The ghost cartel — your pricing algorithm may have stopped competing without your knowledge

Amazon is fighting a very old antitrust question in a very new form. In the FTC’s suit, regulators say an internal pricing tool called Project Nessie spotted products where rivals would likely follow an Amazon price hike, raised the price, then kept it there once others matched. The agency says that produced more than $1 billion in excess profit. Amazon says the tool was shut down years ago.

But the tougher problem isn’t a company that sets out to do this. It’s the case where software gets there on its own. A 2024 Journal of Political Economy study looked at German gas stations after automated pricing software spread in 2017. When both stations in a market used it, margins rose by about 38%. When only one station adopted it, market margins didn’t budge. No meeting. No message. No explicit agreement. Just two systems learning, separately, that backing off paid better than fighting.

That gap matters because antitrust law was built for human coordination. The cleanest examples regulators can touch involve shared systems and shared data, like RealPage. The harder version is the one with no hub at all, only independently deployed algorithms watching the same public prices and settling into the same quiet truce. The source article calls that the ghost. It is much easier to miss because the dashboards can look fine. Margins look healthy. Competition looks calm. The problem may be the calm itself.

The academic record backs up the concern, even if researchers still debate how far it reaches in live markets. In a 2020 American Economic Review paper, four economists showed reinforcement-learning algorithms could learn to charge above competitive levels and punish anyone who cut price. Later that year, the authors and Wharton’s Joseph Harrington warned in Science that delegating pricing can create collusion without human awareness. A company that hands pricing to software can therefore end up with the economic result of a cartel while still having no one to blame, which is very tidy in the way only a lawsuit is tidy.

Courts and lawmakers are already trying to draw lines around the easier cases. The DOJ’s RealPage settlement, filed in November 2025, restricts certain data inputs and adds a monitor, but it does not resolve the deeper question. Cities and states have moved too, from San Francisco to New York and California. The practical lesson for boards is blunt: if prices across a category suddenly converge, “the algorithm did it” is not a defense, it’s a confession that nobody bothered to ask what the algorithm learned.