Soongsil undergraduates address AI’s omission bias and safety gaps

[Financial News] Generative AI’s longstanding limitations—including “moral bias,” a “mismatch between inner thoughts and words,” and “execution risks”—are being addressed by undergraduates at a Korean university, who took the lead in finding solutions.
Soongsil University announced on the 8th that four papers by a team led by Professor Park Chan-jun of its School of Software had been accepted at AACL-IJCNLP 2026, a prestigious international conference in natural language processing. AACL-IJCNLP is a major academic event jointly organized by the Asia-Pacific chapter of the Association for Computational Linguistics (ACL) and the International Joint Conference on Natural Language Processing.
Notably, three of the accepted papers were led by undergraduate researchers from Soongsil University’s Natural Language Processing Lab, who served as first authors or co-first authors rather than being graduate students.
The undergraduate research team focused on analyzing practical problems caused by artificial intelligence large language models (LLMs) in real-world settings.
First, researcher Lee Si-hyeon found that when AI faces a difficult moral dilemma, it exhibits “omission bias”—leaving the situation unattended rather than taking direct action to resolve it. Lee also suggested that this bias could potentially be reduced by prompting AI to recall moral principles on its own before generating a response.
Researcher Chu Gyo-jun identified a gap between AI’s “inner thoughts and words.” Chu found that although the capabilities and information needed to solve a problem are well organized within AI’s internal structure, the accuracy of its answers drops significantly when it is asked to explain them in human language. The research clearly identified a gap between what AI knows and its ability to explain it.
Working with researchers at the University of Tokyo, researcher Cho Seong-hyeon developed a safety system called POLAR that assesses in advance, before an AI agent carries out a user’s command, whether the action can later be canceled or reversed. The technology is designed to block dangerous commands before they are executed, such as accidentally deleting important data or making an irreversible payment.
Professor Park Chan-jun, who supervised the research, said, “It is especially significant that four papers from our team were accepted at AACL-IJCNLP 2026, three of them undergraduate-led studies,” emphasizing that “it is important for students to gain experience defining research problems and designing experiments on their own from their undergraduate years, and completing research through collaboration with researchers in Korea and abroad.” Professor Park added, “Since all three students plan to continue their research in graduate school, I hope this achievement will mark the beginning of their long-term research growth.”
Meanwhile, researchers Lee Si-hyeon, Chu Gyo-jun and Cho Seong-hyeon plan to enter graduate school at Soongsil University after completing their undergraduate studies and continue their research in natural language processing under Professor Park Chan-jun’s supervision.

[email protected] Kim Man-gi Reporter