Korea's Government Built Its Own AI Legal Secretary

South Korea's Ministry of Government Legislation built an AI legal secretary in-house, by its own civil servants, on a homegrown Korean foundation model, and on a retrieval-augmented generation (RAG) corpus of roughly 350,000 documents. The trial went live on 2026-07-14, and the ministry calls the tool a "middle review reference," not a final legal judgment.

According to coverage from YTN and Maeil Business Today (MS Today), roughly thirty ministry officials built the tool in about a month using the whole-of-government AI common base, the Legislation Ministry's existing legal-interpretation system, and off-the-shelf RAG tooling. They did not buy a vendor product. They did not license a foreign model.

The headline figure is 60,000. That is the count of Supreme Court precedents the tool can cite, the slice most likely to land with a cold reader. The full corpus is bigger: about 240,000 statutes and administrative regulations, plus roughly 50,000 local ordinances from Seoul, Incheon, Gyeonggi, Daejeon, and Sejong loaded ahead of the original second-half schedule. Precedents, statutes, and ordinances together total closer to 350,000 documents, not 60,000.

Bernama's English-wire re-report of the Yonhap dispatch carries the same anatomy: a RAG assistant on a sovereign model, accessed through the government intranet "On-AI Lab" (온AI 실험실). The legal-tech trade The Legal Wire frames the move as part of the country's "public-sector AI transition" (AX) push, alongside AI court transcription and AI customs classification pilots elsewhere in government.

A civil servant types a legal question; the system retrieves the relevant Supreme Court precedents, statutes, and local ordinances, and drafts a middle-review answer with citations. The final interpretation, the signed-off document, the call on a hard case, all stay with the human official. Legislation Minister Cho Won-cheol told reporters the work is hard precisely because it requires high expertise, per the Yonhap wire carried by Bernama. The pitch is that a tool that pulls the right precedents and statutes in seconds shortens the drudge work and frees officials to spend more time on the judgment calls.

The trial is roughly one month old as of mid-August, and no accuracy benchmarks, hallucination rates, or independent adoption numbers have been published. The Korean primary reporting (YTN, MS Today) treats this as a deployment claim, not a performance claim, and the English-language aggregator that picked up the wire copy repeats the same framing. Whether the system can actually distinguish a binding precedent from a persuasive one, or catch itself when the corpus has no answer, is the open question the next reporting cycle will need to answer.

A public ministry looked at the legal-AI market, decided not to buy, and staffed a narrow precedent-Q&A tool itself. The original Chosun English report that named the trial uses the 60,000-precedent figure as the headline; the reality is a 350,000-document sovereign RAG build, run by civil servants, that calls itself middle-review material. Watch item: the first published evaluation cycle, and which other ministries ask to plug in.