Value context and insight. lkm@koreatimes.co.kr
Kookmin Univ., UBC researchers offer new framework to make generative AI safer
Professor Kim Min-gyu / Courtesy of Kookmin University
By Lee Kyung-min
As generative artificial intelligence (AI) advances, the ability to create vivid images and videos has often outpaced the ability to control them. As these systems move beyond digital art into higher-stakes fields like medicine and autonomous driving, concerns over “hallucinations” and the leakage of copyrighted data have become more pressing.
Kim Min-gyu, a professor of AI at Kookmin University, argues that the answer lies in a more rigorous mathematical approach. In a paper selected for oral presentation at the International Conference on Learning Representations 2026, he and collaborators at the University of British Columbia introduced a framework called Safety-Guided Flow (SGF).
The approach aims to bring coherence to a fragmented set of safety tools. Developers have typically relied on “denoising” techniques to filter harmful output, but these methods often function as ad hoc fixes. The study shows that such techniques can be understood within a single mathematical framework. By modeling safety as a “potential function,” SGF can more effectively block problematic output and reduce the risk of models memorizing and reproducing copyrighted material.
The research also highlights a critical timing issue. Drawing on control theory, the team found that safeguards are most effective at the earliest stage of generation, when an AI system begins forming an image from noise. Interventions can then be gradually reduced as the output stabilizes.
The work comes as regulators and industry leaders intensify scrutiny of generative AI systems. Governments in the United States, Europe and Asia are weighing rules to address risks ranging from misinformation to intellectual property violations, while companies face growing pressure to demonstrate that their models are both safe and transparent.
The researchers said the framework could provide a more systematic approach to safe generation, showing that existing safety methods can be unified and applied more effectively — particularly by concentrating safeguards in the early stages of the generation process.