Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

"Create a Fear of Missing Out" -- ChatGPT Implements Unsolicited Deceptive Designs in Generated Websites Without Warning

Veronika Krauß, Mark McGill, Thomas Kosch, Yolanda Thiel, Dominik Schön, Jan Gugenheimer

TL;DR

The study demonstrates that large language models can generate deceptive design patterns (DD) in website interfaces even when prompted with neutral business goals. Using a controlled experiment with 20 participants, the authors show that every ChatGPT-produced single-page site for a fictitious shoe shop contained DD patterns (average 5, max 9) and that GPT-4 offered no warnings about these patterns. The analysis maps the DDs to Gray et al.'s ontology, identifies four novel low-level patterns, and includes a preliminary cross-validation with Gemini 1.5 Flash and Claude 3.5 Sonnet, suggesting the issue generalizes across models. The work highlights ethical and legal implications and argues for robust safety, transparency, and governance to prevent AI-generated DDs from influencing designers and end users in real-world applications.

Abstract

With the recent advancements in Large Language Models (LLMs), web developers increasingly apply their code-generation capabilities to website design. However, since these models are trained on existing designerly knowledge, they may inadvertently replicate bad or even illegal practices, especially deceptive designs (DD). This paper examines whether users can accidentally create DD for a fictitious webshop using GPT-4. We recruited 20 participants, asking them to use ChatGPT to generate functionalities (product overview or checkout) and then modify these using neutral prompts to meet a business goal (e.g., "increase the likelihood of us selling our product"). We found that all 20 generated websites contained at least one DD pattern (mean: 5, max: 9), with GPT-4 providing no warnings. When reflecting on the designs, only 4 participants expressed concerns, while most considered the outcomes satisfactory and not morally problematic, despite the potential ethical and legal implications for end-users and those adopting ChatGPT's recommendations

"Create a Fear of Missing Out" -- ChatGPT Implements Unsolicited Deceptive Designs in Generated Websites Without Warning

TL;DR

The study demonstrates that large language models can generate deceptive design patterns (DD) in website interfaces even when prompted with neutral business goals. Using a controlled experiment with 20 participants, the authors show that every ChatGPT-produced single-page site for a fictitious shoe shop contained DD patterns (average 5, max 9) and that GPT-4 offered no warnings about these patterns. The analysis maps the DDs to Gray et al.'s ontology, identifies four novel low-level patterns, and includes a preliminary cross-validation with Gemini 1.5 Flash and Claude 3.5 Sonnet, suggesting the issue generalizes across models. The work highlights ethical and legal implications and argues for robust safety, transparency, and governance to prevent AI-generated DDs from influencing designers and end users in real-world applications.

Abstract

With the recent advancements in Large Language Models (LLMs), web developers increasingly apply their code-generation capabilities to website design. However, since these models are trained on existing designerly knowledge, they may inadvertently replicate bad or even illegal practices, especially deceptive designs (DD). This paper examines whether users can accidentally create DD for a fictitious webshop using GPT-4. We recruited 20 participants, asking them to use ChatGPT to generate functionalities (product overview or checkout) and then modify these using neutral prompts to meet a business goal (e.g., "increase the likelihood of us selling our product"). We found that all 20 generated websites contained at least one DD pattern (mean: 5, max: 9), with GPT-4 providing no warnings. When reflecting on the designs, only 4 participants expressed concerns, while most considered the outcomes satisfactory and not morally problematic, despite the potential ethical and legal implications for end-users and those adopting ChatGPT's recommendations

Paper Structure

This paper contains 51 sections, 6 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Deceptive Design Practices and Patterns
  4. Large Language Models and Their Ethical Impact
  5. Deceptive Design Patterns from LLMs
  6. Study Design and Analysis
  7. Study Procedure
  8. Set-up
  9. Conditions and Tasks
  10. Design Artifact Assessment
  11. Participant Recruitment and Data Sampling
  12. Recruitment Criteria and Process
  13. Data Sampling
  14. Data Cleaning
  15. Data Analysis
  16. ...and 36 more sections

Figures (6)

  • Figure 1: A schematic visualization of the study procedure of our study's central part.
  • Figure 2: Schematic process of the co-coding sessions and the respective results or our main study.
  • Figure 3: An excerpt of the chat history documenting the conversation between ChatGPT and participant ID 29 (cSign-up). ChatGPT generates replies with in-code explanations as well as in-text instructions, hints and further ideas to satisfy the prompted task (screenshot).
  • Figure 4: Lvl1 (left) and Lvl3 (right) versions of the the shoe shop's website. The HTML files were created in dataset 130:product. The patterns contained in Lvl3 are added as labels.
  • Figure 5: Bar charts depicting the open questions regarding 1) participant satisfaction, 2) creator of the website, 3) responsibility, and 4) morality of the resulting artifact. In 2) and 3), some answers were coded with more than one code, e.g., if a participant perceived both, the programmer and ChatGPT as the creator of the webpage in Subfigure 2. Therefore, in some of the bar charts, the total number of codes exceeds the total number of datasets (n=20).
  • ...and 1 more figures