Source-linked AI summary
Dark Patterns at Scale: Findings from a Crawl of 11K Shopping Websites
Arunesh Mathur, Gunes Acar, Michael J. Friedman, Elena Lucherini, Jonathan Mayer, Marshini Chetty, Arvind Narayanan
TL;DR
Dark patterns can steer users toward unintended and harmful decisions, but prior work lacked large-scale evidence about their prevalence and effects. The paper develops automated crawling, clustering, and expert-analysis techniques for shopping websites, finding 1,818 instances across 15 types and 7 categories, deceptive practices on 183 websites, and 22 third-party enablers. Its taxonomy connects observed patterns to interface characteristics, cognitive biases, and potential harms.
Problem
Prior work lacked large-scale evidence documenting dark-pattern prevalence and systematically investigating how different types harm users.
Method
The authors simulate shopping journeys, extract and cluster textual interface designs, inspect clusters for dark patterns, and develop a taxonomy of their characteristics.
Results
1,818 dark-pattern instances represented 15 types and 7 broad categories, with deceptive behavior across 183 websites and 22 third-party enabling entities identified.
Takeaways & Limitations
The findings provide measurement techniques, prevalence data, and terminology to help researchers, journalists, regulators, and tool developers study and counter dark patterns.
Takeaways & Limitations
The study analyzes text-based interfaces and crawls only product and checkout pages, excluding visual patterns and patterns on other page types.
Abstract
from arXiv · showhide
Dark patterns are user interface design choices that benefit an online service by coercing, steering, or deceiving users into making unintended and potentially harmful decisions. We present automated techniques that enable experts to identify dark patterns on a large set of websites. Using these techniques, we study shopping websites, which often use dark patterns to influence users into making more purchases or disclosing more information than they would otherwise. Analyzing ~53K product pages from ~11K shopping websites, we discover 1,818 dark pattern instances, together representing 15 types and 7 broader categories. We examine these dark patterns for deceptive practices, and find 183 websites that engage in such practices. We also uncover 22 third-party entities that offer dark patterns as a turnkey solution. Finally, we develop a taxonomy of dark pattern characteristics that describes the underlying influence of the dark patterns and their potential harm on user decision-making. Based on our findings, we make recommendations for stakeholders including researchers and regulators to study, mitigate, and minimize the use of these patterns.
1 INTRODUCTION
The paper addresses the lack of large-scale evidence about dark patterns by developing automated techniques to identify and characterize them on shopping websites. Its crawl found widespread patterns, deceptive practices, and third-party infrastructure supporting their deployment.
- Motivation: Dark patterns coerce, steer, or deceive users into decisions they might not make if fully informed and able to choose alternatives.They can contribute to financial loss, privacy loss, and compulsive behavior.
- Research gap: Prior work lacked large-scale evidence of dark-pattern prevalence and systematic investigation of how different types harm users.Such evidence could support countermeasures and regulatory efforts.
- Approach: The authors combine web crawling, text clustering, and expert inspection to identify dark patterns at scale.The crawler simulates browsing, clustering extracts interface designs, and experts inspect resulting clusters.
- Findings: 1,818 instances represented 15 dark-pattern types and 7 broad categories across the study’s shopping-web corpus.The instances were found on 1,254 of approximately 11K websites, or approximately 11.1%.
- Findings: 234 deceptive dark-pattern instances occurred across 183 websites, while 22 third-party entities offered tools to implement dark patterns.The paper also reports that many observed patterns were covert, deceptive, information hiding, or bias-exploiting.
- Contributions: The paper contributes measurement techniques, a prevalence dataset, a taxonomy of harmful characteristics, and documentation of third-party enablers.The released artifacts are intended to support research, awareness, regulation, and user-facing countermeasures.
2 RELATED WORK
Related work established dark-pattern taxonomies and theories of manipulation, but often relied on anecdotes, user submissions, or domain-specific studies. This paper extends that work with large-scale evidence and a taxonomy explaining underlying influence and harmful effects.
- Decision-making theories: Research on market manipulation links interface influence to cognitive limitations, framing, accessible information, delayed rewards, and digital marketplace capabilities.These mechanisms have been discussed in relation to financial loss, privacy loss, and reduced independent decision-making.
- Prior taxonomies: Earlier research created taxonomies of manipulative interface designs, including patterns such as Bait and Switch and Confirmshaming.Brignull coined the term in 2010 and used shopping and travel examples to raise awareness.
- Domain studies: Researchers also studied dark patterns across web, mobile, gaming, privacy, and online-shopping contexts.Examples include irresistible app designs, deceptive marketing, and privacy-unfriendly defaults and framing.
- Paper’s distinction: Unlike prior taxonomy work based largely on anecdotes or user submissions, this study documents dark-pattern presence and prevalence in the wild and characterizes their harmful effects.Its taxonomy provides terminology for comparing how different patterns influence users.
- Scope distinction: Impulse-buying features overlap with some dark patterns, but persuasive shopping features such as returns and multiple product images are not necessarily dark patterns.The paper distinguishes general persuasion from designs that adversely influence users through manipulation or deception.
3 A TAXONOMY OF DARK PATTERN CHARACTERISTICS
The paper’s taxonomy describes dark patterns through interface characteristics and the cognitive biases they exploit. It distinguishes unequal choice burdens, hidden effects, deception, obscured information, and restricted choices.
- Taxonomy framework: The taxonomy explains how dark patterns affect decision-making through design characteristics and exploited cognitive biases.It is grounded in literature on online manipulation and prior dark-pattern research.
- Asymmetric: Asymmetric designs impose unequal weights or burdens on explicit choices, such as making cookie acceptance prominent and opt-out less visible.The paper narrows asymmetry to explicit choices presented in the interface.
- Covert: Covert designs hide their steering effect, so users may not recognize that an interface choice is influencing their decisions.A decoy option can make another choice seem more appealing while concealing the intended influence.
- Deceptive: Deceptive designs induce false beliefs through affirmative misstatements, misleading statements, or omissions.A repeatedly renewable discount may appear to be limited-time without actually being so.
- Hides Information and Restrictive: Information-hiding designs obscure or delay necessary information, while restrictive designs limit the choices available to users.Examples include revealing extra charges only at checkout or requiring social-media sign-up to gather user information.
- Cognitive biases: The taxonomy connects patterns to biases including anchoring, bandwagon, default, framing, scarcity, and sunk-cost effects.These biases describe deviations from rational behavior that dark patterns may exploit.
4 METHOD
The method simulates a shopper’s primary purchase path, extracts textual interface elements, clusters recurring designs, and has experts inspect them for dark patterns. The corpus is built from popular English-language shopping websites, so text-only crawling yields a lower bound.
- Detection pipeline: The pipeline automates primary website interactions, extracts textual interface elements, and clusters them for expert analysis.Clustering helps analysts sift through recurring messages and interface designs.
- Crawling: The crawler follows product discovery, cart addition, and checkout to approximate an ordinary shopping experience.The study focuses on shopping websites even though the approach is intended to generalize.
- Scope limitation: Text-only analysis excludes visual dark patterns using style, color, or other non-textual features, making the findings a lower bound.The authors leave detection of these patterns to future work.
- Corpus construction: The corpus was selected to represent popular global shopping websites and was restricted to English so collected text could be analyzed.These criteria shaped the constructed website list.
- Shopping-site identification: 361K websites were screened, and Webshrinker achieved 94% classification accuracy versus Alexa’s 89% on a 500-site sample.Webshrinker also had an 18% false-negative rate versus Alexa’s 93%, despite a slightly higher false-positive rate of 0.4% versus 0.2%.
4.2 Data Collection with a Website Crawl
The study built a Selenium/OpenWPM-based crawl that followed shopping users’ paths from product discovery through checkout, collecting page and interface data at scale.
- Discovering Product URLs: The crawler ranked URLs by product likelihood and identified product pages using an ‘Add to cart’ or similar button.A Logistic Regression classifier achieved 83% average accuracy using five-fold cross validation.
- Discovering Product URLs: 53,180 product pages were gathered from 11,286 shopping websites after crawling the broader shopping-site corpus.The crawler successfully returned legitimate product pages, or none when unavailable, for 86 of 100 evaluated websites.
- Checkout Crawling: The checkout crawler simulated selecting product options, adding products to carts, viewing carts, and checking out.It successfully reached checkout on 66 of 100 sampled product pages.
- Checkout Crawling: The crawl stored page source, screenshots after state changes, and HAR files containing full response contents.These data supported analysis of both static and dynamically changing interface elements.
- Segmenting Pages: Page segments were visible text-containing HTML elements, including transient updates captured through Mutation Summary events.The resulting crawls contained approximately 13 million segments across 53K product URL pages.
4.3 Data Analysis with Clustering
The analysis reduced millions of interface segments into manageable text clusters, which researchers then inspected to identify candidate dark patterns and validate their coding.
- Clustering: The researchers used hierarchical clustering to organize segments for expert scanning rather than to infer latent constructs.This made it easier for analysts to sift through possible dark patterns.
- Feature Representations: The Bag of Words representation retained 10,133 tokens after filtering stop words, punctuation, and infrequent terms.Currency symbols were retained because they indicate product price.
- Feature Representations: PCA retained 3 components that together captured more than 95% of the variance in the data.
- Clustering: HDBSCAN was selected for robustness to noise and flexible minimum cluster sizes, using Manhattan or L1 distance across four passes.
- Clustering: The selected clustering output used min_cluster_size 10 and contained 10,277 clusters.Its output was nearly the same as alternatives, but it produced significantly fewer clusters.
- Expert Inspection: Researchers made two inspection passes, then used dark-pattern and impulse-buying literature to guide coding and validated a sample of 200 clusters with two researchers.
4.4 Detecting Deceptive Dark Patterns
The researchers examined dynamic dark patterns for deceptive practices by repeatedly monitoring relevant websites and applying pattern-specific heuristics.
- Detecting Deceptive Practices: Dynamic dark patterns were monitored once every four hours for five days to detect deceptive practices.The monitoring data were combined with dark-pattern-specific heuristics.
5 FINDINGS
The study found widespread dark patterns across shopping websites, but its prevalence estimate is a lower bound because the crawl and analysis covered only limited pages and text-based interfaces.
- Overall Findings: 1,818 dark-pattern instances were identified across 15 types and 7 broader categories.Table 1 summarizes the categories, types, descriptions, prevalence, and definitions.
- Scope and Limitation: The reported prevalence is a lower-bound estimate because the crawler covered only product, cart, and checkout pages and the analysis considered only text-based interfaces.
5.1 Categories of Dark Patterns
The study identifies 15 dark-pattern types across seven categories, documenting how shopping websites use sneaking, urgency, and misdirection to influence user choices. These patterns include hidden information, deceptive deadlines, and confusing or visually manipulated choices.
- 15 dark-pattern types span 7 broader categories, with their distribution summarized by the number of separate instances found for each type.
- Sneaking: Three Sneaking types—Sneak into Basket, Hidden Costs, and Hidden Subscription—misrepresent actions or hide information users might otherwise reject.Sneak into Basket can add products without consent, Hidden Costs reveal charges late in checkout, and Hidden Subscription can create recurring fees.
- Urgency: 437 websites used Countdown Timers or Limited-time Messages to impose or imply deadlines across product, cart, and checkout pages.The study found 393 Countdown Timer instances and 88 Limited-time Message instances; 157 Countdown Timers were deceptive on 140 websites.
- Urgency: Countdown Timers were sometimes deceptive or covert, while Limited-time Messages hid undisclosed deadlines and were at least partially covert.One Countdown Timer continued displaying an offer after its 60-minute countdown expired.
- Misdirection: Four Misdirection types appeared on 244 websites, steering choices through shameful language, confusing questions, visual presentation, or pressured selling.The taxonomy classified Trick Questions as asymmetric and covert, Visual Interference as sometimes asymmetric, covert, and deceptive, and Pressured Selling as sometimes asymmetric and partially covert.
5.2 Dark Patterns as A Third-Party Service: A Case Study Of Social Proof Activity Notifications
The study finds that third-party entities widely provide infrastructure for dark patterns, especially Social Proof Activity Notifications, and sometimes openly advertise deceptive capabilities.
- Detection: The researchers combined HTTP-response analysis, HAR-file searches, manual website inspection, and domain verification to identify responsible third parties.They used customer name-location pairs, WHOIS records, script domains, search engines, and browser DOM-change breakpoints.
- Prevalence: 22 third-party entities were embedded in 1,066 of the 11K shopping websites examined.The authors also measured these entities across the Alexa top million websites, where they appeared on 7,769 sites.
- Services: Third-party services commonly provided Social Proof Activity Notifications, while broader marketing platforms also enabled Scarcity and Urgency patterns.The entities included plugins for Shopify and WooCommerce.
- Deceptive capabilities: Some third parties openly promoted manipulative tactics such as exploiting fear of missing out, creating urgency, and encouraging impulse purchases.The reported marketing language framed these practices as ways to increase conversions or speed sales.
- Deceptive capabilities: Some providers openly advertised fabricated orders or fake social-proof messages, and Beeketing’s Sales Pop plugin was temporarily removed from Shopify.The plugin allowed retailers to create fake Activity Notifications by entering fabricated sales data.
- Implications: The authors conclude that third-party entities widely enable dark patterns and that some advertise deceptive uses of their services.The dataset and proposed browser-extension mockup support potential countermeasures for detecting and flagging these patterns.
6 DISCUSSION
The discussion connects the findings to consumer protection and regulatory enforcement, while emphasizing that automated measurement can support broader detection of dark patterns. It also identifies methodological and scope limitations that constrain coverage and interpretation.
- Countermeasures: Many dark patterns are manipulative or deceptive, and the dataset could support websites or browser extensions that help users identify them.The paper gives a browser extension that flags dark patterns as one possible countermeasure.
- Findings: The study found deceptive dark patterns and 22 third-party entities that enable Social Proof Activity Notifications, including providers promoting deceptive practices.The authors characterize these practices as influencing consumer behavior for profit.
- Regulation: The authors state that some identified practices are unlawful in the United States and European Union under consumer-protection laws.They specifically connect deceptive representations to the FTC Act and the EU Unfair Commercial Practices Directive.
- Regulation: In the European Union, Sneaking patterns and certain consent practices may violate disclosure and independent-consent requirements.The discussion identifies Sneak into Basket, Hidden Subscription, Hidden Costs, Trick Questions, and Visual Interference as relevant examples.
- Measurement: The crawling and clustering methodology reduces the difficulty of discovering and measuring dark patterns at web scale for regulators and other consumer-protection stakeholders.The authors also present the third-party-entity dataset as a resource for these stakeholders.
- Limitations: The study is limited because it detects text-based patterns, includes gray-area judgments, and may not capture all sites or user susceptibilities.The authors note that visual patterns are excluded, expert and user opinions may vary, and some cognitive-bias connections are more nuanced.
7 CONCLUSION
The paper develops automated techniques to study dark patterns across approximately 11K shopping websites and reports their prevalence, deceptive use, and third-party enablement. It closes by calling for empirical evaluation, countermeasures, and research in other domains.
- Conclusion: Automated techniques simulated user actions and collected text and screenshots from approximately 11K popular shopping websites.The authors linked dark-pattern definitions to cognitive biases to characterize how the patterns affect decisions.
- Conclusion: At least one dark pattern appeared on approximately 11.1% of examined websites, and 183 websites displayed deceptive messages.The study also found that dark patterns were more likely to appear on popular websites.
- Conclusion: The study identified 22 third-party entities, including two that advertised practices enabling deceptive patterns.The paper describes these entities as frequently enabling dark patterns on shopping websites.
- Conclusion: The authors recommend empirically evaluating effects on user behavior, developing countermeasures, and extending detection to other domains.Suggested future domains include websites beyond shopping.
A APPENDIX
The appendix includes an illustration of page segmentation and the corresponding algorithm, which breaks pages into meaningful segments while retaining text-containing content. It also lists a separate classifier-evaluation table.
- Classifier Evaluation: Table 3 is titled “Confusion Matrices From Our Evaluation of Alexa’s and Webshrinker’s Website Classifiers.”The supplied appendix passage identifies the table and the classifiers evaluated, but not the matrix values.
- Page Segmentation: The page segmentation illustration divides a webpage into smaller meaningful building blocks or segments and records only segments containing text.This figure provides the conceptual view of the segmentation process.
- Page Segmentation: The Page Segmentation algorithm filters ignored, invisible, or sub-pixel elements before recursively processing the remaining DOM structure.Ignored tags include script, style, noscript, br, and hr; elements no bigger than 1 pixel are excluded.
- Page Segmentation: For block elements, the algorithm either returns the element, recursively segments its children, or returns an empty list depending on nested blocks, ignored children, and page coverage.A 30% page-occupation threshold determines whether some elements are split into child segments.