South Korea AI Basic Act Guide: Obligations and Timelines

South Korea's AI Basic Act has applied since 22 January 2026, and it reaches foreign AI providers whose products touch Korean users, not just Korean companies. The Framework Act on the Development of Artificial Intelligence and the Creation of a Foundation for Trust (Act No. 20676) is the second comprehensive AI statute after the EU AI Act, but it is built differently: fewer prohibitions, a short list of hard obligations, and modest fines backed by inspection and corrective order powers. This guide explains who is covered, what high-impact and generative AI operators must do, the dates and the grace period, how enforcement works, and what the Act means for teams running AI agents and the machine identities behind them. It is written for security, privacy and AI governance practitioners at organisations that build or deploy AI used in Korea, and for executives who need to know where the real exposure sits.

Key takeaways

• The Act is in force but enforcement is soft for now. The Act and its Enforcement Decree took effect on 22 January 2026, and the Ministry of Science and ICT (MSIT) has announced a guidance period of at least one year in which fact-finding investigations, corrective orders and fines are generally deferred, except in serious cases such as loss of life or human rights violations.
• Three duties carry direct fines. Under Article 43, fines of up to KRW 30 million apply to failing to give advance notice of AI use (Article 31(1)), failing to designate a domestic agent (Article 36(1)) and failing to comply with a cease or correct order (Article 40(3)). Labelling and high-impact duties are enforced through investigations and corrective orders.
• Generative AI output must be labelled. Article 31 requires advance notice, labelling of generative AI output, and clear notice or labelling of realistic synthetic sound, images and video, with lighter treatment for artistic works.
• High-impact AI is defined by sector and effect. Article 2(4) lists ten areas, including energy, drinking water, healthcare, nuclear, criminal biometrics, hiring and loan screening, transport, public sector decisions and student evaluation, where the system significantly affects life, safety or fundamental rights.
• Foreign providers need a Korean agent above set thresholds. Operators with no Korean address or office must appoint a domestic agent if they had prior-year revenue of KRW 1 trillion or more, AI service revenue of KRW 10 billion or more, or more than one million average daily Korean users.

Scope: who and what is covered

The Act regulates "artificial intelligence business operators", split into operators that develop and provide AI and operators that provide products or services using AI. These categories do not map neatly onto the EU AI Act's provider and deployer roles, so a company that fine-tunes a third-party model and ships it inside its own product may hold both roles at once.

• Territorial reach: Article 4(1) applies the Act to conduct outside Korea that affects the Korean market or Korean users.
• Exclusion: Article 4(2) excludes AI prescribed by Presidential Decree that is developed and used solely for national defence or national security.
• Generative AI: AI that produces text, sound, images, video and similar outputs by imitating the structure and features of its input data. Transparency duties apply to these systems regardless of sector.
• High-impact AI: AI likely to significantly affect, or pose a risk to, human life, physical safety and fundamental rights, used in one of the Article 2(4) areas below.
• Large-scale (high-performance) AI: Article 32 applies extra safety duties to systems whose cumulative training compute meets a threshold set by Decree. MSIT set it at 10²⁶ floating point operations or more, combined with cutting-edge technology and potential for widespread, significant impact on fundamental rights.

The high-impact areas in Article 2(4)

Being in a listed area is not enough on its own: the system must also be likely to significantly affect life, safety or fundamental rights. Article 33 lets an operator review this in advance and ask MSIT to confirm whether a system is high-impact. MSIT published high-impact determination guidelines on 22 January 2026 and a revised version on 29 January 2026, and those guidelines, not intuition, should drive classification.

Key obligations

Transparency and labelling

• Advance notice: Users must be told before use that the product or service operates on high-impact or generative AI. The Decree allows this through documentation, on-screen display or posting. A visible interface notice is safer than a clause in the terms of service.
• Output labelling: Generative AI outputs must indicate they were generated by AI. MSIT's transparency guidance expects labels that are both human-readable and machine-readable, which in practice means a visible label plus metadata or an invisible mark. See the AI Model Provenance and Watermarking Guide.
• Deepfakes: Realistic synthetic sound, images or video that are hard to distinguish from real content must be notified or labelled so users can clearly recognise them. According to Shin & Kim's reading of the Decree, a watermark alone is not enough for these outputs. Artistic and creative works may be labelled in a way that does not hinder their display or enjoyment.
• Exemptions: The Decree relieves operators where AI use is obvious to the user, or where the system is used solely for the operator's internal business purposes. An internal coding assistant is a likely candidate for the internal-use exemption; a customer-facing chatbot is not.

High-impact AI duties

• Risk management plan: A documented process for identifying and treating risks across the system's life cycle.
• Explanation plan: Within technical limits, the ability to explain the outcome, the main criteria used to reach it and an outline of the training data.
• User protection plan: Measures to protect users affected by the system.
• Human management and oversight: Named people who supervise the system and can intervene.
• Documentation: Documents that verify the safety and reliability measures. According to Shin & Kim, the Decree requires them to be kept for five years and key elements to be published on the operator's website or premises, excluding trade secrets.
• Equivalence: Article 34(3) treats measures taken under other laws that the Decree recognises as equivalent as compliance, which matters for regulated sectors such as medical devices and finance.

Timeline and current status

MSIT has issued supporting guidelines covering high-impact determination, operator responsibilities for high-impact AI, impact assessment, generative AI transparency and safety for advanced AI, and runs an AI Basic Act Help Desk through the Korea Software Industry Association (KOSA) for case-specific questions. The grace period is an enforcement policy, not a statutory delay: the obligations themselves already apply, and the guidelines can change faster than the Act. Check MSIT's current position before relying on any date beyond January 2027.

Enforcement and penalties

• Regulator: The Minister of Science and ICT. Policy direction sits with the National AI Committee chaired by the President, and an AI Safety Research Institute evaluates risks and develops standards.
• Fact-finding (Article 40(1) and (2)): Where a violation of Article 31(2) or (3), 32(1) or (2) or 34(1) is found or suspected, or a report or complaint is received, MSIT may require data submission and authorise officials to enter premises and examine records.
• Corrective orders (Article 40(3)): MSIT may order an operator to cease or correct a violation. Ignoring the order is itself finable.
• Fines (Article 43): Up to KRW 30 million per violation for the three listed duties. Small by EU standards, but a corrective order can force product changes.
• Agent liability: Under Article 36(3), a domestic agent's violations are treated as the foreign operator's own, so the agent is not a liability shield.

See the EU AI Act Compliance Guide for the EU side of this comparison.

What it means for AI agents and non-human identities

The Act does not mention AI agents, service accounts or API keys. It still lands on them, because an agent that answers customers, screens applicants or acts on accounts is a product or service "operated based on" AI, and the evidence MSIT can demand under Article 40 is mostly identity and logging evidence.

• Notice at every agent touchpoint: An agent that emails, messages or calls Korean users must carry the Article 31(1) notice in each channel, not only in the web app.
• Human oversight needs an owner and a kill switch: Article 34's human management and oversight duty is hard to show without a named owner for each high-impact agent, an approval step before consequential actions, and a way to suspend the agent's credentials quickly. See the Identity Governance for AI Agents Guide.
• Explanation depends on attribution: To explain why a loan or hiring decision was made, you need to know which agent, model version, tool calls and data sources produced it. Shared API keys across agents make that impossible. Give each agent its own workload identity (NIST SP 800-53 IA-4 identifier management, IA-9 service identification and authentication).
• Least privilege limits what a corrective order can catch: An agent scoped to read-only access cannot quietly take high-impact actions outside its documented purpose. Map tool permissions to the documented intended use (AC-6). OWASP's LLM06:2025 Excessive Agency describes the failure this prevents. See the AI Agent Authorisation Guide.
• Logs are your Article 40 evidence: Records of agent actions, approvals and outputs, protected from modification by the agent itself (AU-2, AU-9) and retained long enough to meet the five-year documentation expectation, are what an inspection will ask for.
• Labelling in pipelines: Agents that generate images, audio or video for publication should attach labels and metadata in the generation pipeline.

Compliance steps

1. Inventory AI touching Korea. List every model, AI feature and agent available to Korean users or affecting the Korean market, including third-party models embedded in your products. Record the operator role you hold for each.
2. Test the domestic agent thresholds. Check prior-year total revenue (KRW 1 trillion), AI service revenue (KRW 10 billion) and average daily Korean users over the last three months of the prior year (one million). If any is met and you have no Korean office, appoint an agent with a Korean address, designate in writing and report to MSIT.
3. Classify against Article 2(4). Use MSIT's high-impact guidelines to decide which systems are high-impact. Record the reasoning, and request an Article 33 confirmation for borderline cases.
4. Ship advance notices. Add Korean-language notices in every user channel for generative and high-impact AI. This duty carries a direct fine, so do it first.
5. Implement labelling. Add visible and machine-readable labels to generative outputs and clear labels to realistic synthetic media. Test that labels survive your export, compression and sharing paths.
6. Build the high-impact file. For each high-impact system: risk management plan, explanation plan, user protection plan, named overseers, safety documents and a fundamental-rights impact assessment. Plan for five-year retention and publication of key elements.
7. Check the compute threshold. Model developers should calculate cumulative training compute using MSIT's safety guidance and, if at or above 10²⁶ operations, prepare the Article 32 risk results and risk management system for submission.
8. Rehearse an inspection. Run a mock Article 40 request: can you produce notices, labels, risk plans and agent logs within days? Fix gaps before the guidance period ends.

For small organisations without a Korean entity and below the agent thresholds, steps 4 and 5 are usually the whole job unless a product falls into a high-impact area. Providers selling to Korean public bodies should expect requests for the impact assessment, because Article 35(2) tells state agencies to prefer assessed products.

Detection and measurement

• Notice coverage: Percentage of Korean-facing AI features and agent channels showing the Article 31(1) notice. Target 100 percent.
• Label survival rate: Share of sampled generated media that still carries visible and machine-readable labels after export and sharing.
• Classification freshness: Share of AI systems with a high-impact classification reviewed in the last 12 months or since the last material change.
• Agent attribution: Share of production agents with a unique identity, a named owner and tamper-resistant action logs.
• Evidence turnaround: Time to assemble a complete evidence pack in a mock inspection.

Common failure patterns

• Assuming no Korean entity means no exposure. Article 4(1) reaches conduct abroad that affects Korean users, and the agent duty is aimed squarely at foreign operators.
• Treating the grace period as a delay. The duties apply now; only the enforcement posture is softened, and serious cases are excluded from it.
• Labels stripped downstream. Metadata-only labels vanish when images are re-encoded or screenshotted, leaving realistic content unlabelled.
• Notice only in the app. Agents that send emails or messages on the organisation's behalf go out without the AI notice.
• One shared key for every agent. When a decision is challenged, nobody can tell which agent or model version produced it, so the explanation plan cannot be honoured.

Framework mapping

Practitioner checklist

• Inventory every AI feature, model and agent used by or affecting Korean users, with the operator role held.
• Run the three domestic agent threshold tests each year and appoint and report an agent if any is met.
• Classify systems against Article 2(4) using MSIT's current guidelines and record the decision.
• Show a Korean-language AI notice in every channel, including agent-sent emails and messages.
• Label generative outputs visibly and in metadata, and test that labels survive export.
• Complete the high-impact file: risk plan, explanation plan, user protection plan, named overseers, impact assessment.
• Give each agent a unique identity, a named owner, least-privilege tool scopes and tamper-resistant logs.
• Retain safety documentation for five years and publish key elements without trade secrets.
• Track MSIT guideline revisions, the end of the guidance period and Bill No. 2220395.

Standards and references

• Korea Legislation Research Institute: Framework Act on the Development of Artificial Intelligence and the Creation of a Foundation for Trust (English translation)
• US International Trade Administration: South Korea AI Basic Act
• Digital Policy Alert: Ministry of Science and ICT released revised high-impact AI decision guidelines
• Digital Policy Alert: Bill amending AI Basic Act, Bill No. 2220395
• Shin & Kim: AI Basic Act enforcement and Enforcement Decree overview
• Cooley: South Korea's AI Basic Act, Overview and Key Takeaways
• OWASP: Top 10 for LLM Applications 2025
• NIST: SP 800-53 Rev. 5

This guide summarises South Korea's AI Basic Act, its Enforcement Decree and MSIT guidance as at October 2026 and is not legal advice. Related NHI Mgmt Group resources: Global AI Regulation Map · Japan AI Promotion Act Guide · China Generative AI Measures and AI Labelling Rules Guide · Agentic AI Compliance Guide