South Korea enacts ‘world-first’ AI law requiring watermarks and risk tests

South Korea has launched what is being billed as the most comprehensive set of AI laws anywhere in the world after an AI basic act took effect on Thursday last week, a move the government says is central to its bid to become a leading global AI power. The law forces companies to label AI-generated content: invisible digital watermarks are required for clearly artificial outputs such as cartoons or artwork, while realistic deepfakes must carry visible labels.

Operators of “high-impact AI” systems used in areas like medical diagnosis, hiring and loan approvals must conduct risk assessments and document how decisions are made, and extremely powerful models will require safety reports — though officials acknowledge the threshold for those reports is so high that no models currently meet it.

Companies that violate the rules face fines of up to 30m won (£15,000), but the government has promised a grace period of at least a year before penalties are imposed. Government officials say the law is 80–90% focused on promoting industry rather than restricting it and have described the framework as a flexible, principles-based, “trust-based promotion and regulation” approach.

The legislation has drawn pushback from both startups and civil society.

south korea ai basic act, ai basic act, labeling ai-generated content, invisible digital watermark, visible deepfake labels, digital watermark requirement, high-impact ai systems, risk assessments for ai, ai decision documentation, safety reports for models, 30m won fines, grace period enforcement, ai startups unprepared, startup alliance, lim jung-wook, alice oh, kaist, minbyun human rights lawyers, regulatory blind spots, human involvement exemption, enforcement decree criticism, security hero report, 53% deepfake victims, telegram deepfake chatrooms, trust-based promotion regulation, melissa hyesun yoon, hanyang university