In the evolving landscape of artificial intelligence, the fidelity and fairness of generated content remain crucial challenges. A recent study, spearheaded by Stevens Institute of Technology undergraduate Gursimran Vasir and guided by Associate Professor Jina Huh-Yoo, sheds illuminating light on the distinct flaws present in AI-generated images. This investigation is a pioneering effort that moves beyond the conventional focus on textual AI biases to scrutinize the visual domain, where errors often manifest in surprising and problematic ways.
Vasir’s journey began in a humble setting: a children’s summer camp Photoshop class where she observed firsthand the disconnect between users’ textual prompts and the AI’s image outputs. Children tasked with creating images using Photoshop’s AI interface repeatedly encountered unexpected, skewed, or biased results. This real-world interaction unveiled a troubling inconsistency—AI’s interpretation of language in image generation was not only error-prone but also imbued with cultural and logical distortions. Vasir’s observation that the children struggled to articulate their frustrations due to the absence of a “standardized language” to describe AI failures inspired her to initiate a systematic categorization of these errors.
Delving into social media as a data source, Vasir meticulously analyzed 482 Reddit posts that detailed user experiences of AI-generated image blunders. Her research culminated in a classification model consisting of four core categories: AI surrealism, cultural bias, logical fallacy, and misinformation. Each category encapsulates a unique facet of AI’s visual generation shortcomings, revealing how algorithmic outputs can both reflect and perpetuate human biases and conceptual misunderstandings.
AI surrealism, as described in the study, characterizes images that evoke a subtle yet pervasive sense of “unreality” — images that appear unnaturally smooth, overly perfect in color saturation, or subtly distorted in ways that generate unease. This phenomenon is reminiscent of the uncanny valley effect in robotics and animation but is manifesting in AI-generated visuals, signaling a dissonance between algorithmic perceptions and human expectations of realism. Such surreal qualities undermine user trust and the practical utility of AI-generated images in professional and creative domains.
Perhaps more disconcertingly, cultural bias surfaced as a dominant theme in the analysis. When users prompted the AI for images involving specific roles or iconic figures, the AI’s outputs betrayed ingrained stereotypes and ethnocentric assumptions. For instance, a request for a “cleaning person” invariably produced images of women cleaning, highlighting gender biases deeply embedded within training datasets. A vividly illustrative example involved a prompt for Jesus Christ walking on water; the AI rendered an image incongruously showing Christ surfing on a board in stormy waters rather than the biblical depiction, revealing how cultural and contextual knowledge is unevenly internalized by AI models.
Misinformation emerges as a third critical category, where the AI generates visually plausible but factually inaccurate representations. Users seeking images of particular cities sometimes received depictions that bore no resemblance to the requested locations, reflecting an underlying lack of accurate geospatial or landmark information in the training corpora. This propensity not only risks spreading incorrect visual knowledge but also complicates applications of AI in domains requiring precision, such as education, tourism, and urban planning.
Logical fallacies in AI-generated images constitute the final category. These errors violate basic principles of reality and coherence—rendering anatomically impossible entities or physically implausible scenarios. Reports of images featuring six fingers on a hand or landscapes with two suns demonstrate an absence of logical constraints in the generative process. Such lapses underscore the gap between statistical pattern matching that AI excels at and genuine semantic understanding that remains elusive, limiting the reliability of AI-generated imagery for serious or scientific use.
Associate Professor Jina Huh-Yoo, an expert in human-computer interaction and emerging technologies, underscores the novelty of this work, emphasizing that prior research has largely neglected visual biases in favor of textual AI outputs. This study’s focus on images adds a vital dimension to the ongoing discourse around AI ethics, accountability, and user experience. She praises the initiative and scholarly rigor demonstrated by Vasir, who independently formulated research questions and methodologies, positioning the project as a promising front in the human-centered AI research field.
The importance of this research transcends academic circles, resonating strongly with the design and graphics industries, which increasingly integrate AI-generated content into their workflows. Huh-Yoo points out that industry stakeholders are actively grappling with the implications of AI errors—not only for aesthetics but also for brand integrity and ethical considerations. The insights offered by this study provide foundational knowledge to develop frameworks and tools that can detect and mitigate biases and errors, enabling safer and more effective creative processes.
As AI technologies permeate various domains—marketing, education, travel, and beyond—the imperative for accurate, unbiased content grows more urgent. Gursimran Vasir articulates a vision for a new vocabulary rooted in the precise characterization of AI image flaws, designed to facilitate transparent dialogue between end-users and developers. She argues that such a shared linguistic framework is essential to identify root causes of failure, guide iterative improvements, and ultimately foster AI systems that align with human values and expectations.
The research points toward a broader principle: transparency in AI is not achieved merely through algorithmic sophistication but also through user empowerment. Equipping users with the tools and language to describe AI glitches and biases turns passive consumption into an active partnership in AI development. This democratization of AI literacy promises to accelerate the maturation of generative models so that they better reflect the diverse realities and contexts in which they operate.
Though still in early stages, Vasir’s work charts a compelling roadmap for advancing the reliability and equity of AI-generated images. By documenting the patterns of AI blunders—ranging from the surreal to the logically impossible—this study lays necessary groundwork for future research, policy formulation, and technology design. As the visual language of AI continues to unfold, addressing these flaws will be pivotal in harnessing the full potential of generative technologies responsibly.
In sum, the study “Characterizing the Flaws of Image-Based AI-Generated Content” not only exposes the multifaceted nature of image-generation errors but also serves as a clarion call for interdisciplinary collaboration. Bridging technical, humanistic, and industry perspectives will be essential to craft AI systems that are trustworthy, inclusive, and context-aware. The collaboration between Stevens undergraduates and faculty exemplifies the promising role of academic innovation in this effort.
The work’s presentation at the prestigious ACM CHI 2025 conference in Yokohama, Japan, underscores its international recognition and relevance. The discussions it sparked among researchers and practitioners alike signal a growing consensus that managing AI biases and errors is a shared responsibility with far-reaching societal implications. As generative AI tools become ever more ubiquitous, the clear articulation and remediation of their flaws will be a defining challenge for the technology sector—and for society as a whole.
Subject of Research: Characterization of errors and biases in AI-generated images
Article Title: AI Blunders: Six-Finger Hands, Two Suns and Jesus Christ on a Surfboard in a Stormy Sea
News Publication Date: June 26, 2025
Web References:
– https://dl.acm.org/doi/10.1145/3706599.3720004
– https://chi2025.acm.org/
References:
Gursimran Vasir and Jina Huh-Yoo, “Characterizing the Flaws of Image-Based AI-Generated Content,” Work-in-progress presented at ACM CHI Conference on Human Factors in Computing Systems, April 25, 2025.
Keywords: Artificial intelligence, AI-generated images, AI biases, human-computer interaction, image generation errors, AI surrealism, cultural bias, logical fallacy, misinformation, generative models, AI ethics, user-developer communication
