The Ministry of Science and ICT began full operations this month of a public-private working group tasked with writing a National AI Safety Ecosystem Master Plan due by December. The panel meets as the AI Basic Act’s enforcement decree remains in its first-year grace period, during which administrative fines are largely deferred unless regulators document serious social harm.
Mandate of the working group
MSIT said the consultative body will translate global safety debates into Korean implementation paths: incident reporting, evaluator capacity, and standards for high-impact systems used in finance, healthcare, and public services. Director general Lee Jin-soo told reporters the goal is to align field acceptance with overseas regimes without copying them wholesale.
The master plan is separate from the AI data centre special act, where another MSIT team is drafting lower-level rules on grid assessments, permit timeouts, and definitions of AI-ready facilities. Data-centre developers want clarity before fourth-quarter land bids in provincial industrial zones.
What the Basic Act already requires
The act took effect Jan. 22 with transparency duties for generative models, impact assessments for high-risk deployments, and authority for MSIT to operate an AI Safety Institute. Guidelines on business-operator obligations and advanced AI safety were published through the National Information Society Agency and industry associations.
Foreign providers above revenue or daily-user thresholds must appoint domestic representatives to interface with MSIT; compliance lawyers say registration workflows can take weeks because notarized corporate documents must be translated. A help desk run through the Korea Software Industry Association fields technical questions during the grace year.
Industry stakes
Cloud hyperscalers and domestic conglomerates are shipping enterprise copilots into call centers and factories. Safety working-group members include chipmakers worried about export controls, banks piloting agentic workflows, and startups selling retrieval-augmented customer service. Each wants predictable audit criteria before procurement seasons close.
Telecom operators argued in early meetings that safety testing should not duplicate network security reviews already mandated for critical infrastructure. MSIT staffers responded that model-behavior failures are a different failure mode than packet floods, according to two participants who spoke on background.
December deadline pressure
The ministry promised to circulate a draft for interagency comment in November, then finalize in December ahead of budget hearings. Opposition lawmakers may ask whether the safety plan is funded or aspirational; MSIT has not attached won figures to every workstream yet.
Parallel talks in the National Assembly audit season cover public-sector chatbots and training-data provenance, themes that overlap with the master plan but sit under different committees.
Comparison with overseas moves
Seoul officials cite the EU AI Act and U.S. executive orders as reference points but emphasize Korea’s export-driven manufacturing base. Factory-floor vision models face different tail risks than consumer chat apps, a gap the working group’s subcommittees are supposed to map.
Academics on the panel pushed for shared red-team datasets hosted domestically so smaller firms can test jailbreak resistance without sending proprietary prompts overseas.
What businesses should do now
Legal teams advise documenting model cards, logging human overrides, and cataloguing third-party embeddings even while fines are rare. The grace period is not immunity: MSIT retains power to investigate egregious cases, and sector regulators in finance and health can apply their own licensing rules.
By year-end, Korea intends to show a written safety architecture, not just a startup count. Thursday’s meetings are procedural, but they set the vocabulary regulators will use when the grace period ends.