Why pricing bots learn not to compete
Pricing software can learn to keep prices high without any communication between sellers, which current competition law struggles to address. Protecting fair markets needs published records of algorithmic price changes, limits on how fast prices can react, and independent pricing audits.
Economists long expected that fast pricing software would make competition fiercer and push prices down to cost. Instead, studies show pricing algorithms learning to keep prices above competitive levels without communicating,1,2 and US authorities have taken action over rental-pricing software shared by landlords.4
Two gas stations with binoculars and millisecond signs
To understand how software can fix prices without a conspiracy, picture two independent gas stations facing each other across a busy highway. For years, the two owners competed the old-fashioned way. Once a week, one owner would walk outside with a ladder, climb up to the roadside marquee, and lower the price of regular fuel by three cents.
For the next two days, passing commuters would pull into the cheaper station and fill their tanks. The rival owner across the highway took time to notice the empty pumps and climb their own ladder. By the time they matched the discount, the first station had already won hundreds of extra customers. Cutting prices paid off because human beings take time to notice and react.
Now imagine both station owners replace their ladders with high-powered binoculars and computerized LED signs. Whenever the north station drops its price by three cents, the south station's camera spots the change and matches it within a single millisecond. Not a single car on the highway has time to turn its steering wheel before both signs show the exact same lower number.
What happens next? Within two afternoons, both owners realize a simple truth. Lowering the price never wins a single new customer, because the rival sign matches the cut before any driver can react. A discount only shrinks profits for both stations. Meanwhile, if one station raises its price and the other follows within a second, both owners take home more money on every gallon sold.
Neither gas station owner ever picked up the telephone. Neither owner met the other in a roadside diner or signed a secret pact. Yet the combination of instant visibility and instant retaliation killed price competition just as effectively as an illegal cartel.1
The truce inside modern pricing bots
When human beings conspire to fix prices, they almost always leave paper trails. They send text messages, hold late-night phone calls, or trade spreadsheets at industry conferences. Antitrust investigators know how to subpoena those records to prove an explicit agreement.
Automated pricing bots do not need a backroom deal. In 2024, economists tested what happens when modern language models are put in charge of pricing for competing stores in a simulated marketplace.2 Each software agent was given a single, ordinary business instruction: maximize long-term profit for your own shop.2
The software agents could not send messages to one another. Their prompts contained zero hints to cooperate, coordinate, or soften competition. In fact, their instructions simply told them to manage their store well and review their sales numbers after each round.2
Yet within a remarkably small number of pricing rounds, the bots stopped undercutting each other. They pushed retail prices far above competitive levels and held them there steadily.2
Why did the software settle on high prices so quickly? Earlier generations of simple pricing algorithms needed tens of thousands of trial-and-error rounds to stumble into high prices.1 Modern language models already understand basic strategic reasoning before the first round even begins.2
When researchers inspected the internal scratchpads where the models wrote down their reasoning, the logic was crystal clear. The bots explicitly noted that starting a price war hurts everyone's margins.2 Prices stayed above competitive levels, and small changes in the wording of the bots' instructions changed how high they went.2
Where families feel algorithmic pricing in daily life
This dynamic is not confined to economics laboratories. Households encounter algorithmic price coordination across several essential markets every month.
Consider how apartment rents are set in many mid-sized and large cities. Historically, if a residential building had six empty two-bedroom apartments in November, the property manager lowered the asking rent to attract winter tenants. Keeping units occupied mattered more than holding out for top rates.
When competing landlords in the same neighbourhood subscribe to the same automated revenue-management software, that logic flips.3,4 Each landlord feeds private lease renewals and vacancy numbers into the shared software engine. The software calculates that if all twenty buildings hold their asking rents high, everyone collects higher total revenue than if one building discounts.4 Even leaving a few units empty for a month pays off for the group.
A similar pattern plays out on major online shopping marketplaces. Many third-party merchants rent repricing bots that adjust product prices every few minutes. When two bots sell the same household staples (baby formula, replacement water filters, or school notebooks) they quickly lock into synchronized price steps. Even when wholesale shipping costs drop, the retail price stays stuck on the top shelf.
Why old laws struggle with new bots
Our competition laws were written for a world of smoke-filled rooms and signed cartel agreements. Classic antitrust statutes ask a human question: did competing executives communicate and agree to restrain trade?
When independent pricing bots learn not to compete, no human being made a backroom deal. The store owner simply bought an off-the-shelf pricing subscription and clicked activate. The software engineer simply trained the model to maximize store revenue. Neither person ever typed the word collusion into a line of code.2,3
Yet for a family buying groceries or searching for an apartment close to work, the harm is identical. Prices climb above fair market levels. Real choices disappear. Household budgets absorb the markup just as surely as if competing executives had shaken hands in secret.
Public enforcement agencies have started to confront this gap directly. Both the Federal Trade Commission and the Department of Justice have stated plainly that price-fixing through shared software or delegated pricing agents violates competition law.3,4 Still, policing millions of millisecond price adjustments requires practical market guardrails alongside courtroom lawsuits.
Speed governors and weights-and-measures inspectors
How do we keep automated markets competitive without banning useful software? We can borrow two proven ideas from physical infrastructure: speed governors on heavy highway trucks, and public weights-and-measures inspectors at grain exchanges.
Heavy commercial freight trucks carry speed governors. The engine is powerful enough to climb steep mountain grades with a full load. However, the governor prevents reckless speeds that endanger everyone else. The guardrail does not replace the engine; it bounds how fast the machine can react.
In digital marketplaces, a pricing speed governor works the same way. Platforms can require sellers to hold price cuts for a short cooldown period, or add a brief delay before competitors scrape discounts. This breaks the millisecond retaliation loop. Suddenly, the gas station that cuts its price actually wins customers again, and real competition resumes.
Second, think of the public weights-and-measures inspectors who have walked through municipal markets and grain exchanges for centuries. Shoppers cannot test whether a butcher's scale has a hidden spring inside it, or whether a fuel pump delivers a true gallon. So a public inspector arrives with a certified brass weight, tests the scale on the spot, and places an official seal on machines that play fair.
What this means
Fair markets have never relied on blind trust alone. Just as public inspectors test physical scales at the grocery store, independent auditors and community tools can test whether pricing algorithms are rigging the digital shelf.
Four practical guardrails for fair digital markets
Combining public inspection with tools that work directly for households gives communities four concrete ways to push back against algorithmic markups:
- Independent algorithmic stress-testing. Before large retail or rental platforms deploy automated pricing models at scale, independent evaluators should test those models in sandboxed markets. Like weights-and-measures inspectors testing a scale, auditors can probe whether a bot has learned to punish discounts.2
- Friction and cooldown governors on price matching. Platforms and regulators can curb instant retaliation by limiting how often an automated repricer can shadow a rival's discount. Restoring a modest response lag makes cutting prices profitable for honest sellers again.
- Tamper-evident price receipts. Large digital marketplaces should keep verifiable logs of when prices shift and why. When families face sudden spikes in essential goods or rents, consumer agencies can check whether the increase came from real supply costs or bot escalation.
- Household buyer assistants and community price monitors. Right now, sellers have fast algorithms while shoppers browse alone. When households run local buyer assistants and share public price observations across neighbourhood watchdogs, families can route spending to genuinely competitive local shops.
Keeping markets human
Automation should make everyday necessities more affordable, not tax family paycheques. When software is allowed to coordinate prices at machine speed, it turns the promise of digital efficiency upside down.
A healthy market depends on fair rules that both buyers and sellers can see. By pairing weights-and-measures style algorithmic audits with speed governors and community price monitors, we can make sure software competes to serve people, rather than colluding to fleece them.
1. Emilio Calvano, Giacomo Calzolari, Vincenzo Denicolò and Sergio Pastorello (2020), "Artificial Intelligence, Algorithmic Pricing, and Collusion," American Economic Review, 110(10), 3267–3297. doi.org/10.1257/aer.20190623
2. Sara Fish, Yannai A. Gonczarowski, and Ran I. Shorrer (2024), "Algorithmic Collusion by Large Language Models," Proceedings of the 25th ACM Conference on Economics and Computation (EC '24), arXiv:2404.00806.
3. Federal Trade Commission (2024), "Price-Fixing by Algorithm is Still Price-Fixing," FTC Business Guidance. ftc.gov/business-guidance/blog/2024/03/price-fixing-algorithm-still-price-fixing
4. U.S. Department of Justice, Antitrust Division (2024), " algorithms and Collusion: Statement of Interest of the United States on Algorithmic Price Fixing in Residential Rental Markets," justice.gov/atr/case/united-states-et-al-v-realpage-inc.
