AI dynamic pricing and dark patterns: what online shoppers should know

AI dynamic pricing adjusts prices using signals such as demand, stock and timing. It becomes personalized pricing when the price depends on data about the buyer. In January 2025, FTC staff reported that retail pricing tools can use data such as location and browsing behavior. The FTC (2022) says dark patterns push shoppers into choices they would not otherwise make.

EasWrk is a hosting platform, not a research institute. This essay summarises published sources, which are listed at the end. Where a sentence is the author's opinion, it says so.

The Algorithmic Marketplace

A shelf tag used to fix the price for everyone in the store. Online, the price, the order of products and the wording around them can change for each visit. Three mechanisms drive that: dynamic pricing, behavioral targeting and recommendation ranking.

Dynamic pricing adjusts prices from demand, competitor prices and stock levels, and in some cases from data about the shopper. Behavioral targeting uses a profile built from purchases, browsing and other data to choose what each shopper sees. Recommendation ranking decides which products appear first and how prominent they are.

E-commerce has real benefits. It widens choice, makes prices easy to compare and lets small sellers reach buyers far away. The same tools also allow practices that are hard for a shopper to see: prices that differ between people, interface tricks that steer a purchase, and ranking power held by a few platforms. This essay looks at what published research and regulators have documented about each, and where the evidence is still thin.

AI Dynamic Pricing and Price Discrimination

Dynamic pricing is not new. Airlines have used computer-driven yield management for decades, and hotel prices have long varied by season and booking date. What changed is how often and how finely prices can move, and that some systems can draw on data about the individual shopper.

Economists separate three kinds of price discrimination. First-degree discrimination charges each buyer the most they will pay. Second-degree discrimination varies price by quantity or quality tier. Third-degree discrimination varies price across identifiable groups. AI dynamic pricing becomes price discrimination when the price depends on who the buyer is, not just on market conditions.

Acquisti and Varian (2005) modeled when a seller gains by conditioning prices on a customer's purchase history. In their baseline case, a monopolist facing rational buyers with stable valuations and able to commit to a price policy does not gain by doing so. Add features such as enhanced service for past customers, no price commitment or competition, and conditioning generally becomes profitable. The result shows that purchase data can raise profit under some conditions, not that every seller does it.

Hannak and colleagues (2014) tested 16 e-commerce sites using the accounts and cookies of more than 300 real users, plus controlled fake accounts. They found evidence of some form of personalization on nine of the sites. Home Depot served different search results to mobile browsers, steering them toward more expensive products. Travelocity lowered hotel prices by about $15 on about 5% of hotels for iOS users. The authors left Amazon out, because it hosts many merchants and is hard to measure.

In January 2025, FTC staff published initial findings from a study of surveillance pricing, launched with orders issued in July 2024. Staff concluded that the tools they studied can use a wide range of consumer data, from location to mouse movements, to tailor prices. The FTC said the staff perspective includes only hypothetical examples, the study was still ongoing, and the findings were preliminary (FTC, 2025). The report does not say that any firm broke the law.

AI dynamic pricing example: the $23 million book

In April 2011, evolutionary biologist Michael Eisen found a copy of the reference book The Making of a Fly listed on Amazon for $23,698,655.93 plus $3.99 shipping. Two third-party sellers, bordeebook and profnath, had each set prices by algorithm from the other's price. Once a day, profnath set its price to 0.9983 times bordeebook's. Several hours later, bordeebook raised its price to 1.270589 times profnath's. The two repeated that loop until the price climbed past $23 million (The Register, 2011).

This was not Amazon's own pricing system. It was two sellers' repricing tools reacting to each other with no limit built in. It shows that algorithms in competition can produce prices no one chose.

AI dynamic pricing airlines: yield management and Delta

Airlines were early users. American Airlines described its yield management system, which set overbooking levels and discount seat allocation, in Interfaces in 1992 (Smith, Leimkuhler and Darrow, 1992). In a letter dated 31 July 2025, Delta said it was evaluating Fetcherr's AI pricing tool in domestic and international test markets. Delta also said its "ticket pricing never takes into account personal data" and that prices "are not targeted to individual consumers" (Delta, 2025). Those are Delta's statements. This essay does not test them.

AI dynamic pricing grocery stores: the Instacart test

On 9 December 2025, Groundwork Collaborative, Consumer Reports and More Perfect Union published a study of grocery prices on Instacart. Volunteers added the same 18 to 20 items to carts at the same stores. The authors analysed 193 submissions and found that 74% of items were offered at more than one price to different shoppers. For those items the average gap between lowest and highest price was 13%, and some shoppers saw prices up to 23% higher. The authors say their participants are not representative of the US population (Groundwork Collaborative, 2025).

Instacart said that most customers see the standard price, that the tests do not use personal or demographic data, that customers are randomly assigned to test groups, and that the changes are negligible. The study shows that prices varied between shoppers. It does not show why.

Rules on algorithmic pricing

Most consumer protection law targets deceptive pricing, such as a false "was" price. It rarely addresses an accurate price that differs between shoppers. New York's Algorithmic Pricing Disclosure Act (General Business Law section 349-a) is an early exception. It requires a retailer to display "THIS PRICE WAS SET BY AN ALGORITHM USING YOUR PERSONAL DATA" next to prices set that way. On 8 October 2025, the US District Court for the Southern District of New York dismissed a First Amendment challenge to the Act (Benesch, 2025). Check current status before relying on it.

Dark Patterns in Ecommerce

Dark patterns are interface designs that trick or manipulate users into choices they would not otherwise make. Harry Brignull coined the term in 2010 (FTC, 2022). Examples include hidden costs revealed at the last checkout step, pre-ticked marketing boxes and subscriptions that are easy to start and hard to cancel.

Mathur and colleagues (2019) analysed about 53,000 product pages from about 11,000 shopping websites. They found 1,818 dark pattern instances on 1,254 websites, about 11.1% of those crawled. The instances fell into 15 types in seven categories. A subset of 234 instances across 183 websites were deceptive. The types included countdown timers, low-stock messages, and claims about what other shoppers are doing.

The FTC (2022) notes that online sellers can test design changes more easily, more often and at larger scale than a physical store. Its example is Credit Karma, which A/B tested two claims, "pre-approved" and "Excellent" odds, and kept the "pre-approved" claim that drew more clicks. The FTC alleged that claim was false. The report also says data collection lets companies tailor advertising to a particular consumer's interests. In this essay's view, the same machinery could be used to find which manipulative design works best, and for whom. The FTC report treats that as a risk, not a measured outcome.

Luguri and Strahilevitz (2021) measured the effect in an experiment with 1,963 US adults offered an identity theft protection service. In the control group, 11.3% accepted. With mild dark patterns, 25.8% accepted. With aggressive dark patterns, 41.9% accepted, nearly four times the control rate.

The OECD (2022) offers a working definition of dark commercial patterns and reviews evidence on how common they are, how effective they are and what harm they cause. It also sets out options for policymakers and enforcers.

The FTC's case against Amazon over Prime shows the issue in enforcement. The FTC alleged that Amazon enrolled consumers in Prime without consent using dark patterns and made cancellation hard. The case, filed in June 2023, ended with a stipulated order on 25 September 2025, announced as a $2.5 billion settlement (FTC, 2025b).

Behavioral Targeting and Manipulative Design

Behavioral targeting draws on well-known findings in behavioral economics. Loss aversion, the tendency to weigh a loss more than an equal gain, is the logic behind low-stock warnings and countdown timers. Anchoring, the pull of the first number seen, is the logic behind a struck-through "original" price. Social proof, the habit of following what others do, is the logic behind "bestseller" labels. Those links are this essay's reading of the mechanisms, not findings from one study.

Ezrachi and Stucke (2016) use the term "behavioral discrimination" for exploiting psychological quirks to steer purchases. One of their examples is a travel booking site leading some users toward more expensive hotels by placing them higher in search results. They argue that competition law has paid little attention to this kind of steering.

In this essay's opinion, a store that runs its own checkout and pricing controls what its shoppers see. A seller on a large marketplace does not. The ranking system decides which products get attention, and a seller can offer a better product at a lower price and still go unseen. Sellers who run their own store keep that control, along with the duty to use it fairly. The social media essay covers how similar profiling works on social platforms.

Platform Concentration and Market Power

Amazon acts as both marketplace operator and seller. That dual role drew regulatory attention in Europe. On 10 November 2020, the European Commission sent Amazon a Statement of Objections. It preliminarily found that Amazon's use of marketplace sellers' non-public business data to calibrate its retail decisions distorted competition, and in a second investigation it preliminarily concluded that Amazon's Buy Box and Prime rules unduly favoured its own retail business (European Commission, 2022).

The Commission did not issue a final ruling. On 20 December 2022 it made commitments offered by Amazon legally binding, without finding an infringement. Amazon committed not to use non-public seller data for its retail business, to treat all sellers equally when choosing the Buy Box winner, and to display a second competing offer when one exists (European Commission, 2022).

The Digital Markets Act (Regulation (EU) 2022/1925) came from a Commission proposal of 15 December 2020, COM(2020) 842 final. It entered into force in November 2022 and applies from 2 May 2023. It sets obligations for large platforms designated as gatekeepers. Article 5(4) requires a gatekeeper to let business users communicate and promote offers to customers acquired through its platform, free of charge, and to conclude contracts with them regardless of whether they use the gatekeeper's platform for that (European Union, 2022a). How far enforcement reduces concentration is still unproven.

Rules in the US and EU

In the United States, the FTC enforces consumer protection under its authority to stop unfair or deceptive practices. Its Amazon Prime case and its 2022 dark patterns report show that it treats manipulative design as a matter for enforcement. The New York disclosure law above is a state-level rule aimed at personalized algorithmic pricing.

The European Union covers the same practices through several laws. The Digital Markets Act addresses gatekeeper power. The Digital Services Act (Regulation (EU) 2022/2065, Article 25) says providers of online platforms shall not design, organise or operate their interfaces in a way that deceives or manipulates users or materially distorts their ability to make free and informed decisions. That Article does not apply to practices already covered by Directive 2005/29/EC or the GDPR (European Union, 2022b). The GDPR (Regulation (EU) 2016/679) governs the personal data that behavioral targeting uses. The AI Act (Regulation (EU) 2024/1689) was adopted on 13 June 2024, and its dates of application are staged from 2 February 2025 (European Union, 2024).

The overlaps between these laws are not always clear, and the capacity to investigate complex algorithmic systems is a practical limit. This essay does not cover UK or Canadian rules.

Consumer Strategies and the Limits of Awareness

Practical habits can help. Check the total price before you pay. Look at the cancellation path before you start a subscription. Compare the same item in a logged-out browser and on another device, which is a simple version of the test Hannak and colleagues (2014) ran. Screenshot a price you plan to dispute.

In this essay's opinion, these habits help but cannot be the whole answer. A shopper faces systems that companies tune against shopper behavior. The FTC (2022) notes that dark patterns often work in combination, and the Luguri and Strahilevitz (2021) results show that even mild designs can double acceptance. Rules that require disclosure, limit the most harmful practices and give regulators access to audit algorithmic systems place less of the burden on each shopper.

Small sellers have a part too. If you sell online, the clearest protection against these risks is to show the full price early, make cancellation as easy as sign-up, and avoid countdown timers or stock claims that are not true. The small business AI adoption essay covers how owners can adopt AI tools without losing control of their data. More research is on the research hub.