On November 2,2026, South Korea’s National Court Administration (법원행정처), the management body attached to the Supreme Court, published the results of an investigation into erroneous references produced by AI in court filings. Conducted from October 6 to 16 among judges from all jurisdictions, it collected 360 responses, and about 80% of the responding judges handling civil, family, and administrative cases say they have already encountered citations of false or non-existent rulings. The institution deems it necessary to amend the Civil Procedure Code and the Criminal Procedure Code to create a legal basis, notably for a fine.
From Unfindable Rulings to Fake Judgments
Among the judges who answered the question about the types of errors, 196 (80.7%) have seen a non-existent ruling cited, 159 (65.4%) a falsified ruling, 41 (16.9%) a false or non-existent legal text, and 10 (4.1%) a false or non-existent article or scholarly work. A single judge could check multiple boxes. The document does not specify the number of judges contacted or the response rate.
The examples cited by the administration go beyond a simple mis-copied reference. Parties have filed fake AI-generated judgments, mimicking the formatting of Supreme Court rulings. Others have put forward reasoning unknown to the judges, accompanied by a Supreme Court case number that no search could find. In some cases, fragments of the conversation with the AI or notes on litigation strategy were included verbatim.
This last case is documented in France: in an order dated December 9,2025, the Grenoble Administrative Court ruled on an application where the exchange with the chatbot had been left in the filed text.
Litigants and Lawyers Implicated
Regarding the origin of the false references, 164 respondents (69.5%) pointed to parties acting on their own, and 136 (57.6%) to lawyers, whether they represent a party or provide criminal defense. The press release emphasizes this second figure: invented references also appear in significant numbers in lawyers’ filings.
Two foreign cases are cited as references. On June 22,2023, in New York, federal judge P. Kevin Castel imposed a $5,000 penalty, jointly and severally, on two lawyers and their law firm in the Mata v. Avianca case. On June 6,2025, the Divisional Court of the High Court of England and Wales issued the Ayinde judgment, which outlines the range of possible responses, from public reprimand to referral to the police. In the Al-Haroun case, attached to the previous one, 18 of the 45 references submitted did not exist.
A Legal Basis to Create, Measures Under Study
At this stage, the press release sets neither an amount, a timetable, nor the competent authority to impose the sanction. The term used, 과태료, refers in Korean law to a non-criminal monetary penalty. Regarding measures, 91.7% of respondents consider an institutional response necessary. Among the proposed avenues, 61.4% favor a legal basis for the fine, 61.1% an obligation to verify cited rulings, 60.6% an obligation to declare the use of AI, and 55.4% sanctions, such as disciplinary referral of lawyers in case of repeated false citations.
The AI Hallucination Cases database, maintained by researcher Damien Charlotin, lists as of October 31,2026,2,246 court decisions worldwide, including 1,517 in the United States, 14 in France, and 8 in South Korea. It itself indicates that it does not cover all false citations filed in courts. The 14 French decisions in this database are all dated December 2025 or later.
