At a children’s summer camp, undergraduate Gursimran Vasir from Stevens Institute of Technology noticed some puzzling issues with Photoshop’s AI image search. Many young users were getting unexpected results—images that were oddly distorted or showed clear biases.
For example, when Vasir searched for a “cleaning person,” the AI mostly returned photos of women. A search for “a woman cleaning” amplified the trend, predominantly showing white women in stereotypical scenarios. Vasir recalls that while many kids were frustrated by these mix-ups, they often couldn’t quite put their discomfort into words.
This challenge prompted her to reach out to Associate Professor Jina Huh-Yoo, a specialist in human‐computer interaction, to delve deeper into these inconsistencies. Together, they launched a study—aptly titled “Characterizing the Flaws of Image-Based AI-Generated Content”—which later featured at the ACM CHI Conference on Human Factors in Computing Systems. Vasir examined 482 Reddit posts discussing such glitches and sorted them into categories such as AI surrealism, cultural bias, logical fallacies, and misinformation. AI surrealism included images with unnaturally smooth textures or overly vivid colours, while cultural bias was apparent in scenarios like depictions of Jesus Christ surfing instead of walking on water. There were also clear logical missteps, including extra fingers or landscapes littered with more than one sun.
Professor Huh-Yoo emphasises that this research helps shift the conversation, drawing attention not only to text-based AI errors but also to how visual content can mislead. The study has resonated with industry professionals who are also seeking ways to iron out these quirks in AI-generated material. As you may have experienced yourself, when technology doesn’t perform as expected, it can lead to frustrating miscommunications. In response, establishing a shared vocabulary to discuss these issues could be key to refining future AI systems.
“Developers need to deliver tools that work as intended,” Vasir points out. “When technology falls short, it creates room for misinterpretation and misuse. By standardising the language around these errors, we take a crucial step toward bridging the gap between what users need and what developers build.”
