Artist Lee Jung-woo explains data bias in AI generative models and the absence of 'hesitation' in his working method

In the 'Korea Artist Prize 2026' interview video released by MMCA, artist Lee Jung-woo explained the statistical characteristics of data in AI generative models and the phenomena arising from them. Through this exhibition, the artist continues his attempt to visualize flows that are invisible but act routinely, much like gravity.

The result of 'triggering' created by AI model data training

Lee Jung-woo mentioned the technical limitations and data characteristics experienced during his work using AI. While working on regenerating deleted scenes from the 1946 film 'Viva Freedom!' using AI, he attempted to direct a scene where someone holds a gun and hesitates, but the result was different. According to the explanation in the video, even when the command 'hesitates' was entered into the prompt, the AI produced a result of firing the gun. This phenomenon occurs because, in the process of the AI generative model learning existing data to infer images, there is much more data of firing a gun than data of waiting without shooting.

Furthermore, technical bias also appeared in the scene of a shooter holding a gun with their left hand. The artist stated that he attempted to prevent the left-handed shooter from firing, but the AI showed a pattern of changing the hand to the right hand to fire while holding the gun with the left hand. He explained that this is a result of the statistical flow, where the history of firearm systems and tools is centered around the right hand, overlapping with the AI's image generation process.

The cycle of data created by World Model and Physical AI

The artist also raised the topic of 'World Model', such as AI-based physics engines. While existing game engines digitize the world through physical calculations, a World Model takes the approach of learning physical phenomena through video data and inferring them via pixels. The artist raised questions about what changes will occur when World Models, trained based on YouTube or game broadcasts, overlap with the Physical AI that will encounter us in the future.

The work title 'Cloud-based' refers to this cyclical structure of data. Much like the concept of the cloud that uses data stored on the web, the artist explains that he focused on a kind of cyclical method where the data learned by the World Model is ultimately the data produced and consumed by humans in daily life, and the Physical AI learned through this returns to the world again.

From capturing failures and errors to unfolding things as they are

Lee Jung-woo recalled that his past working methodology was 'a method of capturing moments of failure or error to construct a narrative.' This includes the case of using fake ingredients in his first work 'My First Hamburger' and the experience of capturing the moment a drone crashed in his 2019 work 'Resitanz'. The artist stated that while he used to judge these malfunctions himself and fit them into a story in the past, the flow of his work has since moved toward accepting and unfolding the appearing phenomena exactly as they are.

The artist defined these malfunctions or errors as 'exceptional things' and 'exploration.' It is a perspective that something going astray is not necessarily wrong, but can be room to discover something new. This exhibition also focuses on talking about the ambiguous meanings of malfunctions and the points of exploration.