AI 治理概况
South Korea AI Governance Profile
韩国人工智能治理的初步观察概况,以具有约束力的人工智能基本法、透明度和高影响力人工智能职责、高级模型安全以及公共部门人工智能/数据管理法律层为中心。

概况速览
执行摘要
Key signals for South Korea AI governance, structured for quick institutional review before the full profile analysis below.
地区
Asia-Pacific
治理模式
Comprehensive AI law
监管成熟度
Advanced企业影响
High公共部门准备度
Advanced执法成熟度
Developing企业影响
企业应根据韩国法律定义,确定其系统是否属于生成式人工智能、高影响力人工智能或高级人工智能的范畴。适用的控制措施可能包括通知和标签、风险管理、文档记录、用户保护、人工监督、安全评估和证据保存。跨境服务提供商应审查其服务对韩国用户或市场的影响范围。
阅读完整企业影响分析关键机构
- 科学和信息通信技术部
- 根据《人工智能基本法》设立的国家人工智能治理机构
- 内政安全部
- 韩国标准和人工智能安全机构
观察清单
我们正在关注什么
可能改变治理格局的后续里程碑、尚未解决的监管问题及实施进展。
- 《人工智能基本法》下的执法实践和行政罚款
- 高影响力人工智能的认定和解读
- 透明度和合成内容标签实施
- 高级人工智能安全阈值实践
- 2026年8月28日之后公共部门影响评估的实施
- 第31条 透明度实践和劳动力能力要求
- MOIS public-sector AI ethics self-assessment uptake, training and framework revisions; keep this non-binding framework distinct from statutory impact-assessment obligations
- Follow-on Japan–Korea AISI multilingual evaluation, benchmarks and multimodal red-teaming outputs after the 8–9 October workshop reported on 2026-10-16
- National Assembly progress and official attachment details for the government AI Basic Act / Data Industry Act amendments approved by Cabinet on 13 October; do not treat proposed data provisions as enacted
执行摘要
《人工智能基础法》于2026年1月22日进入主要实施阶段,旨在建立国家人工智能发展和信任框架。该法包含针对特定生成式和高影响力人工智能应用的具有约束力的透明度义务、高影响力系统的义务、高级模型安全条款、制度机制和执法工具。部分修正案于2026年7月生效,而附属立法则进一步明确了实施细则。2026年8月28日,《人工智能和数据管理法》及其相关附属规则的修正案增加了重要的公共部门治理层。企业应明确其系统是否属于高影响力、生成式或高级人工智能的范畴,并将设计文档、人工监督、风险管理和透明度纳入部署流程。
治理架构
韩国的人工智能架构结合了具有约束力的国家人工智能框架法、附属规则以及独立的公共部门人工智能/数据管理层。第31条规定的透明度、第34条规定的高影响力人工智能义务、高级人工智能安全条款以及公共部门影响评估机制是不同的法律层面,不应被合并为欧盟式的单一风险分类体系。
MOIS’s 1 October public-sector ethics framework complements the statutory layer through six values, 68 self-check items and three tailored levels of assessment. Its voluntary self-assessment is not a new statutory duty.
主要政策与框架
| Policy | Issuer | 年份 | Status | Summary |
|---|---|---|---|---|
| 人工智能发展与信任建立框架法案 | Republic of Korea / Ministry of Science and ICT | 2025–2026 | Binding national AI framework statute | Core statute for AI development, trust, transparency, high-impact AI, advanced-model safety, institutional governance and enforcement. Separate national AI-data amendments reported by MSIT on 13 October remain government proposals; Cabinet approval does not establish parliamentary enactment. |
| 人工智能基本法实施令 | Republic of Korea | 2026 | Binding Presidential Decree / subordinate legislation — current version in force 2026-08-20 | Binding Presidential Decree, current version in force from 2026-08-20. Article 25 permits a business to request high-impact AI classification confirmation from the Minister of Science and ICT, submitting the product/service overview, development/training-data overview, materials demonstrating use process/results and other relevant documents. This confirmation procedure is distinct from Article 23 transparency, Article 27 high-impact business duties and Article 28 impact-assessment provisions. |
| 人工智能基本法第31条规定的透明度义务 | Republic of Korea | 2026 | Binding statutory transparency obligations | 涵盖特定生成式人工智能和高影响力用途的事先通知和披露义务,包括适用的合成媒体透明度。 |
| 《人工智能基本法》第34条 人工智能的重要职责 | Republic of Korea | 2026 | Binding high-impact AI duties | 要求相关运营者解决风险管理、可解释性、用户保护、人员管理或监督以及文档记录等问题。 |
| 人工智能基本法第35条影响评估 | Republic of Korea | 2026 | Statutory impact-assessment provision - wording and scope must be applied precisely | 提供影响评估机制;不要将其措辞视为与第 34 条规定的强制性义务相同。 |
| 高级人工智能安全条款和阈值规则 | Republic of Korea | 2026 | Binding safety and implementation rules where statutory thresholds apply | 现行法令材料对受保护的先进人工智能系统进行了定义,该系统使用至少 10^26 FLOPs 的累积训练计算能力,并满足其他最先进的和严重风险条件;但这并非对所有前沿人工智能的普遍定义。 |
| 人工智能和数据驱动型管理激活法案 | Republic of Korea / Ministry of the Interior and Safety | 2026 | Binding public-sector AI and data administration statute | 创建人工智能和数据管理的公共部门治理层,包括公共机构管理和影响评估机制,自 2026 年 8 月 28 日起生效。 |
| Public-sector AI ethics standards | Ministry of the Interior and Safety | 2026 | Final public-sector ethics framework / non-binding self-assessment guidance | MOIS announced final public-sector AI ethics standards on 1 October 2026: six core values, 68 detailed check items and a three-stage tailored self-assessment approach. The framework is practical, non-binding public-sector guidance and voluntary self-checking, not new general legislation or the binding AI Basic Act and Enforcement Decree. |
政策时间线
里程碑按上次审查时点分组,并保留来源日期的精确程度。
已记录的里程碑
2026-08-20
当前执行法令版本反映
Binding Presidential Decree, current version in force from 2026-08-20. Article 25 permits a business to request high-impact AI classification confirmation from the Minister of Science and ICT, submitting the product/service overview, development/training-data overview, materials demonstrating use process/results and other relevant documents. This confirmation procedure is distinct from Article 23 transparency, Article 27 high-impact business duties and Article 28 impact-assessment provisions.
2026-10-01
MOIS finalises public-sector AI ethics standards
MOIS announced final public-sector AI ethics standards on 1 October 2026: six core values, 68 detailed check items and a three-stage tailored self-assessment approach. The framework is practical, non-binding public-sector guidance and voluntary self-checking, not new general legislation or the binding AI Basic Act and Enforcement Decree.
2026-10-13
Cabinet approves proposed AI Basic Act and Data Industry Act amendments
MSIT reported Cabinet approval of government amendment proposals on national AI-data management and use. This backfilled legislative-process milestone does not establish National Assembly enactment or commencement. Detailed provisions and later legislative status remain unverified.
2026-10-16
Japan AISI reports first joint technical workshop with Korea AISI
Japan AISI reported the first Japan–Korea joint technical workshop on 16 October 2026; the event itself took place on 8–9 October with 30 participants. Topics included multilingual/multicultural evaluation, sectoral safety methods, JAI-Trust benchmarks and Korean multimodal red teaming. This is technical cooperation and testing capacity, not a new law, binding standard or memorandum.
企业影响
企业应根据韩国法律定义,确定其系统是否属于生成式人工智能、高影响力人工智能或高级人工智能的范畴。适用的控制措施可能包括通知和标签、风险管理、文档记录、用户保护、人工监督、安全评估和证据保存。跨境服务提供商应审查其服务对韩国用户或市场的影响范围。
Classification confirmation is a separate procedure: under Decree Article 25 a business may request confirmation from the Minister of Science and ICT with a product/service overview, training-data overview, evidence of use process/results and other relevant documents. Keep that classification request separate from Article 23 transparency, Article 27 high-impact business duties and Article 28 assessment provisions.
观察站解读
韩国的模式是一部专门的框架性法规,其目标在于产业发展和信托,并辅以详细的附属规则和公共部门管理法。该法规具有约束力和实质性影响,但其类别和职责应根据韩国法律文本进行解释,而非照搬欧盟人工智能法。
官方资源
| Resource | 来源 | Type | 日期 | Legal force | 为何重要 |
|---|---|---|---|---|---|
韩国国家人工智能框架法规的官方现行记录。 | National Law Information Center | Law | 2025-01-21 / effective 2026-01-22 | Binding | 国家人工智能职责、机构和信托治理的主要来源。 |
Binding Presidential Decree, current version in force from 2026-08-20. Article 25 permits a business to request high-impact AI classification confirmation from the Minister of Science and ICT, submitting the product/service overview, development/training-data overview, materials demonstrating use process/results and other relevant documents. This confirmation procedure is distinct from Article 23 transparency, Article 27 high-impact business duties and Article 28 impact-assessment provisions. | National Law Information Center | Law | 2026 | Binding | Provides the binding high-impact classification-confirmation procedure; read alongside the separate transparency, business-duty and assessment provisions. |
针对特定人工智能用途的官方透明度规定。 | National Law Information Center | Law | 2026 | Binding | 支持对通知和披露要求进行精确分析。 |
高影响力人工智能职责和影响评估规定的官方记录。 | National Law Information Center | Law | 2026 | Binding | 保留了第 34 条强制性控制措施与第 35 条评估措辞之间的区别。 |
官方公共部门人工智能和数据管理法规记录。 | National Law Information Center | Public-sector rule | 2026-08-28 | Binding | 增加了公共部门治理和影响评估层。 |
Specifies notice and labelling methods for relevant high-impact and generative-AI products/services, with stated exceptions. | Republic of Korea / Ministry of Science and ICT | Law | 2026-08-18 | Binding | Read the binding subordinate provision alongside its parent Act; maintain the distinction between separate procedures and duties. |
A business may request ministerial confirmation with product/service, training-data, use-process/results and other relevant materials. | Republic of Korea / Ministry of Science and ICT | Law | 2026-08-18 | Binding | Read the binding subordinate provision alongside its parent Act; maintain the distinction between separate procedures and duties. |
Separately specifies public disclosure of relevant risk management, explanation, user protection and oversight information, with trade-secret exceptions, and five-year evidence retention. | Republic of Korea / Ministry of Science and ICT | Law | 2026-08-18 | Binding | Read the binding subordinate provision alongside its parent Act; maintain the distinction between separate procedures and duties. |
Specifies assessment content, including affected people/groups, fundamental rights, social/economic effects, use, indicators, mitigation and improvement plans. Assessment may be conducted directly or through a third party; do not convert this into a blanket mandatory pre-deployment duty for every AI system. | Republic of Korea / Ministry of Science and ICT | Law | 2026-08-18 | Binding | Read the binding subordinate provision alongside its parent Act; maintain the distinction between separate procedures and duties. |
MOIS announced final public-sector AI ethics standards on 1 October 2026: six core values, 68 detailed check items and a three-stage tailored self-assessment approach. The framework is practical, non-binding public-sector guidance and voluntary self-checking, not new general legislation or the binding AI Basic Act and Enforcement Decree. | Ministry of the Interior and Safety | Framework | 2026-10-01 | Guidance | Adds public-sector implementation guidance without changing the force of the separate statutes. |
Japan AISI reported the first Japan–Korea joint technical workshop on 16 October 2026; the event itself took place on 8–9 October with 30 participants. Topics included multilingual/multicultural evaluation, sectoral safety methods, JAI-Trust benchmarks and Korean multimodal red teaming. This is technical cooperation and testing capacity, not a new law, binding standard or memorandum. | Japan AI Safety Institute | 报告 | 2026-10-16 | Not applicable | One shared event with separate event and publication dates; no new legal obligation. |
MSIT reported Cabinet approval of government amendment proposals on national AI-data management and use. This backfilled legislative-process milestone does not establish National Assembly enactment or commencement. Detailed provisions and later legislative status remain unverified. | Ministry of Science and ICT | 报告 | 2026-10-13 | Proposed | Maintains the distinction between this development and existing binding AI duties. |