I sat through my first antitrust compliance training about a decade ago. The instructor — a former DOJ attorney who clearly missed prosecution — walked us through the classics. Do not meet with competitors to discuss pricing. Do not share rate information. Do not agree on future pricing strategies. Every cautionary tale featured the same cast: executives in a hotel conference room, handshakes that should never have happened, incriminating emails recovered during discovery.
The entire apparatus of antitrust enforcement was built around a simple assumption: collusion requires a conversation.
The rooms are empty now. The outcomes are the same.
Sherman Act Section 1 — the backbone of American antitrust enforcement — requires proof of an “agreement” between competitors . That word was chosen for a species that coordinates through language. It made sense for a century. But AI pricing algorithms now achieve coordinated outcomes through independent optimization. No communication. No agreement. No one deciding to collude. Your compliance program was designed to prevent the first kind of coordination. It has no mechanism — none — to even detect the second.
For someone paying rent or filling a car, the relevant question is whether these systems are already changing the price. The evidence increasingly suggests they are.
Pricing systems and observed outcomes
Start with the case the law could actually handle, because it illustrates what “working” looks like.
RealPage built algorithmic pricing software that aggregated non-public rental data from competing landlords and generated coordinated pricing recommendations. A digital hub-and-spoke cartel, if you want the textbook term. Consumer advocates estimated the coordination inflated rents by more than $3 billion in 2023 alone . The DOJ settled in November 2025, restricting the algorithm from using competitors’ non-public data. There was an identifiable hub, provable data sharing, and an enforceable remedy. The system worked — because the problem looked like the problems the system was built to solve.
Then there is Amazon’s Project Nessie, which occupies an interesting middle ground. This algorithm systematically tested whether competitors employed tit-for-tat pricing strategies, then raised prices to see if they would follow . No shared data. No intermediary. Just an algorithm probing the competitive landscape and exploiting what it found. Not classic collusion, but not innocent price optimization either — deliberate strategic testing of competitor responses without any shared infrastructure. If RealPage was a phone call between conspirators, Nessie was someone reading the room and acting accordingly.
Now consider the case the law genuinely cannot handle. In the German retail gasoline market, competing stations deployed adaptive pricing algorithms that observed publicly posted competitor prices and adjusted in real time. No hub. No shared private data. No intermediary. Each algorithm independently learned that maintaining higher prices was more profitable than undercutting. Researchers found that margins increased 28% in duopolistic regions . Not through coordination — through independent algorithms reaching the same conclusion about what maximized profit. There was no agreement to prosecute, no data sharing to restrict, no intermediary to regulate. The system did not fail. It simply had nothing to grab onto.
The academic work explains why this keeps happening. Calvano and colleagues demonstrated in the American Economic Review (2020) that Q-learning algorithms consistently learn supracompetitive pricing — prices above competitive levels — without any communication between them . What struck me about this research was not just the convergence on high prices — it was the sophistication. The algorithms developed punishment-cooperation cycles: when one agent deviated to a lower price, others punished it with temporary price drops before returning to the cooperative high price. One-shot experiments confirmed this was strategic, not accidental. These algorithms reinvented cartel enforcement mechanisms from scratch, without anyone teaching them to do so.
Punishment for deviators. Temporary price wars followed by return to cooperative equilibrium. One wonders whether a group of petroleum-exporting nations, meeting periodically in Vienna since 1960, would find any of this surprising.
Research on LLM-based pricing agents adds another concern. Fish, Gonczarowski, and Shorrer showed that LLM-based pricing agents reach supracompetitive prices rapidly — not after thousands of learning iterations, but quickly, from general training . The researchers also found that seemingly innocuous differences in prompt phrasing substantially changed how aggressively the algorithms priced. The mechanism is harder to detect because it emerges from general language capability, not from anyone writing “collude” into the code.
The German gasoline case connects the pricing problem to a policy intended to help buyers: the very transparency regulators fought to create became the mechanism of harm.
The Transparency Paradox
Economic theory says market transparency benefits consumers . More information leads to more competitive pricing. For decades, regulators pushed for exactly this — public price postings, comparison tools, open markets. The logic was sound: if buyers can see all the prices, sellers have to compete. This was genuinely good policy for a world of human-speed decision making.
That theory assumed human-speed processing. It did not anticipate what happens when algorithms observe and react to every competitor price change in milliseconds.
When machines can monitor every public price signal and adjust instantaneously, the theoretical consumer benefit inverts. The German gasoline market had everything a transparent market should have: public pricing, easy comparison, real-time information. The algorithms weaponized all of it . Full transparency, public data only — yet researchers documented those 28% margin increases. The thing that was supposed to help consumers became the tool that hurt them, not because transparency is bad, but because the assumptions underlying its benefits — assumptions about processing speed and strategic behavior — no longer hold.
The same transparency that helps consumers compare prices also helps pricing algorithms monitor competitors.
The regulatory response reveals the confusion. The Preventing Algorithmic Collusion Act (S. 232) only prohibits algorithms using nonpublic competitor data — it solves for RealPage but not for the German gasoline pattern. Meanwhile, states cannot agree on the problem: California’s AB-325 makes no distinction between public and non-public data, while Connecticut’s HB8002 explicitly permits public data use . The same algorithm would be legal in Hartford and illegal in Los Angeles. The OECD’s 2025 G7 report classifies algorithmic collusion into four categories, with autonomous tacit collusion — the kind that needs no shared data at all — identified as the most challenging for every jurisdiction .
Europe is no further ahead, despite its reputation for moving first on tech regulation. Article 101 TFEU — the EU’s equivalent of Section 1 — requires proof of an “agreement” or “concerted practice” between undertakings. The European Court of Justice established decades ago in Wood Pulp II (Joined Cases C-89/85 et al., 1993) that parallel pricing alone does not prove coordination — there must be some form of contact.
The closest the EU has come to testing this against algorithms was Eturas (Case C-74/14, CJEU, 2016), where Lithuanian travel agencies used a common booking platform that capped discount rates. The Court found a concerted practice — but only because a human-readable system notice informed the agencies, and they failed to distance themselves. Shared platform, identifiable message, traceable conduct. That gives investigators conduct they can trace.
The Commission’s own Competition Policy for the Digital Era report — the Crémer Report, commissioned by Vestager in 2019 — flagged autonomous algorithmic collusion as a gap in existing tools and recommended the Commission consider whether Article 101 needed updating. Six years later, neither the Digital Markets Act (Regulation 2022/1925) nor the AI Act (Regulation 2024/1689) addresses it.
As Vestager herself put it in her 2017 Bundeskartellamt address: companies cannot escape responsibility for collusion by hiding behind a computer program. The problem is that no one has yet defined what “responsibility” means when the program colludes on its own.
Consider a scenario that is hypothetical but grounded in demonstrated mechanisms: three mid-market SaaS companies each deploy an LLM-based pricing agent. Same general-purpose model. Same public market data. No shared information, no common vendor. Within weeks, all three converge on nearly identical pricing tiers — not because they communicated, but because the same model architecture reached the same conclusions about optimal pricing . A customer complains to the FTC. The investigation finds no agreement under Sherman Act Section 1, no nonpublic data under the proposed federal bill. Permitted under Connecticut law. Possibly illegal under California’s. The regulatory patchwork produces no enforcement action despite measurable consumer harm.
This is not a forecast. It is a description of mechanisms that have already been demonstrated in controlled settings, applied to a market structure that already exists.
Maybe This Is Just Markets Working
The Ninth Circuit wrote it in Gibson v. Cendyn (August 2025): parallel adoption of the same vendor’s pricing software does not establish conspiracy . Independent pricing decisions, even when informed by similar analytics, are not unlawful . This is real precedent, and it articulates something economically coherent. If algorithms independently process the same public data and reach similar prices, that might simply be market efficiency doing what market efficiency does. The price converges on the right answer. No villain required.
The regulatory system is responding, too . The DOJ brought RealPage. States passed legislation. The OECD convened G7 competition authorities. The framework is bending, not breaking.
Those arguments leave the problem of autonomous pricing on public data unresolved.
Every successful enforcement action so far has targeted one pattern: shared non-public data flowing through an identifiable hub. The harder problem — autonomous convergence on public data alone — has zero successful enforcement actions globally. Zero. The regulatory patchwork itself, where the same algorithm is legal in one state and illegal in the next, is evidence that the system has not yet defined the problem it is trying to solve.
And Calvano’s evidence is specific about what “convergence” actually looks like. The algorithms do not just reach competitive equilibrium — they systematically overshoot it to supracompetitive levels and maintain those levels through learned punishment strategies . What the German gasoline market produced at scale was not efficient equilibrium but 28% margin increases. What the rental market produced was $3 billion in excess cost to tenants. Efficient equilibrium is the theory. Documented consumer harm is the data.
The Four Gaps
When I map this out, I find four gaps in how the institutions recognize and respond to harm. Each concerns a familiar task — proving an agreement, establishing intent, detecting the behavior, or assigning responsibility — that becomes harder when an algorithm sets the price.
The Agreement Gap. Antitrust law requires proof that competitors agreed to coordinate . Algorithms converge on supracompetitive prices through independent optimization — same outcome, no agreement. The legal threshold was written for actors that coordinate through language. The new actors do not use language, and they do not need to.
The Intent Gap. Enforcement requires demonstrating anti-competitive intent. But AI pricing agents learn collusive strategies autonomously — nobody programmed them to collude. The algorithm discovered the strategy through optimization. Harmful outcomes exist. Intentional wrongdoing, in any meaningful sense, does not.
The Detection Gap. Traditional cartels are inherently unstable. Participants defect. Whistle-blowers emerge. Someone gets nervous and calls a lawyer. Algorithmic collusion has none of these failure modes. The punishment-cooperation cycles Calvano documented maintain high prices indefinitely . There are no meetings to intercept, no communications to subpoena, no conscience to appeal to. The cartel that cannot be betrayed is the cartel that cannot be caught.
The Accountability Gap. Compliance frameworks require a responsible human decision-maker. AI pricing distributes decisions across the developer who set the objective function, the data team who provided training data, the operations team who deployed the system, and the algorithm that learned the strategy. Nobody is accountable because everybody is partially responsible — which, in practice, means nobody is responsible at all. I explored this problem from a different angle in The Phase Collapse — when AI compresses project phases, accountability becomes the hard boundary that AI cannot cross. The same dynamic applies here: the more distributed the decision-making, the harder it becomes to assign responsibility.
These four gaps do not exist in isolation. They reinforce each other. The agreement gap means there is nothing to detect. The intent gap means there is no one to hold accountable. The detection gap means the behavior persists. The accountability gap means no one is motivated to look.
Are We Ready?
The regulatory infrastructure we rely on took decades to build. It is sophisticated, tested, and well-resourced. Its methods were developed for people who needed to communicate, reach an agreement, and intend to coordinate their prices.
Now a different kind of competitor operates in the same markets. Human-deployed but exceeding human oversight, these algorithms do not need rooms, agreements, or intent to produce outcomes indistinguishable from collusion. Seventy-two percent of enterprises have already deployed AI agents without formal oversight frameworks. Sixty-two percent experienced at least one agent-driven operational incident in the past twelve months . The governance gap is not a forecast. It is an audit finding.
This is ultimately a question about what we build safety into. As I argued in Trust Is Not an Instruction, relying on behavioral compliance — telling systems to behave well — is fundamentally different from structural safety, where the architecture itself prevents harm. The antitrust framework is behavioral: it assumes actors will either follow the rules or get caught breaking them. Algorithmic pricing breaks neither rule nor law. The structure needs to change.
I do not have a clean answer for this. I am not sure anyone does. The standard regulatory playbook — identify the bad actor, prove the conspiracy, impose the remedy — assumes a world where coordination requires communication. We are entering a world where it does not. And the people best positioned to navigate this transition are not the technologists building the algorithms or the lawyers interpreting the statutes — they are the experienced professionals who understand both the markets these algorithms operate in and the judgment required to verify whether algorithmic outputs serve or subvert the public interest.
The research describes algorithms reaching coordinated prices without an agreement, while the legal framework still depends on proving one. For the customer paying the higher price, that distinction offers little comfort. Whether our institutions can close that gap will shape the markets people rely on.
References
“DOJ Settles Algorithmic Price-Fixing Case Against RealPage.” Wilson Sonsini Goodrich & Rosati, 2025. https://www.wsgr.com/en/insights/doj-settles-its-algorithmic-price-fixing-case-against-realpage.html
Calvano, E., Calzolari, G., Denicolo, V., & Pastorello, S. “Artificial Intelligence, Algorithmic Pricing, and Collusion.” American Economic Review, 110(10), 2020. https://www.aeaweb.org/articles?id=10.1257/aer.20190623
Fish, S., Gonczarowski, Y. A., & Shorrer, R. I. “Algorithmic Collusion by Large Language Models.” arXiv:2404.00806, 2024; presented at AEA 2025. https://arxiv.org/abs/2404.00806
“Ninth Circuit Clarifies Antitrust Implications of Algorithmic Pricing (Gibson v. Cendyn Group).” Arnold & Porter, August 2025. https://www.arnoldporter.com/en/perspectives/advisories/2025/08/antitrust-implications-of-algorithmic-pricing
Assad, S., Clark, R., Ershov, D., & Xu, L. “Algorithmic Pricing and Competition: Empirical Evidence from the German Retail Gasoline Market.” Referenced via: https://medium.com/@milesk_33/the-silent-failures-when-ai-agents-break-without-alerts-23a050488b16
“Algorithmic Collusion: Reviving Section 5 of the FTC Act.” Columbia Law Review, 2025. https://columbialawreview.org/content/algorithmic-collusion-reviving-section-5-of-the-ftc-act/
Preventing Algorithmic Collusion Act of 2025, S. 232, 119th Congress. https://www.congress.gov/bill/119th-congress/senate-bill/232/text
“Algorithmic Pricing Bans Go Coast to Coast: California, New York, Connecticut.” Arnold & Porter, October 2025. https://www.arnoldporter.com/en/perspectives/advisories/2025/10/algorithmic-pricing-bans-go-coast-to-coast
“Algorithmic Pricing and Competition in G7 Jurisdictions.” OECD Competition Committee, 2025. https://www.oecd.org/en/publications/algorithmic-pricing-and-competition-in-g7-jurisdictions_f36dacf8-en.html
“Algorithmic Pricing: Understanding the FTC’s Case Against Amazon.” Carnegie Mellon University, October 2023. https://www.cmu.edu/news/stories/archives/2023/october/algorithmic-pricing-understanding-the-ftcs-case-against-amazon
“The Agentic Governance Collapse.” AIGN Global, 2026. https://aign.global/ai-governance-insights/aign-global/the-agentic-governance-collapse/