Characterizing the Flaws of Image-Based AI-Generated Content

Gursimran Vasir, Jina Huh-Yoo

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

The advancement of foundation models provides opportunities to efficiently generate multimodal content with reduced manual labor and time. Such AI-generated content (AIGC), however, can often be inaccurate, misleading, surreal, or hallucinating. Researchers developed multiple ways to systemize our understanding of the bias, errors, and failures of AI systems. However, little work has been done to characterize the flaws of image-based AIGC. In this work-in-progress, we analyzed 482 Reddit posts on various flaws of AIGC experienced by AIGC tool users. We found four themes describing the flaws of AIGC—logical fallacy, AI surrealism, misinformation, and cultural bias. We compare the results with the existing text-based AIGC framework on errors to discover unique flaws that image-based AIGC creates. We discuss implications toward a framework to describe, understand, and interpret flaws in AIGC in the broader context of understanding our social-technical world.

Original languageEnglish
Title of host publicationCHI EA 2025 - Extended Abstracts of the 2025 CHI Conference on Human Factors in Computing Systems
ISBN (Electronic)9798400713958
DOIs
StatePublished - 26 Apr 2025
Event2025 CHI Conference on Human Factors in Computing Systems, CHI EA 2025 - Yokohama, Japan
Duration: 26 Apr 20251 May 2025

Publication series

NameConference on Human Factors in Computing Systems - Proceedings

Conference

Conference2025 CHI Conference on Human Factors in Computing Systems, CHI EA 2025
Country/TerritoryJapan
CityYokohama
Period26/04/251/05/25

Keywords

  • AI bias and fairness
  • AI-generated content
  • AI-generated images
  • Artificial Intelligence
  • Chatbot
  • hallucination
  • Human-AI Interaction
  • large language models

Fingerprint

Dive into the research topics of 'Characterizing the Flaws of Image-Based AI-Generated Content'. Together they form a unique fingerprint.

Cite this