SurfingBearHome › AI News › AI Policy & Regulation
Korea’s AI Framework Act subordinate legislation takes effect — what to prepare for generative AI labelling and high-impact AI management
The enforcement decree and subordinate notices have been running since 21 July. An opportunity — preferential treatment in public procurement — arrived alongside genuine confusion about how far the labelling duty actually reaches.
The AI Framework Act (formally the ‘Framework Act on the Development of Artificial Intelligence and Establishment of a Foundation of Trust’) took effect on 22 January 2026, and with its subordinate legislation — the enforcement decree and associated notices — in force since 21 July, Korea’s AI legal framework is now operating in earnest across industry.
This second wave has two strands. One puts detail behind the generative AI labelling duty and the high-impact AI management regime. The other expands industrial support measures — public procurement, startups, workforce development, and assistance for underserved groups. Regulation and support switched on the same day.
The opportunity: AI products get preference in public procurement
The government will require state bodies to give priority consideration to AI products when procuring goods and services, and is introducing an ‘AI product and service verification scheme’ under which the Minister of Science and ICT confirms whether AI is actually used.
- Companies holding a certificate get relaxed eligibility requirements for multiple-supplier contracts.
- They can receive additional points in qualification screening for lump-sum contracts.
Alongside this, a legal basis was established for national and local government to subsidise the cost of AI services for people who struggle to access them — people with disabilities, older people, women returning to work, and jobseekers. Training and employment support for AI specialists, startup support through the fund-of-funds, and standards for supplying public data for AI training are all being advanced together.
The burden: how much AI makes something ‘generative AI output’?
Major companies including Naver, Kakao, SK Telecom, KT, LG Uplus and Upstage are responding by building AI safety organisations and risk management systems, and by expanding labelling of generative AI output.
Confusion within industry continues, though. Two issues have been reported.
- How much generative AI use triggers the labelling duty — the threshold is not clear.
- Whether the same standard applies when a human has edited or reviewed the output — the point that snags most often in practice.
It is also unclear which services in healthcare, finance, recruitment and education qualify as high-impact AI, raising concerns that companies may have to spend more on legal review and risk assessment before launching a service. It has also been noted that for startups, with less staff and capital than large companies, compliance cost can act as a barrier to market entry.
A grace period of at least a year
The government plans to run a grace period of at least one year while it gathers industry feedback and puts detail into guidelines covering high-impact AI cases and generative AI labelling methods. In other words, this is not the moment to comply with a finished standard — it is the moment to understand your own internal position while the standard is being written.
For context, major jurisdictions are moving in different directions. The US and Japan are weighted towards industrial growth and technological innovation, while the EU is phasing in an AI Act that assigns obligations according to the risk level of the AI system.
What this means for Korean teams — SurfingBear editorial
For a team adopting international AI products, what this actually changes is two items on your evaluation checklist.
First, any tool whose output leaves the building — content generation, customer-facing copy, video and voice generation — should be evaluated on the assumption that it may fall within the labelling duty. The fact that human review may or may not exempt you is precisely the point: you cannot currently rest an argument on ‘we reviewed it, so we are fine’.
Second, AI used in recruitment, healthcare, finance or education needs a high-impact determination first. Among the products SurfingBear lists, those touching recruitment assessment or medical data require that judgement up front, and we flag it on each product page under ‘what to check before adopting in Korea’.
In terms of sequencing, compliance work is not the first step. Inventorying the AI tools already in use inside your company is. Labelling duty or high-impact determination, you cannot start without knowing what the subject actually is.
This article is not legal advice. Whether an obligation applies depends on how your service is structured, and needs legal review.
Related event
The law firm DLG is running a practical compliance guide seminar on the AI Framework Act at Dreamplus Gangnam on Tuesday 25 August 2026 (Lawtimes report). Doing the internal AI tool inventory first makes the questions you bring to something like this far more specific.
Sources
- Yonhap News, “AI Framework Act subordinate legislation takes effect… a test for Korea’s AI industry”, 2026.07.21 — read the original
- Korean Law Information Center, “Framework Act on the Development of Artificial Intelligence and Establishment of a Foundation of Trust” (AI Framework Act) — statute text
- Lawtimes, “DLG: practical compliance guide seminar on the AI Framework Act”, 2026.08 — read the original
The regulatory detail and quotations above are compiled from the reporting and statutory information cited. The ‘What this means for Korean teams’ section is SurfingBear editorial interpretation, is not a claim made in the original reporting, and does not constitute legal advice. Content may change as guidelines are made more specific during the grace period.
To start with an inventory of your internal AI tools
The 2-week AX assessment starts by establishing which tools are used where in your current operations. Compliance or automation — either way, the list comes first.