A bill to amend the AI Framework Act (인공지능 발전과 신뢰 기반 조성 등에 관한 기본법) moved from committee referral to committee review on 2026-09-08. As drafted, it would require an AI business that generates synthetic audio, image, or video output to embed a visible or invisible watermark — code, characters, or symbols inserted into the output itself — disclosing that the content was AI-generated, and it would separately prohibit removing, forging, or altering that watermark. This is a pending bill, not current law; no effective date has been set.

What the draft changes

The current AI Framework Act already requires an AI business to notify or label AI-generated audio, image, or video output so that a user can clearly recognize it as such. The drafters' stated concern is that, because the law does not specify *where* that disclosure must sit, content can be reprocessed or redistributed in a way that strips the AI-origin notice before it reaches the end user, letting synthetic output circulate as if it were human-made or real. The amendment addresses this by moving the disclosure into the content itself: a watermark embedded in the output, rather than a label attached alongside it, plus a new prohibition on tampering with that watermark (proposed Article 31(3)–(4) and Article 42(2)).

Who would be affected: AI businesses that generate synthetic audio, image, or video output, and separately, any party that damages, forges, or alters the required watermark on such output.

Penalty as drafted: up to 2 years' imprisonment or a fine of up to KRW 20 million for violating the watermark or anti-tampering requirements.

What this means for you

  • Do not treat this as in force. The bill is at the committee-review stage; no effective date has been set, and the text can still change before (or instead of) passage.
  • If you generate synthetic audio, image, or video output for the Korean market, check your current AI-content labeling approach against the draft. A separate on-screen or metadata label may not satisfy an embedded-watermark requirement if this passes as drafted.
  • Track technical feasibility now. Embedding a visible or invisible watermark directly into generated output (rather than as an attached label) is a different engineering task than current disclosure practices for many providers; assess lead time before the bill advances further.
  • Watch for the anti-tampering provision separately. It would apply to any party removing or altering the watermark, not only to the AI business that generated the content — relevant if you redistribute or reprocess AI-generated media from third parties.
  • Re-check status before shipping a new AI-generation feature into Korea, since committee review can move to a vote, stall, or be folded into a broader revision.

Source: (no URL provided in source data)