South Korean experts urge AI Act revisions: clarify supplier-operator liability and adopt C2PA with invisible watermarks

As South Korea's Artificial Intelligence Framework Act enters its first year of enforcement following its early 2026 implementation, industry experts are calling for adjustments to the law's liability structure and technical execution methods. At the 63rd Breakfast Forum hosted by the Korea AI Industry Association (AIIA) in Seoul on October 20, Park Sang-chul, Associate Professor at Seoul National University Law School, and Kim Dong-hwa, Executive Director at MarkAny, presented recommendations on legislative issues and generative AI content trustworthiness.

Park highlighted a mismatch in high-impact AI regulations where the entity bearing legal obligations often differs from the actual system operator. For instance, startups supplying AI systems to public institutions may be assigned impact assessment or disclosure duties but cannot effectively perform them on behalf of the operating agencies. He argued that shifting all burdens to operators could result in excessive regulation across industries using AI, suggesting a need to align obligations with the specific roles of developers and operators.

Regarding AI-generated content labeling, both experts noted that visible watermarks are easily removed via image editing, rendering them ineffective. Kim proposed a shift towards recording provenance at the generation stage by combining C2PA (Coalition for Content Provenance and Authenticity) standards with invisible watermarks. C2PA is a technical specification that records the creator and modification history; even if metadata is stripped during distribution, invisible watermarks embedded within the content can assist in tracking origin.