AI Regulation Tracker  /  Regulator guidance

South Korea's MFDS Publishes Approval Guideline for Generative-AI Medical Devices

The Ministry of Food and Drug Safety issued a review and approval guideline for medical software that uses generative AI, telling manufacturers exactly what to identify, submit, and prove. MFDS calls it the world's first such guideline.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published July 7, 2026  ·  Last updated July 7, 2026

South Korea's Ministry of Food and Drug Safety published a formal approval and review pathway for medical devices that use generative artificial intelligence, a document the agency describes as the first of its kind from any regulator. The Guideline on Approval and Review of Generative-AI Medical Devices, filed as document sequence 15628 and registered as guidance number 1416-01, was published on January 24, 2025. It tells manufacturers what to identify, how to draft their applications, and what they must prove before a generative-AI medical device can reach the Korean market.

What the guideline covers

The guideline addresses AI software-as-a-medical-device that relies on generative AI. MFDS gives concrete scope examples: software for medical-image reading, diagnostic assistance, and treatment-planning support. The first task it sets for a maker is a threshold one. A company must determine whether its product is in fact a generative-AI medical device, because that classification governs which application-drafting method and dossier apply. From there the guideline lays out the method for preparing the approval application and the evidence package the ministry expects.

The core obligation is demonstrating safety and effectiveness. The guideline asks makers to account for data bias, accuracy, and ethical risk, and to do so across the full product lifecycle rather than at a single approval checkpoint. That lifecycle framing matters for generative systems, whose behavior can shift as models and inputs change after clearance.

How it was built, and what it does not do

MFDS developed the guideline through an expert consortium it convened in March 2024 to analyze full-lifecycle risk for these products. The result is guidance, not a statute. It sits under the Medical Devices Act framework and functions as civil-petitioner guidance, the practical instruction set a company follows when it approaches the ministry. It does not create a new penalty regime and it does not replace the underlying device-approval law. What it does is convert a general expectation, that AI medical software be safe and effective, into a specific submission method for the generative-AI case.

A note of precision on the "world first" claim. That framing comes from MFDS and the Korean trade press, and it is worth attributing to the agency rather than treating as an independent fact. The substance that a professional can rely on is narrower and firmer: Korea has a published, dedicated generative-AI device guideline with a defined dossier, and the effective date on the document is January 24, 2025.

The cross-border angle for US firms

For a US digital-health or SaMD maker, the practical significance is timing. The US Food and Drug Administration has AI-related device guidance and mechanisms such as predetermined change control plans, but it has not published a dedicated approval guideline for generative-AI medical devices. A US company eyeing Korea therefore encounters a concrete submission template abroad earlier than it has one at home. That does not bind US operations, and it does not change FDA requirements. It does mean a firm planning a Korean launch can and should read the MFDS guideline directly, because it defines the classification question and the evidence expectations that a Korean application will turn on.

What to do now

The immediate action is diagnostic. A regulatory-affairs lead should retrieve the guideline from the MFDS site, compare the product against the scope examples, and decide whether it is treated as a generative-AI medical device under the Korean definition. If it is, the drafting method and dossier in the guideline become the working blueprint, and the safety and effectiveness case should be organized around data bias, accuracy, and ethical risk across the lifecycle. Because this is guidance rather than a fixed rule, teams should also confirm they are working from the current version on the ministry's page before they build a submission around it.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.