South Korea's AI Labeling Law: Who Has to Label, and Why Labels Get Lost After Upload
South Korea's AI labeling law already applies to companies that provide generative AI services, but not to the people who repost the output. The AI Basic Act took effect on January 22, 2026, and requires providers to tell users a service is generative AI and to mark its outputs. What happens to that mark after a file is shared again is still an open question in Seoul.
As of early November, two ministries were still debating whether to extend labeling to the distribution stage, and which law should carry it.
What does Korea's AI labeling law require today?
According to Korean press coverage of Article 31 of the AI Basic Act, a generative AI provider has two duties. It must give advance notice that a product or service runs on generative AI, for example in its terms or on screen. It must also mark outputs as AI-generated.
The method depends on the output. Deepfakes, meaning outputs hard to tell apart from real audio, images or video, need a label users can clearly see or hear. Other outputs, such as animation or webtoons, may use an invisible digital watermark or machine-readable metadata. According to Korean press coverage of the government guideline, a provider that chooses an invisible method should also tell users at least once, by text or voice, that the content is AI-generated.
The Ministry of Science and ICT says the duty falls on providers. People who use AI output in their own content or post it on social media do not carry a legal duty under the current act. The government is running a guidance period of at least one year before fines apply.
Why can a label disappear after a file is shared?
The current labeling duty covers only the moment of generation. Ajunews reported on November 1 that there is no duty to keep a label intact when a user uploads the output to another platform or someone else redistributes it.
Lawmakers reviewing the watermark bills in September described the same problem: marks being damaged or lost as content moves online. An invisible watermark only helps if someone downstream has a way to read it.
Who would be responsible at the distribution stage?
That is still being argued: the two ministries agree the label loss needs fixing but disagree on where the rule should live.
Position | Proposed approach | Status (as reported Nov 1, 2026) |
|---|---|---|
Korea Media and Communications Commission | Keep generation rules in the AI Basic Act; add platform duties to keep labels in the Information and Communications Network Act | No agreed draft |
Ministry of Science and ICT | Review generation-stage methods and distribution duties together, to avoid gaps between laws | Draft for the National Assembly not yet submitted |
Network Act amendment (Rep. Cho In-cheol, January) | Labeling duty for people who make or edit and distribute AI content; technical and managerial measures for platforms | Under continued review since September 15 |
A National Assembly subcommittee asked both ministries to bring a revised plan by early October. Ajunews reported that they had not finished talks by the end of October.
Our reading: whichever law wins, the practical question for rights holders and platforms will be the same. Can a label that was attached at generation still be read after two or three uploads? That is a technical question as much as a legal one, and it is worth testing before any rule arrives.
How does this compare with the EU and Taiwan?
The direction is similar: mark at the source, then worry about what survives distribution. We covered the EU AI Act Article 50 roles and Taiwan's three-layer labeling plan earlier. In Korea, the generation-stage duty is already in force while the distribution-stage duty is still a draft question.
What can a content team check now?
You can check whether your own files keep their provenance after the platforms you actually use. Upload a marked file, download it again, and see whether the visible label, the metadata and any invisible watermark are still readable.
Content Credentials (the C2PA standard) record origin and edit history in a file, and an invisible watermark can point back to that record if the metadata is stripped, a link known as soft binding. Contents Defence is on the C2PA Conforming Products List in image, audio and video, and Image Defence pairs an invisible watermark with C2PA issuance and verification for images.
What this article cannot tell you
This is a summary of Korean press reports and public statements up to November 13, 2026, not legal advice. The final shape of any distribution-stage rule, and which law it lands in, is not decided. We have not seen the ministries' internal drafts.
FAQ
Does Korea's AI Basic Act apply to individual creators?
Not under the current act. The ministry says the labeling duty applies to providers of generative AI products and services, not to users who post AI output.
Can an invisible watermark meet the labeling duty?
For outputs that are not deepfakes, yes, invisible watermarks or metadata are allowed. Press coverage of the guideline says users should also be told once by text or voice. Deepfakes need a visible or audible label, with a limited exception for artistic works.
When will fines start?
The government has set a guidance period of at least one year from January 22, 2026, before fines apply.
Want to see whether your images keep their provenance after upload? Request a Contents Defence review.