Trustworthy and Responsible Agentic AI: SNU CTAI Symposium

On July 6, the Center for Trustworthy AI (CTAI) at Seoul National University (SNU) hosted the Trustworthy and Responsible Agentic AI Symposium at Oakwood Premier COEX Center. Held as an official satellite event of the 2026 International Conference on Machine Learning (ICML), the symposium brought together researchers and industry leaders from Korea, the United States, and Japan to examine the challenges posed by agentic AI. The event unfolded across two sessions moderated by Kim Yoon, President and Chief Strategy Officer of TwelveLabs, an AI video intelligence platform. The symposium also featured a poster session showcasing CTAI's ongoing research, along with live AI interpretation for attendees.

Attendees looking around at ICML research materials

CTAI Director Lee Eun Ju, a professor in SNU's Department of Communication, opened the event by framing trust as something that must be maintained across three pillars: technology, law, and people. "Trustworthiness and responsibility can no longer be afterthoughts," Professor Lee said, urging participants to listen generously and engage with differing viewpoints.

Rethinking Human Agency and Market Power

The first session was opened by S. Shyam Sundar, Evan Pugh University Professor and Director of the Center for Socially Responsible AI at Penn State University, with his presentation "Human Agency in the Age of Agentic AI." Professor Sundar warned that agentic systems risk deepening problems of over-trust and under-trust in AI by shifting interactions away from back-and-forth dialogue toward outright delegation, reducing the need for users to stay engaged at all. Drawing on his Human-AI Interaction Theory of Interactive Media Effects (HAII-TIME), the professor proposed the Four Cs framework, which categorizes the roles of AI in communication as creator, converser, curator, and co-author. AI can serve as a mediator, helping people find common ground in democratic deliberation but can just as easily reinforce conspiracy beliefs.

The second talk was delivered by Ko Haksoo, Professor at SNU School of Law and former Chairperson of South Korea's Personal Information Protection Commission. Professor Ko examined how autonomous agents that search, compare, negotiate, and execute transactions on behalf of users are reshaping market structures. He described a concrete scenario: a user asks a travel agent for a flight from London to New York. The travel agent queries separate flight, hotel, and ground transportation agents, which in turn retrieve information from multiple services using standardized protocols such as the Model Context Protocol (MCP), an open standard that enables AI agents to communicate with external tools and data sources. The entire chain runs autonomously with the user receiving only the final result. Ko cautioned that agents must accurately interpret and act on user preferences, while users must clearly communicate what they want and remain willing to review what the agent has done to keep costs low across the whole chain.

Professor Ko presenting on Market Dynamics in an Agentic AI Economy

Algorithms and Industry Strategy

The second session shifted to technical and strategic dimensions. Kim Youchul, Head of Strategy and Vice President at LG AI Research, spoke on the topic of "LG AI Research's Approach to Trustworthy Agentic AI." He argued that as AI evolves from generating content to taking actions, safety must be built in multiple layers. As part of this layered approach to AI safety, he introduced K-AUT, the Korea-Augmented Universal Taxonomy, a risk classification framework designed to adapt global principles to the Korean context while also anticipating emerging risks associated with agentic AI. Developed in collaboration with UNESCO and informed by input from 50 in-service evaluators, the framework comprises 296 risk criteria. Kim then demonstrated how such risks can be systematically evaluated through an internal safety evaluation system. Using the system, only 21.2 percent of the cases evaluated were classified as truly safe, underscoring the need to assess agentic AI not only by task completion but also by safety, reliability, and accountability.

Kim explaining LG AI Research's agent risk taxonomy

Sung Nako, Executive Vice President and Head of Hyperscale AI at Naver Cloud, led the closing discussion. When asked about the tension between adding safety layers and preserving user convenience, panelists argued that autonomy should remain the core design principle, but the appropriate degree of autonomy depends on the domain. In healthcare, for instance, full autonomy is unacceptable, and human decision-making must intervene.

On accountability in complex multi-agent systems, panelists suggested that if an AI agent causes harm and no party can cover the consequences, the agent's capabilities should be scaled back. The discussion also touched on multi-agent governance itself. While multiple agents can theoretically correct errors and prevent mistakes, the risk of groupthink and group polarization means collective deliberation can become its own source of danger.

A Discourse on Trustworthy Agentic AI Lies Ahead

Organizers and speakers of the symposium

Taken together, the symposium underscored that a trustworthy agent begins with a trustworthy model, secured through layered safeguards. Safety and performance must advance together as AI shifts from an era of providing knowledge to an era of taking direct action. Since AI capabilities are evolving faster than the technical, societal, legal, and policy frameworks designed to govern them, closing this gap will require coordinated efforts across academia, industry, and government. With this symposium being a starting point for the tougher questions ahead, further discussion will continue at ICML 2027.

Written by Yu Hee Young, SNU English Editor, [email protected]