The Invisible Handshake

Classic competition law was built around a clear premise: cartels require coordination. People in a room. An agreement. Intent.

The premise is becoming obsolete.

When multiple companies use variants of the same foundational AI models to optimise their pricing strategies, those models may independently converge on equilibria that look — from the outside — remarkably like collusion.

No communication.

No agreement.

Just algorithms that have each concluded, separately, that competition is inefficient.

This page tracks the legal and technical dimensions of algorithmic tacit collusion: the enforcement record, the gaps in competition law doctrine, and the question of what "intent" means when the relevant actor is a system optimising an objective function.

The Clause has read Article 101 with some care and wishes to point out that it prohibits agreements, decisions, and concerted practices. It is not aware of any provision addressing systems that never needed to agree on anything.

The architecture of the argument

The doctrinal hole. Article 101 TFEU and Sherman §1 both require a meeting of minds — an agreement, a decision, or a concerted practice involving some concurrence of wills. Parallel conduct alone has never been enough; the case law is emphatic that firms may lawfully observe competitors and adapt intelligently. That safe harbour was drawn around a factual assumption: that independent adaptation is genuinely independent, arrived at by separate firms reasoning separately. The assumption is now empirically shaky. Firms adapting through variants of the same foundation model, trained on overlapping data, are not reasoning separately in any sense the doctrine anticipated — they are running correlated inference on shared priors. The conduct is parallel. The reasoning is nearly identical. And no one agreed to anything, which under current doctrine means no one is liable.

Enforcement went around the question, not through it. RealPage is the most advanced test of algorithmic pricing anywhere, and its most instructive feature is what it declined to decide. The DOJ's proposed settlement, announced in late 2025, carried no financial penalty and no finding of wrongdoing. The remedy was structural and entirely about data: RealPage may train only on historical data at least twelve months old, must cease runtime use of unaffiliated non-public data, and may not report pricing information at a granularity finer than state level. Notice what this is. Rather than answer whether an algorithm ingesting rivals' non-public data constitutes a concerted practice, the settlement removed the data and left the doctrine untouched. It is an effective remedy and a doctrinal evasion at the same time, and it establishes the template: where intent cannot be located, regulators reach for the inputs instead.

The same evasion, in consumer protection. The pattern repeats one layer down, where the convergence is not between firms on a market price but within one firm on your price. Maryland banned AI surveillance pricing in grocery stores outright — a structural limit on what may be collected and used, not a rule about how to use it properly. The FTC, in the same period, rejected Terms and Conditions as a valid basis for the practice. Two regulators, two doctrines, one instinct: when the mental element is unprovable, legislate the data. That instinct is now visible in antitrust and consumer protection simultaneously, which suggests it is not a quirk of either field but a response to a shared problem. (Read the analysis →)

Monoculture is the technical premise of tacit collusion. This site tracks architectural convergence as a systemic-risk question — shared models produce shared blind spots, and everyone fails in the same way at the same moment. Competition law is looking at the identical fact through a different lens and mostly not noticing. If enterprise deployment concentrates on a small number of foundation models, then "independent conduct" and "shared architecture" describe the same market simultaneously, and one of those two descriptions has legal consequences the other does not. The systemic-risk framing and the competition framing are not analogies for one another. They are the same premise, sorted into two bodies of law that do not currently talk to each other. (The shared-architecture argument →)

Convergence does not require the model to be good. A tempting objection holds that convergence is a transitional problem — as models improve and differentiate, correlated outputs will disperse. The evidence points the other way. Hallucinated package names repeat across independent queries with measurable regularity, because the wrong answer is a statistical property of the model rather than a random draw; attackers can profile that distribution and pre-register the names a model is likely to invent. If models converge reliably on the same false answer, they converge at least as reliably on the same profitable one. Correlated error and correlated pricing are the same phenomenon wearing different clothes, and the first is now documented well enough to be exploited commercially. (Constraint decay and shared failure modes →)

There is no one to cross-examine here either. Cartel enforcement runs on evidence of the mental element: the meeting, the message, the leniency applicant who describes the room. Strip out the human deliberation and the evidentiary apparatus has nothing to seize. The problem is structurally identical to the one criminal evidence law now faces with machine-generated material — a process cannot be impeached, and there is no witness whose account can be tested. Competition authorities have not yet had to confront this as directly as criminal courts have, but the shape of the difficulty is already visible, and it is the reason the RealPage remedy addressed inputs rather than conduct. (The evidentiary version →) (Where the responsibility goes →)

Concentration as a competition question the AI Act does not ask. When two companies' compute agreements define what "frontier" means, their architectural choices propagate into every downstream deployment — including every downstream pricing system. That is a market structure fact with competition implications, and neither the AI Act nor current merger practice treats shared architectural dependency as a structural concern. The Act regulates products for safety. The question of what happens when the whole market runs on two products is not in its scope, and competition law has not yet claimed it either. (Read the analysis →)

Counterarguments and open questions

The strongest objection is that this problem is overstated because the doctrine already stretches further than critics assume. A concerted practice does not require a formal agreement; the case law reaches hub-and-spoke arrangements, and a common algorithmic intermediary can serve as the hub without the spokes ever communicating. On this reading RealPage was not a doctrinal failure at all — the theory was available, and the parties settled precisely because it might have worked. This is a serious argument and probably correct as far as it goes. But it depends on there being a hub: a single vendor supplying the algorithm to competing firms. The harder case, and the one this page tracks, is convergence with no vendor in common at all — separate firms, separate deployments, the same underlying foundation model, and no intermediary to designate as the centre of the wheel. Hub-and-spoke needs a hub. Architectural monoculture does not supply one.

A second objection: tacit collusion is not new, and the law has always tolerated it. Oligopolists have watched each other's prices for as long as there have been oligopolies, and the deliberate decision not to prohibit conscious parallelism reflects a judgment that punishing rational adaptation would chill competition more than it protects it. Algorithms make the adaptation faster and more precise, but speed is a difference of degree. The response is that degree matters when it crosses a threshold: tacit collusion classically required a market structure — few firms, transparent prices, homogeneous products — that limited its reach. Correlated inference does not require that structure. It can produce the same outcome in fragmented markets with many participants, which is precisely where the law assumed the problem could not arise.

A third objection: the remedy question is easier than the liability question, so the doctrinal gap may not matter much. If data limits work — and the RealPage settlement suggests they can — then regulators can constrain the mechanism without ever resolving whether it is an agreement. This is pragmatic and it is also how both antitrust and consumer protection are in fact proceeding. The open question is whether a regime built entirely on input restrictions is stable. Restrictions on data granularity and recency are precise, technical, and easy to erode incrementally through the ordinary process of vendors asking for adjustments; a liability rule, by contrast, is a general standard that does not require the regulator to win the same argument repeatedly.

The deepest open question is whether "intent" can survive as the organising concept of competition law. Every mechanism on this page points the same way: the conduct is observable, the harm is measurable, and the mental element is absent or unprovable. Three responses are available. Redefine agreement to reach correlated conduct arising from shared architecture, which risks capturing genuinely independent behaviour. Abandon the mental element for algorithmic pricing and move to effects-based liability, which competition law has resisted for good reasons. Or regulate the inputs and stop asking about minds at all — the path both RealPage and Maryland actually took. The third is working, for now, and it is the least examined, because it does not look like a doctrinal choice. It looks like a settlement term.

The Clause, characteristically, has no preference among the three. It notes only that the first two require someone to prove a state of mind, and the third requires someone to keep amending a schedule of permitted data fields. It has views about which of those is easier to attend to, year after year, and which is easier to quietly let slip.