South Korea's Courts Confront AI-Invented Case Citations

ActuIA reports on 2 November 2026 that a survey by South Korea's judicial administration found roughly 80% of responding civil, family and administrative judges had received written submissions citing false or non-existent decisions attributed to AI. The institution is asking for an explicit legal basis, including a fine, in the codes of procedure. No amount and no timetable have been specified.

What the report actually establishes

The figure covers judges who answered the survey, not the whole bench, and the report does not break it down by court, by case type or by period. It also does not say whether the false citations originated with lawyers, parties representing themselves, or both, nor how many filings were affected or how judges handled them. The proposed fine is a request for a legal basis, not an adopted measure: the mechanism, the amount and the parliamentary or regulatory path remain open.

That gap matters. A fine is only as useful as the detection it depends on. Citation checks are cheap to automate, but someone must run them and someone must be accountable when a check fails.

Why fabricated citations land in court

General-purpose language models generate plausible text rather than verified references: asked for authority, they can produce a citation with the right shape and the wrong substance. This is documented across jurisdictions, and courts in several countries have had to address lawyers who filed briefs with invented references. Korean procedural codes, like most, say nothing explicit about machine-assisted drafting, which leaves judges to handle it case by case.

For teams building products where citations carry weight, the lesson is not new but it is now adversarial. Retrieval over a verified corpus, deterministic citation lookup and a refusal path when a source cannot be confirmed are engineering decisions, not nice-to-haves. A system that cannot say "no source found" will eventually say something worse.

What a team can check this week

• Take ten outputs from any assistant you use for research or drafting and verify every citation manually; log the failure rate and the type of error.
• Confirm whether your retrieval layer is restricted to sources you control or license, and whether the model is allowed to answer without one.
• Decide who signs off on a document before it leaves the building, and write that into your internal process rather than assuming it.

A legal framework will take time to settle. The operational standard does not wait: if a citation cannot be traced to a real, checkable source, it should not reach a court, a client or a regulator. That is a process question a small team can answer before any legislator does.