Korea AI Basic Act: who must label AI video, provider or user?

The sources I read say Korea's AI Basic Act puts its penalties on AI developers and service operators, not on private users, broadcasters or publishers. A business that renders video through an API sits in a grey zone those summaries do not settle, so decide with counsel whether you are acting as a service operator toward Korean users.

This is based on Korea JoongAng Daily's explainer, its report on watermark rules and enforcement gaps and Shin & Kim's update, all read on 2026-10-02. None of them is the statute, and none discusses API customers by name.

Who does the law name?

JoongAng Daily quotes the principle that responsibility rests with AI developers and service operators, not with private users, broadcasters or publishers. A second piece says the rules apply only to businesses, not individual users.

Foreign providers must appoint a local representative only if they pass a threshold: 680 million dollars in global revenue, 10 billion won in domestic revenue, or 1 million daily Korean users, per JoongAng Daily. The paper notes this effectively limits the requirement to the largest foreign firms.

Where does an API customer land?

If you only use the output yourself, you look like a user. If you build a product that lets Korean users generate or receive AI video, you look like a service operator, and the labeling and advance-notice duties Shin & Kim lists would be yours. The sources do not draw this line, so it is a question for counsel, not for a blog.

A useful working rule is to ask whether your own customers would say that you generate content for them. If yes, plan as an operator.

What would labeling look like in a Sume pipeline?

Shin & Kim says ordinary outputs may use a human-recognizable label or a machine-readable method such as a watermark, while deepfakes need clearly recognizable labels. The visible route is within reach of a Sume job: Video captions burns authored cues into a finished clip.

Sume does not offer a machine-readable watermark step in the docs I read, and Sume's docs describe what each tool returns and say nothing about a watermark or embedded provenance record on outputs, so treat the output as unmarked until you have checked the delivered file yourself. Keep job ids from Jobs and results so you can show which clips came from which request, and note that the Sume basics page describes where Sume's primary path stands today.

What do enforcement gaps mean for you?

JoongAng Daily reports that watermark removal tools are easy to find, and that one app with an 'AI eraser' feature has been downloaded over 500,000 times. It also notes that deepfakes made with foreign apps largely escape the law's jurisdiction.

That is a description of the regulator's problem, not a safe harbor for you. A business with Korean customers is inside the jurisdiction, and the grace period only defers penalties. If a competitor removes labels and gets away with it, your own exposure does not shrink.

The reputational side also stands apart from the statute. A platform or a client can ask where a clip came from, and 'we stripped the label' is a poor answer. A visible notice plus a job record is a defensible position whatever the enforcement rate turns out to be.

Questions to settle before launch

Use this as a working list and adjust it to your own pipeline and counsel's advice.

• Do Korean end users receive AI-generated video from your product?
• Who gives the advance notice that generative AI is used?
• Where does the visible label live, and does it stay on screen throughout a deepfake-like clip?
• Do you meet a local-representative threshold?
• Who keeps the records if the ministry asks?