AXHUB5 cards

The AI Framework Act,
what should our team prepare?

In effect since January 22 this year.
Fines are suspended during the grace period,
but the obligations themselves already exist.
Now is the time to prepare quietly.

Prepared by cross-checking the statutory text (Korea Law Information Center) against law-firm and consulting commentary

AI Framework Act1 / 5

01

Does it apply to us too?

There's no size threshold.
It applies if you provide users a product or service containing high-impact AI or generative AI.
Overseas providers count too, if they affect domestic users.

Chatbot support, AI recommendations, generated images — if it's a feature users touch, it's a candidate.

A one-line self-check "Among the features users directly touch in our service, is there anything AI generates?"
If yes, the rest of this deck applies to you.

Source AI Framework Act (in effect 2026-01-22), scope of application — confirmed against statutory text and law-firm commentary

AI Framework Act2 / 5

02

First duty — tell them in advance

Before a user uses it, you must let them know it runs on AI (Art. 31(1)).
Break it and the fine is up to 30 million won.

It's nothing grand.
It starts with one line of notice on the screen.

Example notice Instead of hiding it —
"This answer is a draft generated by AI. For anything important, confirm with an agent."

Source AI Framework Act §31(1) prior-notice duty, fine up to 30 million won for non-compliance

AI Framework Act3 / 5

03

Second — mark that it's AI-made

Label the output of generative AI (Art. 31(2)).
There are two ways.
Visible — a logo or a line of text.
Machine-readable — a watermark or metadata.

For virtual audio, images, or video hard to tell from the real thing, go one step further — mark it so users clearly recognize it (para. 3).

Practical fork Ordinary generated output → text label or watermark, either works
Deepfake-type → the standard is "clearly recognizable," so visible is safer

Source AI Framework Act §31(2)(3) labeling duty for generated output and deepfakes

AI Framework Act4 / 5

04

High-impact means more accountability

Hiring and loan screening, healthcare, energy, student assessment —
areas that shape people's rights and safety are "high-impact AI."
If you fall here, explanation duties, impact assessment, and a management system are added (Art. 34).

There's some relief too.
The safety-assurance duty (Art. 32) is set at frontier training scale (10²⁶ FLOPs or more), so most companies aren't in scope.

Deciding question "Is our AI used to hire people, lend money, or judge illness?"
If not, card 4 is reference only; cards 2 and 3 are the main point.

Source AI Framework Act §34 high-impact provider duties · §32 safety duty (10²⁶ FLOPs threshold)

AI Framework Act5 / 5

05

Three things to do during the grace period

Fine enforcement is suspended (at least a year; actual fines expected after 2027).
But that's a reprieve on preparation, not on the obligation.

Law firms' advice converges on three.
A full inventory of where you use AI.
Notice and labeling reflected on screen.
Documenting internal standards, including cross-department collaboration.

Teams already halfway there If you built the "three-line rules" (data, review and accountability, records) from card No.8, the third item already exists. Just add notice and labeling on top.

Source Grace-period operating policy · the 4 common preparation recommendations from law firms and consultancies

AXHUBclosing

Regulation can be
a reason to get organized

In the end this law asks for two things.
Say that you're using it, and mark what it makes.

Fewer teams than you'd think can say right now where they use AI.
Even without the law, that list is worth making.

See the source material on axhub.net

Sources: statutory text (Korea Law Information Center) · law-firm and consulting commentary (2026) — checked 2026-07-05, details in the AXHub case library

AXHub card No.13 — the statutory provisions were checked against the original text, but this card is information, not legal advice. Individual matters need expert confirmation.