South Korea just became the first country in the world to pass comprehensive AI safety legislation. While everyone else is still arguing about frameworks and principles, Korea shipped actual law.
The "Basic Act on the Development of Artificial Intelligence and the Establishment of a Foundation for Trust" went into effect January 23, 2026. The name is bureaucratic, but the implications are significant: this is the first national regulatory framework specifically targeting AI-related harms like disinformation, deepfakes, and algorithmic discrimination.
Here's what it actually does: The law creates a "high-risk AI" classification for models that affect people's safety or lives. These systems must warn users they're AI-generated and hold developers accountable for harms. All AI-generated content must include watermarks—a baseline transparency requirement that sounds obvious but no other country has mandated at scale.
The government gets authority to investigate violations and impose sanctions. More interestingly, major international AI firms have to establish local representatives if they meet certain thresholds: $681 million in annual revenue, 10 billion won in domestic sales, and over 1 million daily users in Korea.
Right now, only Google and OpenAI meet those thresholds. That's deliberate—the law targets the major players without drowning smaller companies in compliance costs. Non-compliance carries fines up to 30 million won, which sounds modest but represents a enforcement mechanism that can scale.
What makes this different from EU or US approaches: The EU's AI Act is comprehensive but isn't fully in force yet. The US has executive orders and voluntary commitments but no federal legislation. China has regulations but they're focused more on content control than safety. Korea's law is narrower in scope but actually operational.
Critically, the law includes provisions promoting AI development alongside safety requirements. The government must update policies every three years to balance innovation with harm prevention. This isn't a ban—it's a framework that says you can build AI, but you're responsible for what it does.
Watermarking can be stripped. High-risk classifications will be fought over. International coordination remains a mess—what Korea requires doesn't bind OpenAI's operations elsewhere. And fundamentally, regulating AI capabilities we don't fully understand is really hard.
