If everyone is using AI for pricing, who wins?

The question in the title has an asymmetrical answer: pricing algorithms transfer wealth from consumers to companies, with efficiency gains that do not compensate for that redistribution.

Imagine you arrive at your local gas station. The price of gasoline is $2.45 per liter. Half an hour later, your neighbor arrives at the same station and pays $2.52. There was no change in the cost of fuel. There were no discussions between the gas stations in the area. Only an invisible and silent algorithm decided to raise the price.

Humans used to set prices. Now, algorithms do it faster, smarter, learning from millions of transactions per second. We told ourselves this would make markets more efficient, more competitive. But when every company’s algorithm converges on the same “perfect” price… do you really think that price is perfect for you?

That’s the tension at the heart of Artificial Intelligence, Algorithmic Pricing and Collusion (Calvano, Calzolari, Denicolò, Pastorello), they pit Q-learning (reinforcement learning) algorithms against each other in a workhorse model of repeated oligopolistic competition. In simpler terms, they create a simulated market where a few companies repeatedly compete by setting prices, and they let the algorithms learn through trial and error, focusing solely on profits, without instructions to collude, without communication, and without prior knowledge of the environment.

The result was disturbing: The algorithms consistently learned to set prices above the competitive equilibrium, the level predicted by classical economic theory when firms compete independently. This wasn’t a mistake or an isolated case. It happened with two, three, or four competitors. It happened even when the firms had different costs and under conditions of uncertainty.

How did they do it?

Q-learning is a form of reinforcement learning: it doesn’t attempt to predict demand like traditional forecasting, but rather to learn which action (which price) maximizes cumulative reward (profits) over time. These algorithms designed sophisticated strategies of reward and punishment.

If one algorithm lowered its price to gain market share, the other would immediately “punish” it by lowering it even further. After a few rounds of mutual punishment, both would gradually return to higher, cooperative prices. No words, no secret meetings, no contracts. Just code optimizing profits.

Over the years, this topic has been the subject of research. Tanja Klein replicated the phenomenon in 2021, and Zach Brown and Alexander MacKay also confirmed it in 2023 using real e-commerce data.

And in 2024, researchers from MIT and Harvard published something even more alarming: they tested large language models (such as GPT-4, Gemini, and Claude) as pricing agents. The models also systematically converged to supra-competitive prices, even with seemingly innocuous modifications to their instructions. The models justified their decision not to lower prices because they were “concerned about price wars,” a strategic rationale that no programmer had explicitly coded.

From the Lab to your Pocket

1. German Gas Stations: The World’s Cleanest Natural Experiment. Between 2017 and 2019, the German fuel market generated the world’s cleanest dataset on fuel prices, thanks to a transparency system that mandates real-time reporting of changes. Using this data, economists Stephanie Assad, Robert Clark, Daniel Ershov, and Lei Xu studied what happened when gas stations adopted algorithmic pricing. Their study, published in the Journal of Political Economy (2024), found that in markets with two competitors, when both stations adopted algorithms, profit margins increased by 28%. In contrast, when only one station adopted the algorithm, prices remained virtually unchanged.
2. RealPage: When 4 Million Apartments Share the Same “Brain” RealPage developed YieldStar, a pricing software that uses machine learning and confidential data shared by landlords to recommend optimal rents, ultimately coordinating prices across more than 4 million apartments in the U.S. Evidence cited by the White House Council of Economic Advisers estimated that in the Washington, D.C. area alone, tenants paid an average of $112 extra per month attributable to this system. Internal customer data showed significant increases, including income improvements of 3–7%, and extreme cases of up to 25% in less than a year. In August 2024, the Department of Justice sued RealPage under the Sherman Act for allegedly facilitating coordinated price fixing among competitors. In January 2026, the company reached a settlement that prohibits the use of non-public, real-time data from competitors and establishes three years of judicial oversight, without fines or an admission of guilt.
3. Amazon’s Project Nessie: The Hidden Algorithm. Between 2014 and 2019, Amazon secretly operated “Project Nessie” which was revealed in 2023 when the FTC (Federal Trade Commission) released portions of its antitrust lawsuit against the company. The algorithm detected products where competitors had price-matching bots and raised prices to see if those systems followed suit; if they did, it maintained the higher price. According to internal documents cited in the lawsuit, in 2018, “Nessie” helped generate approximately $334 million in additional profit and was used to raise prices on ~8 million products. In total, internal estimates suggest it may have generated more than $1 billion in additional profit during its operation. Amazon also paused the algorithm during high-visibility events like Prime Day or Black Friday, suggesting they understood the reputational risk of using it.

How Much Do Consumers Lose? The Transfer of Wealth in Numbers

The most precise study on the net effect of algorithmic pricing comes from Brown and MacKay (2023). They analyzed hourly prices for allergy medications at five major US retailers: Amazon, Walmart, Target, CVS, and Walgreens, using 3.5 million price observations over several years.

Their findings are revealing:

· Average prices: 5.2% higher under algorithmic competition

· Corporate profits: 9.6% higher

· Consumer surplus: 4.1% lower

· Reduction in quantity sold: barely 1%

The implication is clear: the main effect of algorithmic pricing is not a loss of economic efficiency but a direct transfer of wealth from consumers to companies.

A 2023 study of the job platform ZipRecruiter found that algorithmic personalized pricing reduced overall consumer welfare by about 25%. Although 63% of users ended up paying less than they would have under a flat price, the aggregate impact was negative. The study explains that consumers with a lower willingness to pay received discounts, but those with a higher willingness to pay were charged much more. The result shows how personalized pricing can benefit many individuals but worsen the overall market outcome.

In January 2025, the FTC launched a study on “surveillance pricing”, the use of personal data to set individualized prices. They found that at least 250 large companies in the United States use data such as your real-time location, purchase history, demographics, and even your mouse movements to adjust prices specifically for you.

Counterargument: Is everything bad?

Although considerable evidence and data suggest that algorithmic pricing generally reduces overall consumer well-being, it’s also fair to say that in some cases it can be beneficial. Let’s look at some examples.

• A 2016 study of UberX, using nearly 50 million trips, estimated that surge pricing generated approximately $6.8 billion in consumer surplus in the US in 2015 by incentivizing more drivers during peak demand.

• In the airline industry, Williams (2022) found that dynamic pricing on monopoly routes increased overall welfare by about 1% and raised leisure ticket prices by 7.2% by allowing some seats to be sold very cheaply while charging more to business travelers.

• A recent study of electronic price tags in 114 US supermarkets found no relevant evidence of price gouging and documented food waste reductions of up to 21%.

However, a key limitation is that many positive results come from studies funded by companies or using data provided by the companies themselves. Independent academic evidence, especially with data from complete markets, tends to find more risks or negative effects for consumers.

Regulators wake up: the global response

In the US, the Department of Justice warned in 2025 that algorithms facilitating the exchange of price signals could violate antitrust laws and anticipated more litigation. State activity grew rapidly: from 10 bills in 2024 to more than 51 in 24 states in 2025.

California passed AB 325, which prohibits common algorithms in anticompetitive agreements and has a national impact due to the size of its economy. At the federal level, the Preventing Algorithmic Collusion Act was also proposed to limit the use of competitor data and require transparency.

In Europe, the regulatory response is also accelerating, with the European Commission confirming multiple active investigations into algorithmic pricing by 2025. Italy is evaluating its use in airlines through market research. In the UK, the CMA launched a specific project following the Ticketmaster dynamic pricing scandal in 2024. The European approach is focused on coordination risks, consumer protection, and algorithmic transparency.

So… who wins when everyone uses AI for pricing?

The most objective answer is: it depends on the market structure.

If the market is concentrated and interactions are frequent, the mechanism proposed by Calvano et al. suggests that companies can obtain higher margins without explicit coordination, and the consumer is the likely loser due to higher prices. In that context, the “winner” is not necessarily the company with the most sophisticated model, but rather the algorithm that best leverages market conditions and data.

If the market is fragmented, adaptation is easy, and transparency is high, automation can intensify competition and improve efficiency, often benefiting consumers through lower prices or a better match between inventory and demand.

And a third “winner” is emerging: Regulation is progressing, and as algorithmic pricing becomes more mainstream, companies that can demonstrate transparency, security, explainability, and legal use of data could gain a strategic advantage, even if their model is not the most sophisticated.

Keys for the Future

Key 1 — Designing “Anti-Collusion” Algorithms In 2024, Asker, Fershtman, and Pakes showed that algorithm design changes the competitive outcome. Standard Q-learning tends to converge toward collusive pricing, while models with basic economic logic generate more competitive prices. The implication is clear: profitable algorithms can be built without pushing the market toward collusion. But this requires explicit intent in the design, not just revenue optimization.

Key 2 — Platforms as De Facto Regulators Johnson, Rhodes, and Wildenbeest (2023) show that “demand-steering” rules on platforms can break collusive dynamics. Giving more visibility to sellers who lower prices incentivizes competition, even if some try to coordinate. The debate will be whether regulation compels them to use this power to foster competition.

Key 3 — Transparency and Mandatory Algorithmic Audits The regulatory trend points toward registering algorithms, allowing external audits, and disclosing personalized pricing to consumers. Academics are pushing for mandatory third-party audits, similar to financial audits.

Key 4 — Strict Limits on Competitive Data Sharing The RealPage case showed that sharing sensitive data in real time between competitors is the greatest systemic risk. Regulators are already prohibiting this type of data sharing. The key question is whether this rule will expand to more industries. It will likely become the minimum compliance standard for algorithmic pricing.

Key 5 — Evolution of Antitrust for the Algorithmic Era Lawyers are exploring using tools like Section 5 of the FTC against tacit collusion without an explicit agreement. In Europe, there is debate about treating algorithmic collusion as an abuse of dominance, without the need to prove human coordination. This could lower the legal bar for enforcement. Future legal risk may depend more on market outcomes than on human intent.

Main sources consulted: Calvano et al. (2020, AER) | Brown & MacKay (2023, AEJ) | Assad et al. (2024, JPE) | DOJ v. RealPage (2024–2026) | FTC v. Amazon (2023) | Fish et al. (2024) | Johnson et al. (2023, Econometrics) | Asker et al. (2024, JEMS) | Ezrachi & Stucke (2016) | +40 academic papers and reports from competition authorities.