Korea AI label tiers: invisible watermark vs visible for deepfakes

Korea's AI Basic Act sets two tiers of AI labeling. Content that viewers can easily tell is artificial, such as animation or webcomics, may use an invisible digital watermark, while deepfake content that closely resembles real people or events must carry a clear, visible label. For a video, the visible label has to be shown throughout, per one report.

This is from Korea JoongAng Daily's watermark report, its explainer and Shin & Kim's update (all read 2026-10-02). They are press and law-firm summaries, not the decree.

What are the two tiers?

Shin & Kim says standard outputs may use either a human-recognizable label or a machine-readable method such as a watermark, while for deepfakes only labeling clearly recognizable by users is permitted. JoongAng Daily states the visible-for-deepfakes rule by format: audio needs an announcement, images need a visible watermark, and video must display a watermark throughout.

How do I decide which tier a clip falls in?

Ask the viewer's question: would an ordinary person think this is a real person or event? A photoreal avatar presenting a product, or a cloned voice, leans toward the deepfake tier. A stylized cartoon mascot does not. The sources do not give a test beyond 'closely resembles real people or events', so when in doubt, choose the visible label.

The cost of over-labeling is a small on-screen notice. The cost of under-labeling is, after the grace period, a corrective order.

How do I keep a label on screen throughout?

A burned-in element is the one that survives every export. With Sume, Video captions burns cues with text, start and end, so you can pass one cue that spans the whole clip, for example start 0 and end equal to the clip length. Timeline compose is the other route: it keeps one still on screen for the entire video layer, up to a 300 second ceiling.

What this does not give you: a machine-readable watermark. Sume's docs do not describe one, 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.

What about mixed clips?

Real clips mix styles: a cartoon host over a real product shot, or an animated intro before a photoreal presenter. The sources give no rule for mixed content, so the safe reading is to apply the stricter tier to the whole clip if any part could pass as a real person or event.

Audio needs separate thought. JoongAng Daily says deepfake audio needs an announcement. A clip with a cloned or realistic synthetic voice over unrelated visuals may therefore need a spoken line even when the picture is clearly artificial. Check the voice before you check the picture.

Finally, remember the grace period. The tiers describe what will be expected once penalties apply, and the ministry has said enforcement is guidance-led for at least a year. That is time to test your label placement on real devices, not a reason to skip it.

Practical steps

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

• Set end to the real clip length in seconds, read from a probe.
• Keep the notice out of the platform's UI safe zones.
• Re-check the delivered MP4 on a phone before you publish.
• Keep the job id as your record of how the label was added.

curl -X POST https://api.sume.com/v1/video-captions \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: ai-notice-001" \
  -d '{
    "video_url": "https://example.com/clean.mp4",
    "cues": [
      { "text": "AI-generated", "start": 0, "end": 20 }
    ]
  }'