Industry Field's 'Application Standards' on the Proving Ground
On January 22, 2026, the 'Framework Act on the Development of Artificial Intelligence and Establishment of Trust Foundation,' the so-called AI Basic Act, which unifies the growth and safety of the AI industry into a single framework, came into effect.
The significance of this implementation lies not in strengthening regulation but in organizing the uncertainty that has repeatedly arisen in the generative AI proliferation phase and presenting minimum operating standards that industry and society can commonly reference.
The government describes this as a three-pronged framework of 'promotion, trust, and minimal regulation,' emphasizing a design that prioritizes predictability over punishment.
The first effect of the AI Basic Act is the fixation of governance. With the national-level AI strategy committee system, policy center designation, safety research institute operations, and a unified training data provision system being structured through the act and enforcement decrees, industry has for the first time encountered an environment where investment and development decisions can be made on the premise of "how far the state supports, and from where it checks." This carries a clear policy signal in that industrial infrastructure was presented before individual technology regulation.
The substantive content of regulation is compressed into three branches. The first is transparency. Services utilizing generative AI or high-impact AI must notify users in advance of AI usage, and results with concerns about major social side effects such as deepfake-type outputs must carry markings that users can clearly recognize. With invisible watermarks also permitted depending on content types such as webtoons and animation, the design centered on 'responsibility to inform' rather than blocking technology has become clear. However, since the level and method of marking is broadly open, the practical issue most likely to be encountered in the field first will be 'what constitutes sufficient notification.'
The second is high-impact AI management. The case of AI being used in areas that can significantly affect life, safety, and fundamental rights becomes the judgment criterion, with sensitive sectors such as medical, energy, transportation, finance, employment, education, and public services as the main subjects of review. The direction of policy here is clear. Even if AI makes recommendations or classifications, the risk burden is evaluated as low when humans perform final decision-making. Accordingly, from a corporate perspective, how to design 'human intervention' and how to prove this through logs and documents emerges as the core element of compliance.
The third is the obligation to ensure safety for ultra-high-performance AI. With a structure combining quantitative criteria such as a cumulative computation level of 10 to the 26th FLOPs and qualitative requirements such as concerns about widespread impact on fundamental rights, this is closer to a safety framework in preparation for future ultra-large foundation model competition than something applied to many domestic companies right now. The government also emphasizes that in reality no model currently meets the relevant criteria, clearly indicating its 'preemptive safeguard' character.
In industry, concerns that the ambiguity of criteria could impede innovation speed and evaluations that predictability has improved thanks to the minimal regulation principle are mixed. In civil society, criticism is also raised about the lack of strong disciplines such as remedies or prohibited AI. In response, the government has chosen to cushion the shock by announcing a minimum grace period of one year or more, guidance-centered operations, factual investigations only in exceptional cases, and consultation provision through dedicated support desks.
Ultimately, the first day of the AI Basic Act is closer to a day when the state presented the basic grammar of how to operate AI and secure trust, rather than a day when regulation began in earnest.
The key going forward is the boundary-setting of high-impact AI, the effectiveness of generation marking methods, and the refinement of procedures that companies can comply with without excessive costs. The act went into effect today, but the accumulation process of first applications without precedent is likely to determine the success or failure of the system over the next year.
