidelines should distinguish dangerous features from meaningful safeguards.
High-risk features would include a common commitment to follow recommendations, coercion or penalties for deviation, access to current and detailed rival-specific information, tools that expose deviations, communications among users about pricing strategy, uniform limits on discounts or price cuts, and broad adoption in a concentrated market.
The analysis should also consider the provider’s incentives, how often users receive and accept recommendations, whether outputs are individualized, and whether the system influences output or occupancy as well as price.
The agencies should pair that list with a rebuttable safe zone for systems that preserve independent rivalry. Relevant safeguards would include no user access to rival-specific data, aggregation and anonymization, appropriate delays, individualized objectives and outputs, meaningful freedom to reject recommendations, no penalties for doing so, no vendor-facilitated meetings among competitors, strong firewalls and audit logs, and documented efficiency justifications. Nonpublic data should count as one factor, not a forbidden category.
Third, federal enforcers should use statements of interest to clarify doctrine, rather than only to expand liability. The DOJ and FTC filed a 2024 statement of interest in hotel-pricing litigation explaining that an algorithm cannot immunize conduct that would otherwise violate antitrust law. That principle is sound.
Future filings should add its necessary limit. A common algorithm does not erase Section 1’s agreement requirement, and pooled nonpublic data do not by themselves establish a naked restraint. Carefully chosen filings could steer courts away from per se treatment of unfamiliar arrangements before economic evidence and experience support it.
Fourth, Congress should consider targeted preemption. A federal law could displace state and local rules that impose liability solely because firms use the same pricing algorithm, receive similar recommendations, or contribute nonpublic data without proof of agreement, coercion, or a likely anticompetitive mechanism.
States would remain free to enforce antitrust and consumer-protection laws against actual collusion, deception, discrimination, and unfair practices. The goal would be a coherent national baseline for interstate software and data services, not immunity for algorithms.
Preemption makes sense because fragmented state bans can impose costs beyond state borders. A prohibition in one large jurisdiction may dictate product design nationwide, discourage entry, and deny consumers elsewhere access to useful tools. Congress often adopts national rules when interstate commerce and networked technologies make 50 conflicting regimes unusually costly. Algorithmic pricing warrants the same consideration.
Finally, enforcement should remain empirical and open to revision. Agencies should study actual effects on prices, output, vacancies, quality, and entry before imposing permanent design rules. Remedies should address demonstrated sources of harm, such as rival-specific disclosures, coercive acceptance requirements, or monitoring tools, rather than banning entire categories of data.
Sunset provisions, regulatory sandboxes, and retrospective reviews would allow regulators to revise their approach as the technology and evidence develop.
Collusion Still Requires Collusion
New Jersey’s FAIR Act begins with a legitimate concern and ends with the wrong legal design. Algorithms can facilitate collusion. So can trade associations, consultants, phone calls, and private meetings. Antitrust law should target the agreement and the mechanisms that sustain it, rather than the general-purpose tool that carries the information.
The costs will extend beyond landlords losing a convenient product. Broad bans can make prices less informative, shield large incumbents from smaller technological rivals, encourage inefficient public disclosure, slow AI research, and fracture a national software market into incompatible state regimes.
Renters may pay through fewer available units, poorer matches, higher operating costs, and reduced investment. Meanwhile, zoning restrictions, construction costs, interest rates, and limited developable land will keep doing what they were doing before the algorithm arrived.
A sound policy would prosecute genuine hub-and-spoke cartels, provide clear federal guidance, examine actual conduct, create safe harbors for independent decision tools, and preempt state laws that mistake data processing for agreement.
Preserve rivalry. Demand evidence. And do not confuse better arithmetic with a cartel.
