When Korea’s Framework Act on the Development of Artificial Intelligence and the Establishment of a Foundation for Trust—widely known as the AI Basic Act—entered into force on January 22, 2026, it did more than formalize national AI policy. It placed transparency at the heart of regulation, requiring that users be able to recognize when AI has shaped the products, services, or content they receive.

Article 31 of the statute makes that obligation explicit: providers of high-impact or generative AI–based products and services must give prior notice that AI is operating the service and must clearly label outputs generated by AI.

Article 31 operationalizes this through two requirements that matter most for content industries: prior disclosure when a product or service is AI-driven, and labeling when users are presented with generative AI outputs. The regulatory design is not aimed at policing internal experimentation. It is aimed at governing user-facing deployment, where confusion, reliance, and harm scale.

For media organizations, the practical test is therefore narrow but consequential: Is AI being delivered to the public as a feature or service? If yes, the organization is no longer in the realm of internal productivity tooling. It has crossed into a regulated interface with end users.

Industry and legal commentary around Article 31 commonly frames transparency duties around three buckets: disclosure that AI is operating the service, labeling of AI-generated outputs, and heightened notice for synthetic or deceptive content such as deepfakes. Applied to newsroom products, this can be translated into four compliance-relevant service types. The statute does not list “four types” as such; the categories below are a functional mapping of Article 31’s obligations onto common media features.

1) Conversational News Products (Chatbots, Q&A, “Ask the Newsroom” AI)

If a user prompts a system and receives AI-generated answers, and AI automatically structures the service. This is the clearest case for advance disclosure: Readers should not have to infer whether they are interacting with automated generation or human editorial judgment.

The distinction with internal drafting tools is critical. Internal summarization may never leave the newsroom. A chatbot is the newsroom—productized.

2) Automated Summaries and AI Briefings

Auto-summaries appended to articles and AI briefings delivered through apps or newsletters fall into the labeling problem space. Article 31’s core concern here is not whether journalists used AI somewhere in the pipeline; it is whether the output consumed by readers is AI-generated and therefore should be identifiable as such.

This is where many publishers will face compliance friction: summary modules are often treated as “UX enhancements,” but regulatorily they resemble generative outputs distributed at scale.

3) Personalized Generation and Rewriting (User-Requested Versions)

Services that rewrite, restructure, or generate alternative article versions based on user requests sit at the center of the regulatory risk curve. Once a product moves beyond curation and recommendation into text generation or transformation, the reader is no longer simply browsing. The reader is receiving a manufactured version of content.

That shift raises the transparency expectation: readers may need to be informed that what they are reading is not merely edited by an editor, but generated or materially altered by AI. In compliance terms, the question is not editorial intent. It is AI determinism over the delivered output.

4) Synthetic Media (Images, Video, Audio; Deepfakes)

Synthetic media is where transparency becomes inseparable from integrity. If a person’s face, voice, or likeness is generated or reproduced, the risk is not only legal exposure but institutional credibility. Journalism depends on verifiability; synthetic artifacts increase the probability of misperception by design.

The policy rationale of Article 31—reducing user confusion—points toward stricter scrutiny in this domain. For media, deepfake handling is no longer a niche ethics issue. It is a frontline compliance and trust issue.

What Likely Stays Out—But Not “Risk-Free”

Translation, grammar correction, research organization, internal drafting, and design assistance generally sit on the internal-tool side of the line. Still, these uses are not automatically insulated. If the final output creates a plausible risk of user confusion—particularly where synthetic elements are embedded—publishers may still need additional standards to avoid misrepresentation.

Why This Becomes a Real-World Test Quickly

As AI adoption matures, publishers move from internal efficiency to product strategy: turning features into user-facing services. That is the inflection point where compliance stops being theoretical. The AI Basic Act is designed to regulate the public-facing layer, where users rely on what they receive and where disclosure failures can scale.

The Act’s underlying question is not “Did the newsroom use AI?” It is “Did the user receive AI-mediated outputs without being able to tell?”

Bottom line: for media organizations, the first-order compliance test is whether AI-based functions are offered to users and whether users consume AI-generated outputs. Internal use remains largely a governance and process issue. Reader-facing AI services—chatbots, automated summaries, generative rewriting, and synthetic media—are where disclosure and labeling obligations become live.

In the early enforcement period, the most defensible posture is a feature-by-feature assessment: map each user-facing capability to Article 31’s disclosure and labeling logic, then implement clear, consistent notice mechanisms before product scale makes remediation harder.

Legal basis and guidance: Korea IT Times analysis based on the Republic of Korea’s AI Basic Act (Act No. 20676; effective 22 January 2026), Article 31, as amended by Act No. 21311 (20 January 2026). 
Additional guidance: KOSA, “Article 31 Transparency Guidance” (public materials).
 

Kim Yoo-shin More by this reporter