South Korea is taking a harder line on AI-generated content than anywhere else in the world. A new bill would require AI service providers to embed permanent watermarks - codes, letters, or symbols - directly into every piece of generated content. Not platform labels. Not metadata. Actual watermarks baked into the files themselves.
The legislation targets a genuine problem: under current rules, AI providers only need to notify users with captions or icons in their interfaces. Screenshot the image, download the audio, copy the text, and those labels disappear. The content can then circulate online "as if it were created by a human or depicted real events," as the bill's sponsors put it.
The proposed solution is technically ambitious. Watermarks embedded in images would survive editing and compression. Audio watermarks would persist through format conversion. Text watermarks would remain even when copied and pasted. And anyone who removes, alters, forges, or damages these watermarks faces up to two years imprisonment or fines of 20 million won (about $13,500).
This is where things get complicated. The technology for watermarking images is reasonably mature - several AI companies already embed invisible markers in their output. Audio watermarking exists but is more fragile. Text watermarking is still largely experimental, with most techniques either too brittle (they break when you rephrase anything) or too detectable (they make the text sound weird).
So South Korea is essentially mandating a technology stack that only partially exists yet. That's bold, and it might push innovation. Or it might just push AI companies to avoid the Korean market entirely.
I appreciate what they're trying to do. The flood of AI-generated content with no way to identify it as such is a real problem, especially for news, political discourse, and anything where authenticity matters. But requiring watermarks that survive hostile removal attempts is an arms race, and arms races in technology tend to favor the attackers.
The technology is impressive - when it works. The question is whether anyone can build watermarks robust enough to meet these requirements without breaking the content itself.
