The Case for Reversing the Ban on AI Rent Pricing in San Francisco

Abstract

This memo argues against the DOJ’s ban on algorithmic pricing using private data by offering an economic analysis of RealPage’s effect on the rental unit market. First, it examines the central argument from this study that machine learning algorithms learn to collude naturally, and invalidates it based on its applicability to the real estate market. Next, it argues in favor of free information, dynamic pricing, and ensuring landlord profits drawing on economic analysis and lived experience in San Francisco. Finally, it gives a real solution to San Francisco’s high prices.

Context

In October of 2024, The Department of Justice (DOJ) banned usage of AI-assisted algorithmic pricing tools, namely RealPage, in San Francisco. These tools analyze non-public data in its interface from each participating firm and landlord to recommend optimized rent prices. The DOJ argues that this is collusion and constitutes antitrust behavior. This memo argues that the pricing recommendations are simply reflective of San Francisco’s tight housing market, and that tools like RealPage are neutral and economically beneficial.

What is RealPage?

According to DOJ v. RealPage: Nov of 2025, “RealPage is a provider of commercial revenue management software and services for the conventional multifamily rental housing industry.” RealPage is a revenue management software service (RMS) and an AI revenue management (AIRM) product that uses algorithmic pricing, which is the use of automated computer programs to set prices using real-time data according to Truth on the market. The DOJ argues RealPage “relies on private, ‘competitively sensitive’ market information from landlords; RealPage’s software has also included features designed to limit rental price decreases and otherwise align pricing among competitors.” One such feature is the ‘sold out guardrail,’ which is a runtime operation feature that “increases the recommended rental price for a floor plan when the floor plan reaches its target occupancy.” It uses what is called ‘surrogate data,’ which is “data used from a property other than the one whose rent is being recommended to supplement the transactional data (current and historical data from the Subject Property).

The DOJ argues collusion exists if non-public information is used in a shared database for recommendations by an algorithm. It claims “whether firms effectuate a price-fixing scheme through a software algorithm or through human-to-human interaction should be of no legal significance” (DOJ). This raises concerns that prices will be artificially inflated by these algorithmic tools. However, evidence from other parts of the country that use the same software tell a different story.

Prices fall elsewhere

According to MultifamilyDive, in Austin, Columbus, Miami, and Nashville, software like RealPage offers led to reductions in average price for rental units. In Austin specifically, rent prices fell 8.8% between May 2024 and May 2025 in conjunction with implementation of these software tools, in large part due to a construction boom and a large supply of housing in comparison to demand. Additionally, according to RealPage market analytics data, US rents fell 1.7% in the last quarter of 2025. The algorithms recommended cutting prices to reduce vacancy, demonstrating that they reflect market conditions rather than trying to create them.

Do machine learning algorithms collude naturally?

The landmark study

“Artificial Intelligence, Algorithmic Pricing, and Collusion” is a landmark study that argues that machine learning algorithms will naturally collude to a higher optimal price. It is cited across numerous articles examining collusion in real estate pricing, and the DOJ relied on its logic implicitly or inadvertently in its argument against the usage of RealPage. This study is the most effective at using models and measuring outcomes to argue for collusion happening naturally between autonomous machine learning agents, so deconstructing its inapplicability sheds light on the misdirection of the DOJ’s ban on pricing tools that behave similarly.  Applying a microeconomic analysis reveals that the design of the study negates the application of its results to the real estate market.

The study examines Q-learning programs, processes in which an AI agent engages in experiential learning to arrive at the best course of action. The agents interact within an Bertrand oligopoly setting, which means each agent chooses a price and treats its rivals prices as fixed (where the product is homogeneous among all agents). In this model, an agent selling at a lower price will clean out the market, so each firm (agent) must reduce its price to equal marginal cost, so that no firm makes a profit. However, this is only one equilibrium (or Nash equilibrium). There exists a second Nash equilibrium in which firms cooperate to raise their prices simultaneously without undercutting each other, which maximizes profits. The study finds that Q-learning pricing algorithms unambiguously learn to collude, though the study concedes that the real-life applicability of these findings need to be studied more.

While the study doesn’t disclose the number of agents used in the Q-learning process, the environment’s description as a Bertrand oligopoly model implies that there were only a handful of agents (representing firms). In this setting, it is natural to predict that few suppliers will learn to maximize their profit through cooperation. 

Economic implications

The DOJ argues, implicitly, that since each supplier (landlord, firm) has access to each other’s information, that this constitutes collusion. Recall, however, that the agents in the study had full knowledge of each other’s pricing, and even with this knowledge, profits were not maximized until after hundreds of thousands of trials. The feature of the study that allowed for collusion was not the information that the agents had access to, but the structure of the market itself that only allowed for a handful of suppliers. According to the robustness section of the study, the addition of agents up to four saw decreases in profit gain, confirming that “collusion is harder to sustain when the market is more fragmented.” When applied to the San Francisco rental market, which has about half a million rental units according to Bay Area Census, the condition for collusion falls apart.

A large number of buyers and sellers is a hallmark of both a perfectly competitive market and a monopolistic one with the exception of the type of product and the availability of information. While the differentiation versus homogeneity of units is unclear, making information more transparent, as RealPage does, creates a perfectly competitive market condition, not an anticompetitive one. 

Perfect access to information, which is notoriously difficult in the real estate market, is made more possible when landlords can rent based on the true sold price for their competitors’ units. A common complaint from Yardi and Zillow users is that they only know what their neighbor’s units were listed for, not the eventual agreed-to price. A pricing software’s access to private information allows for optimized pricing, and since market conditions for collusion do not exist in the rental unit market, collusion does not follow.

Market-wide price hikes are preferable to dead weight losses

The boldest claim of this memo comes from careful economic analysis, lived experiences, and an examination of the past and present conditions for San Francisco’s housing crisis. This argument has two stages: the benefits of dynamic pricing, and the perils of keeping landlords from breaking even.

Dynamic Pricing

This section largely draws upon Mercatus Center Fellow Cody Taylor’s arguments in his article “The Case for Algorithmic Pricing: Consumer Welfare, Market Efficiency, and Policy Missteps.” In his article, Taylor begins with a case study of sorts into TicketMaster’s notorious dynamic pricing scheme that pushes welfare from the consumer to the producer egregiously through virtual lines, egregious price hikes, and general monopolistic behavior as the only major platform for ticket sales. However, in his examination of this case, he posits “asymmetric access to information” is the underlying reason for the considerable consumer loss, where “businesses that use algorithmic pricing tools know more than consumers do about the external and internal factors that impact pricing.” Taylor uses the example of higher charges for an Uber during peak rush, since there is a sharper demand to be capitalized upon. If the consumer knows that they will need an Uber at peak rush hour, they will schedule ahead to avoid the surge prices or catch a ride from a friend instead. Taylor argues that this creates ‘allocative efficiency,’ which applies directly to the real estate market in San Francisco. Consumers’ willingness and ability to pay higher prices creates welfare within the market for both consumers and producers, while surge prices also incentivize higher consumption during low-demand periods. Surprisingly, real estate markets benefit from dynamic pricing similarly to the Uber model. According to the article,

“In a study of personalized pricing on the employment platform ZipRecruiter, economists Jean-Pierre Dubé and Sanjog Misra find that, without accounting for distributive effects, consumer welfare falls by about 25 percent when pricing is personalized. However, they qualify this find­ing by stating that “this decline in total surplus comes from less than half of the consumers,” and further show that 63 percent of the consumers under personalized pricing face lower prices than in the uniform case.

Most consumers face lower prices overall when dynamic pricing is employed, which is corroborated by the data cited above. However in cases in which demand is greater than the supply, welfare for the producers and consumers can be created from RealPage’s ‘sold out guardrail’ feature, which is simply an instantaneous version of what market conditions elicit from landlords anyway. However, welfare was disproportionately siphoned away from the poorest consumers according to the preceding study, who should be protected via subsidies and personalized rent control to preserve the market benefits as much as possible.

The phenomenon of asymmetric information that producers (landlords) can exploit is actually reduced with the introduction of algorithmic pricing with access to private information across a large swath of sellers. Keeping information unclear, as the DOJ has unintentionally done by banning clearer access to better pricing, creates losses of efficiency within the market and drives prices up further. 

Landlords must profit

It is paramount that a market-optimal price be charged for rental units so that landlords can maintain a good standard of living for their tenants. In exploring the effects of too-low prices on maintenance and renter welfare,examining rent control’s effects on building maintenance in San Francisco reveals the danger of artificially deflated prices.

Since prices for maintenance, including insurance, utilities, garbage services, and taxes, rise constantly while rent prices are capped, landlords make trade-offs: they delay non-essential repairs and maintenance, reduce amenities and services, and eventually exit the rental market altogether. Supply of price-controlled units will disappear in the long term according to microeconomic theory. In San Francisco, “landlords treated by rent control reduce rental housing supplies by 15 percent” according to a study from Stanford. Reducing the supply of housing on the market raises prices further. Artificially reduced pricing not only reduces living quality for renters, but eventually raises prices across non-rent controlled units.

The rent control case shows that if landlords can’t make a living without cutting corners, prices rise anyway. The ban on RealPage creates a similar effect, since it deprives landlords of access to information to charge the most market-optimal price.

The solution

In his article, Taylor claims “the negative effects associ­ated with algorithmic pricing tools are wrongly attributed to the tools themselves, when in fact they are the result of the broader competitive context the tools are used within.” Let us now contextualize the market-wide price hikes in San Francisco as a result of RealPage’s pricing recommendations.

According to local news outlet KTVU, in January of 2026, San Francisco had the highest year-over-year rent growth across the US, with an average one-bedroom apartment renting for $3,790. According to Zillow, the average rent for a one-bedroom apartment in the US as of April 16, 2026 is $1,321. The story of how San Francisco became so expensive is relevant to the solution proposed.

San Francisco sits at the crossroads of five different pressures that keep supply low and demand high, all of which stem from the city’s high desirability and high exclusivity. First, zoning restrictions that delineate residential and commercial districts were historically designed to prioritize single-family housing to price minorities and low-income residents out of the zoned neighborhoods. The result is low-density housing across most of the city with high concentrations of high-density housing in historically redlined areas. Next, obtaining permits to develop housing takes about a year and a half on average, which meant the state was not on track to meet state-mandated housing targets before the moderate process streamlining in 2024. This comes as a result of the colloquially-known term NIMBYism (not in my backyard), a sentiment shared by the majority of San Francisco residents who don’t want affordable housing built near their residence. Next is the issue of increased demand following the tech boom: hundreds of thousands of jobs have been created around the Bay Area as a result of high tech growth out of Silicon Valley, while housing has been lagging for the previously stated reasons. The final reason is environmental pressures, which bleed into slow and arduous permit laws: the California Environmental Quality Act (CEQA) in particular significantly slows or blocks new housing development. Each of these factors depress the supply of housing. It is clear that San Francisco needs more housing to cope with its outsized demand. In a competitive market, this is the only way to reduce prices efficiently.

Taylor’s article states: “because pricing algorithms are neutral tools that can be used to both benefit and harm consumers, bans and limits on the use of pricing algorithms are the incorrect policy response,” which this policy memo echoes. Outright banning an efficient and pro-competitive pricing tool puts San Francisco at a disadvantage economically and developmentally. Economically, landlords and residents leave welfare on the table as discussed previously. Developmentally, the housing market is cut off from integrating an AI-assisted tool that may prove beneficial in more areas than one to improve market efficiency.

One such alternative use for AI algorithmic software is use it to improve administrative efficiency in generating approvals for more housing units. AI can administer screenings on new development plans at a fraction of the time and cost of human screenings for black-and-white cases like missing documents, code violations, and zoning conflicts. This significantly reduces the time spent on screening and opens up more time and availability to review plans from a wider variety of submissions. According to the San Francisco Chronicle, the median approval time for planning and building takes more than one calendar year, and this is after the process was reformed to improve wait times in January of 2024. Since the bottlenecks occur at bureaucratic tasks like routing time between offices and following prescribed intake steps, AI is well-equipped to perform these tasks and cut down wait time significantly. Next, according to the same article, San Francisco is notorious for its complex permit process for new housing developments. Implementing a new AI tool at the government and individual level that can parse through and explain various overlapping codes and regulations can further aid in promoting clarity and speed in planning and project creation. On the supply side, landlords currently sit on empty units, costing them thousands per month, without knowing the legal price they can charge with shifting rent control laws. Saving time and legal fees with a clarifying tool can reduce prices charged on the markets as a result. On the government side, immediate access to laws and legislation during the approval process significantly cuts down on legal fees and wait time during second-round reviews and approvals. 

Conclusion

This policy memo argues for the reversal of the DOJ’s ban on RealPage by arguing that the software is a neutral tool that raised prices unilaterally because of market conditions and not collusion. It argues that rather than reducing prices, banning an algorithmic tool with access to private information unintentionally creates market inefficiencies from a microeconomic perspective. In examining the evidence that AI algorithms collude, we find that the structure of the study is not applicable to a competitive real estate market, since it relies on an oligopolic model. This memo breaks down the positive effects of dynamic, personalized pricing and its applicability to San Francisco’s real estate market, and it offers a real solution to the city’s high prices.

Special thanks

• Joby Tapia, Tapia Property Solutions

• Barry Chiswick, George Washington University

• Alden Abbott, Mercatus Center

• Karan Khana, Build Fellowship

Bibliography

Bay Area Metro. “Housing Units – San Francisco.” Bay Area Census. Using 2020 U.S. Census Bureau data. Accessed April 26, 2026. https://census.bayareametro.gov/housing-units?location=san_francisco.

Calvano, Emilio, Giacomo Calzolari, Vincenzo Denicolò, and Sergio Pastorello. “Artificial Intelligence, Algorithmic Pricing, and Collusion.” American Economic Review 110, no. 10 (2020): 3267–3297. https://doi.org/10.1257/aer.20190623.

Diamond, Rebecca, Tim McQuade, and Franklin Qian. “The Effects of Rent Control Expansion on Tenants, Landlords, and Inequality: Evidence from San Francisco.” American Economic Review 109, no. 9 (2019): 3365–3394. https://doi.org/10.1257/aer.20181289.

Dubé, Jean-Pierre, and Sanjog Misra. “Personalized Pricing.” [Add journal / publisher / year if available].

Duffy v. Yardi Systems, Inc.

Federal Trade Commission and U.S. Department of Justice. Statement of Interest of the United States and the Federal Trade Commission. November 2023. https://www.justice.gov/d9/2023-11/418053.pdf.

GeeksforGeeks. “Q-Learning in Reinforcement Learning.” Last modified October 31, 2025. https://www.geeksforgeeks.org/machine-learning/q-learning-in-python/.

KTVU FOX 2. “Rent in This Bay Area City Jumped 18% in the Last Year, Data Shows.” March 24, 2026. https://www.ktvu.com/news/sf-rent-prices-zumper-data.

Leonard, Christian. “New SF Data Reveals How Long the City’s Notorious Permit Process Really Takes.” San Francisco Chronicle. May 30, 2025. https://www.sfchronicle.com/realestate/article/sf-permit-tracking-data-20349788.php.

Manne, Geoffrey A., and Dirk Auer. “The Case for Algorithmic Pricing.” Truth on the Market.

Mankiw, N. Gregory. Principles of Microeconomics. 10th ed. Boston: Cengage, 2024.

RealPage Analytics. “4th Quarter 2025 Multifamily Update.” RealPage Analytics Blog. January 8, 2026. https://www.realpage.com/analytics/4q-2025-data-update/.

SPCM College. “Bertrand Model.” PDF lecture note. Accessed April 26, 2026. https://www.spcmc.ac.in/uploads/1709645386_BERTRANDMODEL.pdf.

Taylor, Cody. “The Case for Algorithmic Pricing: Consumer Welfare, Market Efficiency, and Policy Missteps.” Mercatus Center. May 14, 2025. https://www.mercatus.org/research/policy-briefs/case-algorithmic-pricing-consumer-welfare-market-efficiency-and-policy.

In re RealPage Rental Software Antitrust Litigation, MDL No. 3071.

United States v. Greystar Management Services, LLC. Complaint filed January 7, 2025.

United States v. RealPage, Inc. Complaint filed August 23, 2024; amended January 7, 2025.

U.S. Department of Justice. “Justice Department Requires RealPage to Stop Sharing Competitively Sensitive Data and Using Algorithms that Inflate Rents.” News release, November 2025.

“San Francisco Housing Shortage.” Wikipedia. Last modified April 2026. https://en.wikipedia.org/wiki/San_Francisco_housing_shortage.

Zillow Rental Manager. “United States Rental Market Trends.” Filtered for one-bedroom apartments and apartment/condo units. Accessed April 26, 2026. https://www.zillow.com/rental-manager/market-trends/united-states/?bedrooms=1&propertyTypes=apartment-condo.