A Comprehensive Study on Dark Patterns

TL;DR

This paper introduces a comprehensive framework and taxonomy for dark patterns, evaluates detection tools and datasets, revealing significant gaps and suggesting directions for improved identification and classification of manipulative UI elements.

Contribution

The study develops the Dark Pattern Analysis Framework and a detailed taxonomy of 68 dark pattern types, and evaluates detection tools and datasets for coverage and effectiveness.

Findings

Detection tools identify only 45.5% of dark pattern types

Datasets cover only 44% of dark pattern types

Unified datasets reveal gaps in current detection capabilities

Abstract

As digital interfaces become increasingly prevalent, certain manipulative design elements have emerged that may harm user interests, raising associated ethical concerns and bringing dark patterns into focus as a significant research topic. Manipulative design strategies are widely used in user interfaces (UI) primarily to guide user behavior in ways that favor service providers, often at the cost of the users themselves. This paper addresses three main challenges in dark pattern research: inconsistencies and incompleteness in classification, limitations of detection tools, and insufficient comprehensiveness in existing datasets. In this study, we propose a comprehensive analytical framework--the Dark Pattern Analysis Framework (DPAF). Using this framework, we developed a taxonomy comprising 68 types of dark patterns, each annotated in detail to illustrate its impact on users, potential scenarios, and real-world examples, validated through industry surveys. Furthermore, we evaluated the effectiveness of current detection tools and assessed the completeness of available datasets. Our findings indicate that, among the 8 detection tools studied, only 31 types of dark patterns are identifiable, resulting in a coverage rate of just 45.5%. Similarly, our analysis of four datasets, encompassing 5,561 instances, reveals coverage of only 30 types of dark patterns, with an overall coverage rate of 44%. Based on the available datasets, we standardized classifications and merged datasets to form a unified image dataset and a unified text dataset. These results highlight significant room for improvement in the field of dark pattern detection. This research not only deepens our understanding of dark pattern classification and detection tools but also offers valuable insights for future research and practice in this domain.