The Enforcement Gap

Competition law does not prohibit high prices. This is the single most important thing to understand about the legal position, and it surprises most people.

The prohibition targets agreement. Section 1 of the US Sherman Act reaches contracts, combinations, and conspiracies in restraint of trade. Article 101 of the Treaty on the Functioning of the European Union reaches agreements between undertakings and concerted practices. Both require a meeting of minds between competitors.

Charging a high price alone is legal. Charging the same price as a competitor is legal. Watching a rival's prices and matching them is legal. What is illegal is agreeing with the rival to do so.

The reason is practical. In a concentrated market, firms can reach high prices simply by each recognising their interdependence, without any communication. Economists call this conscious parallelism or tacit collusion. It is economically harmful and, in most jurisdictions, lawful, because there is no agreement to prohibit and no conduct a court could order a firm to stop.

Algorithmic collusion lands directly in that pre-existing hole.

Bichler, Durmann, and Oberlechner put the problem in one sentence: even if a competition authority were to identify a potential case of tacit collusion, the current state of the law could make such practice irreproachable in the absence of explicit communication or contact among the companies using such autonomous algorithms.

Irreproachable is the operative word. Not hard to prove. Not requiring new evidence. Potentially lawful, because the conduct element the statute requires is genuinely absent.

The difficulty compounds when you ask who would be liable if it were unlawful. The firm did not instruct the software to collude. The vendor sold a profit-maximising tool. The developers implemented a standard algorithm from a textbook. Intent, which most enforcement regimes rely on somewhere, is distributed until it disappears.

And the usual remedy fails too. Cartel enforcement works partly because participants can defect and report each other for leniency. There is nobody to defect here, and no agreement to confess to.

So the gap is not evidential. It is definitional, and it cannot be closed by better investigation.

Faced with that gap, regulators have concentrated on cases that do not require crossing it, and there is a clear pattern in which ones work.

The UK poster case noted by Bichler and colleagues involved online retailers using simple pricing algorithms as part of a horizontal cartel to coordinate prices on Amazon. There was a cartel; software executed it. Conventional law applies straightforwardly.

The European Commission's 2023 guidelines stipulate that explicit agreements to use the same pricing algorithm violate Article 101. Again, the violation is the agreement, not the algorithm.

The generalisation is worth stating as a rule. Enforcement succeeds where there is a human agreement somewhere in the picture, or a hub-and-spoke structure in which competitors do not talk to each other but each shares information with a common intermediary that feeds back into everyone's prices.

That second structure is not hypothetical, and it produced the most consequential case in this area so far.

On 23 August 2024, the US Department of Justice filed a civil antitrust suit against RealPage, a revenue management software company, in the Middle District of North Carolina, joined by the attorneys general of North Carolina, California, Colorado, Connecticut, Minnesota, Oregon, Tennessee, and Washington.

The allegation, brought under Section 1 of the Sherman Act, was that RealPage used nonpublic, competitively sensitive information from competing landlords in generating its pricing recommendations, and discouraged independent pricing decisions, thereby facilitating price alignment.

Note the shape of the theory. It does not depend on proving that an algorithm learned to collude. The alleged mechanism is the flow of confidential competitor data through a common intermediary, which is a recognisable hub-and-spoke arrangement that existing doctrine can reach.

On 24 November 2025 the DOJ filed a proposed settlement. RealPage did not admit liability and agreed to no fine, but accepted conduct restrictions for a seven-year term, subject to court approval.

The terms are the interesting part, because they read as a template.

The proposed settlement imposes detailed guardrails on RealPage's rent-pricing tools: how they may use competing landlords' nonpublic data, how pricing features are designed, and how competitively sensitive information may be collected and shared.

One restriction is unusually concrete and worth understanding. RealPage is forbidden from providing rental pricing data to its subscribers at a granularity narrower than state-wide.

The reasoning is that competitive sensitivity is a function of resolution. Knowing average rents across an entire state is market intelligence: it cannot tell you what the building across the street is charging. Knowing average rents for your immediate submarket, where you may face three competitors, is close to knowing their prices. Same category of data, entirely different competitive effect.

So the settlement does not ban the software, the data, or the recommendations. It regulates the resolution at which competitor information flows.

That is a genuinely new regulatory instrument, and it suggests where this area is heading: not toward banning algorithmic pricing, but toward constraining what data may enter it and at what granularity.

Three structures, three legal outcomes. The first two are reachable under existing doctrine because a human agreement or a data intermediary exists. The third, independent algorithms converging without any connection between firms, is the one that falls outside, and no enforcement action has closed it.

The policy conversation is active across jurisdictions. Bichler and colleagues note that the OECD and multiple countries including France, Germany, Denmark, Japan, Norway, and Sweden have published policy papers, and that the US Preventing Algorithmic Collusion Act of 2024 emphasises transparency.

Four categories of proposal recur, and they trade off differently.

Transparency requirements, making algorithmic decision-making visible to regulators. Useful for investigation, but a learned policy is not self-explanatory even when fully disclosed, which is why the interpretability problem is not merely technical.

Design constraints, guidelines discouraging collusive strategies. Klein's proposal fits here: require simultaneous rather than sequential price updates, and prohibit conditioning on price history. Precise, but it forbids ordinary responsiveness too.

Real-time monitoring, detection algorithms identifying collusive pricing patterns. Klein's cycling result is the warning: naive detectors looking for stable high prices miss the cycling case.

Regulatory audits, compliance testing requiring firms to disclose pricing algorithms.

Every one of these imposes costs on lawful, pro-competitive uses. Klein's own caution applies to all of them: pricing algorithms have many pro-competitive effects, and prohibiting their use would be excessive.

For anyone deploying pricing software, the RealPage settlement and the case law point to a workable set of distinctions.

The organising question is not whether you use an algorithm. It is where its inputs come from and whether you remain free to deviate.

Two gotchas. First, vendor assurances are not a defence: the DOJ theory targeted the data flow through the vendor, so a vendor's compliance claim does not resolve your exposure. Second, no admission is not vindication: RealPage settled without admitting liability and without a fine, yet accepted seven years of conduct restrictions. Read the restrictions, not the absence of a penalty, as the signal about what regulators consider problematic.

Four claims, stated at the strength the evidence supports.

The mechanism is demonstrated, in simulation. Independent Q-learning agents in a standard oligopoly model reach supra-competitive prices sustained by finite punishment and gradual return to cooperation, with no communication and nobody programming the strategy.

The generalisation is genuinely contested. Large exploration parameters eliminate it, transfer across environments often breaks it, convergence conditions are poorly understood, and there is little evidence that deployed pricing software uses reinforcement learning at all.

The field evidence shows an outcome, not a mechanism. German fuel margins rose 9 percent with adoption in non-monopoly markets and 28 percent in duopolies where both stations adopted, a pattern consistent with softened competition, but nobody observed the algorithms.

The legal gap is definitional. Competition law requires agreement, tacit coordination without communication is largely lawful, and enforcement to date has succeeded only where a human agreement or a data intermediary existed.

The useful mental model: this is not a settled harm awaiting enforcement, nor a phantom. It is a demonstrated possibility whose real-world prevalence is unknown, sitting in a legal category built for a different problem.