A company that uses AI to produce text and images for its own work is not subject to the AI Basic Act's transparency duties, according to the official support desk. A company that builds someone else's generative AI into a product or service for customers must tell them, before they use it, that AI is involved. The Act calls the second kind of company an "AI business operator," and that includes companies that never built a model themselves. The statute sets a fine of up to 30 million won for skipping the advance notice, but the government is currently in a guidance period, and investigations and fines are postponed.
This reflects public sources as of October 26, 2026. I read Articles 2, 31, 40 and 43 of the Act and Article 23 of its Enforcement Decree on the Korean national law portal, the Ministry of Science and ICT's January 22 press release on its transparency guidance, and the Q&A pages of the AI Basic Act Support Desk. None of this is legal advice, and it describes no particular company or customer. I could not open the full PDF of the guidance itself, so I relied on the press release and the support desk's answers. The table and the four-step order later in the article are DMS.Labs proposals built from these sources. They are not statements in the law or in government guidance. The sources are Korean-language pages, and the translations of statute wording below are mine.
A wooden desk by a window with several printed sheets, a highlighter and an open notebookView original
Who do the transparency duties apply to?
They apply to AI business operators. Article 2, item 7 of the Act splits these into two kinds: AI developers, who develop AI and supply it, and AI user-operators, who use AI supplied by a developer to provide an AI product or AI service. Item 8 defines a "user" as someone who receives such a product or service.
The ministry's press release also describes the duty-bearer as an AI business operator that provides AI products and services directly to users. The support desk adds that a free or non-profit service still counts if it develops or uses generative AI to provide a product or service.
Read that way, one company can stand in two places. When staff use a generative AI tool to draft a report, the company is a user. When the same company attaches that AI to a chat function on its website and offers it to customers, it is a user-operator. That split is my reading of the text and not a ruling on any specific case.
Do companies that use AI-made text and images have to label them?
Under this Act, the support desk says no. It says companies that use generative AI as a work aid, for example to prepare reports or portfolios, have no labeling duty, and that YouTubers and individual creators who make films or videos with generative AI are generally "users" with no labeling duty either. It states outright that users have no duties under the AI Basic Act. The example in the ministry's press release points the same way: a studio that makes and distributes a film with a video-generation AI is not providing an AI product or service. It is a user that applied AI in its work, so it falls outside the duty.
One caveat comes with that. The support desk adds that a user who employs AI-generated material in a way that violates other laws, such as the Information and Communications Network Act, the Public Official Election Act, the Youth Protection Act or the Act on Fair Labeling and Advertising, is sanctioned under those laws. Having no duty under the AI Basic Act does not mean anything made with AI may be used freely. This article does not cover those other laws.
What must you tell customers when you offer an AI feature?
You must tell them that AI is involved before they use the product. Article 31, paragraph 1 requires an AI business operator that provides a product or service using high-impact AI or generative AI to notify users of that fact in advance.
Article 23, paragraph 1 of the Enforcement Decree allows four methods. You can state it on the product itself or in a contract, user manual or terms of service. You can show it on the user's screen or device. You can post it, in an easy-to-notice way, at the place where the product is provided. Or you can use another method the Minister of Science and ICT recognizes.
Two further duties sit alongside the advance notice. Paragraph 2 requires output from generative AI, or from a service built on it, to be marked as generated by generative AI. The Decree allows either a method people can perceive or a machine-readable one, and if the machine-readable route is used, the fact must also be announced at least once through a notice or audio. Paragraph 3 covers virtual sounds, images and video that are hard to tell from real ones: they must be disclosed or marked in a way the user can clearly recognize. If the output is an artistic or creative work, the marking may be done in a way that does not spoil the viewing.
On how visible a mark must be, the support desk says a logo or similar sign somewhere on a chat screen is enough, and a separate watermark on every reply is not needed. A mark that an ordinary user can reasonably notice in size and color will do, and it says text such as "AI" or a symbol can serve.
Article 23, paragraph 4 of the Decree also lets these rules be set aside in some cases: when the use of AI is already obvious from the product name or on-screen wording, when the AI is used only for the operator's internal work, and in cases the Minister sets by public notice.
How do these cases read in practice?
The table below is a DMS.Labs aid for matching the provisions above to common situations. It is a way to choose which provision to open first, and it does not give rulings.
Scroll horizontally to view a wide table.
| Situation | Where to look first | Direction it points |
|---|---|---|
| Staff make reports and images with AI for their own work | Support desk Q&A | User side; the support desk says the Act's transparency duties do not apply |
| AI is used only to summarize internal documents | Decree Art. 23(4) | Internal-use-only cases may be exempt |
| You add a generative AI chatbot to your website | Act Art. 31(1), Decree Art. 23(1) | Appears to call for advance notice before use |
| Your service generates text or images and hands them to customers | Act Art. 31(2) | Appears to call for marking the output |
| You provide virtual images or video that are hard to tell from real ones | Act Art. 31(3) | Falls under clear notice or marking |
| You use AI in hiring or loan screening | Act Art. 2, item 4 | Check separately whether it is high-impact AI; outside this article |
When do fines apply, and how long does the guidance period last?
Article 43, paragraph 1 names three cases that carry a fine of up to 30 million won. One is failing to give the advance notice in Article 31, paragraph 1. Another is failing to appoint a domestic representative under Article 36, paragraph 1. The third is failing to obey a stop order or corrective order under Article 40, paragraph 3.
The marking duties in paragraphs 2 and 3 have no such direct entry. Instead, Article 40, paragraph 1 lets the Minister collect documents or investigate when a violation is suspected or reported, and paragraph 3 lets the Minister order a stop or correction once a violation is found. Ignoring that order falls under the third case in Article 43. That link between the articles is my own reading. A July Yonhap News TV report quotes a vice minister saying a violation of the marking duty carries a fine of up to 30 million won, and I could not confirm which route he meant.
The guidance period is a ministry policy. The January 22 press release says that under a policy of running a guidance period of at least a year, fact-finding and fines are put off for the transparency provisions as well. The support desk says the postponement is one year and may be extended after that, depending on how ready industry is. The Act took effect on January 22, 2026, so one year ends on January 22, 2027. Because an extension is possible, I would not treat that date as a settled deadline. The same page says that for companies that fall short of the guidance, the government plans to offer explanation and consulting through the support desk instead of fines.
I also need to state what I could and could not check in the statute. The Act's pages I read were labeled as the version in force from January 22, 2026 (Act No. 21311), and the Decree was the version in force from August 20, 2026. Reports say follow-up amendments took effect on July 21, but I did not compare the articles one by one against those amendments.
What would this look like for an eight-person interior design office?
Here is a hypothetical. An interior design office has eight employees. It is an imagined situation and not a real customer or site. The office uses generative AI in two places. Staff make draft images with AI to send to clients, and the website carries an AI chat window that takes inquiries about visits.
A reception desk with a tablet stand, a small card and a potted plant at an office entranceView original
The draft images are a work aid, so under the guidance above they read as the user side, with no labeling duty under this Act. If a draft is made to look like a finished photo of completed work, other laws such as the labeling and advertising act could come into play, and this article does not judge that. I suggest sending drafts labeled as drafts.
The chat window is different. It puts an outside AI in front of customers as a service, so the office is a user-operator, and customers should see that it is AI chat before they start talking. Writing "AI chat" at the top of the window or in the terms of service uses one of the methods the Decree opens up. While the conversation runs, a visible AI mark somewhere on the screen is enough, as the support desk describes.
What should you sort out first?
DMS.Labs suggests four steps. First, list the AI your company uses in two groups: what stays inside the company and what customers touch. Next, pick out each screen on the customer side that contains generative AI. Chat windows, automatic replies, automatically produced documents and AI images belong here. Then decide where the advance notice will go for each item: the terms of service, the on-screen wording or a posting at the premises will do. Finally, write down the date you checked and the version of the law you read. Amendments and an extended guidance period are both possible, so you will want to know later which version you relied on.
For hard cases, the support desk lists consultation channels by topic. Where you need a legal conclusion, check it with a professional.
What to read next
For ways to attach AI to everyday work and customer services, see the DMS.Labs automation page, which is provided in Korean. The limits of watermarks and detection tools for AI output are covered in SynthID Detector Is Now Open to All, and a way to hand automation permissions out in stages is in Giving AI automation permission in three steps. If you are considering adding an AI feature to a customer service, you can describe your situation on the contact page.
Sources consulted
Checked on October 26, 2026. Laws can be amended. The two photographs in this article are AI-generated illustrations, not records from a real site. All sources below are in Korean.
- National Law Information Center, Framework Act on the Development of Artificial Intelligence and Establishment of Trust (AI Basic Act), Article 2: web page
- National Law Information Center, same Act, Article 31 (transparency duties): web page
- National Law Information Center, same Act, Article 40 (fact-finding): web page
- National Law Information Center, same Act, Article 43 (fines): web page
- National Law Information Center, Enforcement Decree of the same Act, Article 23 (transparency duties): web page
- Ministry of Science and ICT, press release on the AI transparency guidance (January 22, 2026), copy on the KDI Economic Information and Education Center site: web page
- Same press release, copy on ITS Korea: web page
- AI Basic Act Support Desk, consultation topics (generative AI transparency): web page
- Yonhap News TV, report on watermark duties and the high-impact AI management system (July 21, 2026): web page
