South Korea will become the first country in the world to enforce an “AI Framework Act”. From now on, AI businesses must label results produced by AI so users can recognise them. Fact-finding investigations will be conducted only exceptionally when there are fatalities, human rights violations or national damage.

The Ministry of Science and ICT said on Jan. 21 that the Framework Act on the Development of Artificial Intelligence and the Establishment of a Trust Foundation will take effect from Jan. 22. The European Union has previously drawn up the AI Act, which is similar in nature to the AI framework act, but it is scheduled to take effect in 2027, making South Korea the first country to enforce such a law.

The ministry said it placed emphasis on promoting the AI industry and, under a principle of minimum necessary regulation, minimised obligations and sanctions for AI businesses. It said wide-ranging measures to promote the AI industry were reflected.

AI businesses only required to label generative AI output

The ministry set out the requirement in an enforcement decree so companies can notify users in various ways that content is AI-generated. When AI-generated content is provided within a service environment, it allows flexible labelling distinct from cases where the content is taken outside the service.

When AI-generated content is provided only within a service environment, it can be flexibly labelled through a user interface or logo display. For example, interactive services such as chatbots will recognise notices before use or logo displays on the screen. Games and metaverse services will allow methods such as notices at login or marking characters as AI.

When AI-generated output is taken outside the service, more certain labelling must be applied. When AI-generated text, images or video are downloaded or shared, it must be labelled in a “human-recognisable way (visible or audible watermarks, etc.)”, or after providing text or audio guidance, a “machine-readable method (metadata, etc.)” must be applied to the AI-generated output.

Generative AI output and deepfake output that could cause social side effects must be labelled in a clearly recognisable way, taking into account factors such as the user’s age. For AI output such as animated webtoons that is not deepfake output, the rules allow not only visible methods but also invisible digital watermarks.

The AI framework act introduces a high-impact AI system. High-impact AI is limited to cases used in 10 areas specified by law. The areas are energy, drinking water, healthcare, nuclear power, criminal investigation, hiring, loan screening, transportation, public services and education, which are directly linked to life, safety and fundamental rights.

The enforcement decree specifies the criteria for determining high-impact AI and the responsibilities of high-impact AI businesses. The criteria take into account whether AI is used in areas specified by law and the severity of risks. At the same time, cases in which humans intervene in the final decision-making process are judged to be controllable and are excluded from high-impact AI.

The decree also sets out obligations to ensure safety. The aim is to prevent major damage, assuming a situation in which highly advanced AI becomes uncontrollable. To fall under the safety obligation, all of the following must be met: cumulative compute used for training of at least 10 to the 26 floating-point operations (FLOPs); application of cutting-edge technology; and a risk level that could broadly and significantly affect people’s fundamental rights.

The party responsible for transparency obligations was clearly defined as an “AI business” that directly provides AI products and services. Overseas businesses providing AI products and services to users in South Korea are also subject to the obligations. Users are not subject to the obligations because they are in a position to use AI products, services and generated content provided to them.

At least a 1-year grace period; fact-finding only when serious problems occur

The ministry will defer regulation for at least 1 year to minimise confusion for companies and give the field time to prepare. During that period, it will run fact-finding and administrative fine guidance periods. Fact-finding investigations will be conducted only in extremely exceptional cases, such as when serious social problems occur including fatalities or human rights violations.

Transparency and safety obligations, criteria for determining high-impact AI, and methods for fulfilling business responsibilities can be checked through a transparency guideline published on Jan. 21. The ministry plans to supplement the guideline during the grace period to resolve uncertainties.

The ministry will also operate a support desk made up of experts who participated in drafting subordinate regulations to help companies prepare for compliance. It is a channel for experts to provide detailed consulting to companies. To prevent the leakage of trade secrets, consultation content will be kept confidential and anonymous consulting will also be provided.

With the AI framework act taking effect, the National AI Strategy Committee will be converted from a body based on a presidential decree into a statutory committee based on law. The AI framework act set out the legal basis for the National AI Strategy Committee and the basis for establishing subcommittees within it, a support group and AI accountability officers.

To develop the industry, the enforcement decree specified legal support measures and procedures including AI research and development, building training data, supporting AI adoption and use, supporting startups, promoting AI convergence, securing specialised personnel, and supporting the construction of AI data centres.

Kim Kyung-man (김경만), director general for AI policy at the ministry, said, “As we are taking a path we have not been on before, we are trying to approach it at the minimum baseline of global discussions and establish minimal norms.” He said, “The AI framework act is not regulation, but a perspective of setting a benchmark to move forward globally.”