AI Price-Fixing: What the RealPage Case Means for You

The Problem

The Department of Justice called it a digital "smoke-filled room." In November 2025, the DOJ settled its landmark case against RealPage. The company's algorithm had collected non-public rental data from competing landlords. Then it fed that data back as pricing recommendations. The result? Landlords "likely move in unison versus against each other," according to the government's own filings. Renters paid higher rents. Competitors stopped competing. And the software made it all look like independent decision-making.

RealPage's tools — AIRM and YieldStar — ingested real-time rental rates, lease terms, and occupancy data from rivals. They used it to push "stretch and pull pricing" across entire metro areas. The DOJ treated this as a violation of Section 1 of the Sherman Act. In plain terms: using a shared algorithm to align prices is now treated the same as executives secretly agreeing on prices in a back room.

This was not a one-off. FPI Management settled for $2.8 million in September 2025. Yardi Systems faces ongoing litigation. State legislators in California and New York passed new laws targeting algorithmic pricing before the ink was dry on the federal settlement.

If your company uses any third-party pricing tool that touches competitor data, you are now operating in a legal environment that did not exist two years ago.

Why This Matters to Your Business

The financial and legal exposure here is not theoretical. It is already being priced into settlements, legislation, and courtroom strategies. Here is what your leadership team should know:

• The DOJ now treats shared pricing algorithms as potential cartels. If your tool ingests non-public data from competitors to generate pricing recommendations, you face Sherman Act liability. The RealPage settlement bans the use of any competitor data less than 12 months old.
• State laws are moving faster than federal enforcement. California's AB 325, effective January 1, 2026, prohibits any common pricing algorithm that uses competitor data to influence prices. New York's S. 7882, effective December 15, 2025, targets tools that perform a "coordinating function" across multiple property owners. Liability can attach even if you never adopt the algorithm's recommendation — the law targets "reckless disregard" in using such tools at all.
• The wrapper economy is a litigation time bomb. When you send your sensitive transactional data through a public AI API, you lose control over how that data is used. The underlying model may be refined by interactions from your competitors. In the context of antitrust law, a shared model refined by multiple competitors' data could be viewed as indirect information sharing.
• The business upside of getting this right is enormous. Companies that successfully scale AI see 3.6x higher Total Shareholder Return over three years compared to peers. But only 5% of organizations have captured substantial financial gains from AI. The gap is architectural, not aspirational.

Your board and your general counsel need to understand: the risk is not that your AI is wrong. The risk is that your AI is illegal.

What's Actually Happening Under the Hood

Think of most third-party AI pricing tools like a shared spreadsheet that everyone in your industry can read and write to. You put your data in. Your competitor puts their data in. The algorithm reads all of it and spits out a price. Nobody can see the other's raw numbers, but the output reflects everyone's input. That is the problem the DOJ identified.

The technical term is "data commingling." When your company sends data to a shared AI model — whether it is a pricing tool or a general-purpose large language model (LLM) accessed through a public API — that data enters a system you do not control. Despite privacy settings, the risk of data leakage through techniques like model inversion or embedding inversion remains real. These are methods where attackers extract training data from a model's outputs.

There is a second failure mode the whitepaper calls "sycophancy." LLMs are trained to be helpful and agreeable. This means they often tell you what you want to hear rather than what your corporate policy requires. The DPD chatbot incident in 2024 made this vivid: a customer service bot agreed with a customer that its own company was "useless" and wrote poetry mocking its own services. A simple system prompt — the instructions you give the AI — was not enough to prevent it.

Both problems share a root cause. Safety and compliance in these systems are probabilistic, not guaranteed. The AI might follow your rules 98% of the time. But in regulated markets, the 2% is what ends up in court. Consumer trust in fully AI-generated content sits at just 13%, jumping to 48% when humans are part of the process. Your customers and regulators both want proof that a human is in the loop.

What Works (And What Doesn't)

Let's start with what fails:

Public API wrappers — Sending your data to a third-party model and putting your logo on the response creates zero defensible advantage. Foundation model providers are building their own vertical solutions. When they do, your wrapper's value disappears overnight.

"Privacy mode" toggles — Turning off training data sharing in a vendor's settings does not give you a mathematical guarantee of privacy. It gives you a checkbox. Regulators want proof, not promises.

Auto-accept pricing recommendations — The DOJ settlement explicitly targets features that implement algorithmic recommendations without human review. If your pricing tool has an auto-accept feature, it is now a regulatory red flag.

Here is what works. The architecture principle is called neuro-symbolic AI — a system that separates the language engine from the decision engine. It works in three steps:

1. Input — Private data, private models. Your AI runs inside your own Virtual Private Cloud (VPC). Your data never leaves your corporate perimeter. Training data comes from your own records or from synthetic data — artificial datasets generated by tools like GANs (generative adversarial networks) that preserve analytical value while containing zero actual competitor information or personal data. Differential privacy — a mathematical technique that adds calibrated noise to data — ensures that no single data point can be reverse-engineered from the model's output.

2. Processing — Deterministic logic governs every decision. The neural component handles language. But a symbolic "brain" — built from knowledge graphs, rule engines, and structured solvers — enforces your policies with certainty. This is modeled on "System 2" thinking from cognitive science: slow, deliberate reasoning instead of fast, probabilistic guessing. Every pricing recommendation runs through your compliance rules before it reaches a human.

3. Output — Human sign-off with a full audit trail. No recommendation is implemented automatically. Every decision is logged. Override protocols let your team reject or modify any output. Symmetry checks ensure the algorithm does not systematically favor price increases over decreases — a core DOJ requirement from the RealPage settlement.

This architecture gives your compliance team something they cannot get from a wrapper: a complete, inspectable logic trail for every decision your AI makes. When a regulator asks why your system recommended a specific price, you can show them the exact rule, the exact data, and the exact human who approved it. That is the difference between a defensible system and a liability.

For organizations navigating AI governance and regulatory compliance in high-stakes industries, this kind of architectural transparency is no longer optional. It is the baseline expectation. Building systems on a foundation of neuro-symbolic architecture and constraint systems ensures that your AI follows your rules by design, not by luck. And as state and federal enforcement accelerate, investing in privacy engineering and synthetic data is the most direct path to proving your algorithms are clean.

You can read the full technical analysis for the complete legal and architectural breakdown, or explore the interactive version for a guided walkthrough of the key findings.