Fabricated Rulings Reach Court Filings
South Korea's Supreme Court has released its first survey of judges on false legal citations associated with artificial intelligence, finding that 196 judges handling civil, family and administrative cases, or 80.7% of those answering that question, had encountered references to nonexistent precedents. Some court filings included fabricated judgments made to look like actual Supreme Court rulings.
Contents
The survey, conducted from October 6 to 16 and released on November 2, drew responses from 360 judges across courts nationwide. It also found that lawyers, as well as people presenting their own cases, had submitted documents containing false legal references.
The National Court Administration, the Supreme Court's administrative body, plans to consult the Ministry of Justice and other agencies about changes to the Civil Procedure Act and Criminal Procedure Act. Those discussions would address legal grounds for fines and other responses to inaccurate submissions.
Judges expressed strong support for action: 91.7% said institutional measures were necessary, including 64.8% who described them as very necessary. The administration said checking false citations takes time from judges and opposing parties, costs money and can delay people's right to a timely trial.
What the Survey Numbers Show
The most common problem was a reference to a judgment that did not exist. Judges also encountered references to actual decisions whose contents had been changed or whose meaning had been misrepresented. A legal precedent is an earlier court decision cited to support an argument about how the law should apply.
The survey allowed respondents to select more than one type of false citation. Its findings for civil, family and administrative judges were:
- 196 judges, or 80.7%, had encountered citations of nonexistent precedents.
- 159 judges, or 65.4%, had encountered false accounts of precedents.
- 41 judges, or 16.9%, had encountered false or nonexistent laws and regulations.
- 10 judges, or 4.1%, had encountered false or nonexistent academic papers and other publications.
The percentages use the response group for each survey question, with 360 judges participating in the survey as a whole. Judges could report several kinds of errors, so the categories overlap.
Nonexistent precedents were reported by 37 more judges than false accounts of precedents, 196 compared with 159. Together, those findings show that the checking task has two parts: finding the cited judgment and reading it to confirm that it supports the argument being made.
How False Citations Appeared
Judges described submissions containing AI generated documents that copied the format of Supreme Court judgments. Other filings paired an unfamiliar legal argument with a Supreme Court case number that could not be found in a search.
A case number normally helps a reader locate a specific proceeding or judgment. When that reference leads nowhere, the judge or opposing party must spend additional time trying to identify the decision behind the claim.
Some documents referred to real rulings but changed what the court had decided or used the judgment in a way that departed from its meaning. Checking the case number alone would leave that problem unresolved. The reader would also need to compare the argument in the filing with the actual judgment.
Other submissions contained text from conversations with an AI tool or wording about litigation strategy. Such passages gave judges a visible indication that AI had been used in preparing the documents.
The survey examined concerns about AI hallucinations, a term for incorrect information presented as though it were true. In a court document, an invented case can appear convincing because it comes with familiar legal wording, a case number and the format of a judgment.
Lawyers and Litigants Both Submitted Errors
Asked who had submitted false precedents or other inaccurate material, 164 judges, or 69.5% of respondents to that question, identified the parties themselves. Another 136 judges, or 57.6%, identified lawyers acting as representatives or defense counsel.
The 164 figure refers to judges who encountered false citations submitted by litigants. Across the survey, judges were describing inaccurate material they had received during proceedings.
The response groups overlap because a judge could encounter false citations from both a person conducting their own case and a lawyer. The figures measure judges reporting those experiences; the survey's focus was their exposure to the problem rather than a count of every faulty filing.
Lawyers' involvement gives the proposed verification requirement particular weight. Professional representation already places someone with legal training between a client's argument and the court. Yet more than half of respondents to the question had encountered inaccurate material submitted by lawyers.
People conducting their own cases face a different practical challenge. They must prepare arguments and respond to the other side without a lawyer's assistance. The administration warned that these litigants may find it especially difficult to identify and challenge an invented judgment or a misleading account of a real one.
Who Pays for Checking an Invented Case?
The National Court Administration said false citations can shift extra work onto the opposing party. Someone receiving a document with an unfamiliar precedent may have to search for it, check the quoted content and prepare a response explaining the error.
A false citation can therefore create costs beyond the document in which it appears. A represented party may need more legal work; a person without a lawyer must spend their own time investigating a reference that may never lead to a real judgment.
Judges face a similar burden. Time spent checking whether a decision exists or whether its contents have been changed is time added to the court's examination of the dispute. The administration connected that extra work to slower proceedings and delays in access to justice.
The concern is especially concrete when a false reference looks like an authoritative Supreme Court decision. Its apparent authority gives the opposing party a reason to investigate it carefully, while the court must resolve the citation problem before assessing the legal argument built on it.
Four Measures Draw Majority Support
Judges supported several possible responses, with the leading options attracting closely matched levels of backing. Establishing legal grounds for fines received support from 215 judges, or 61.4%. A duty to verify cited precedents and other material drew support from 214, or 61.1%.
Requiring people to disclose their use of AI received support from 212 judges, or 60.6%. Sanctions involving lawyers, including referrals for disciplinary action, were supported by 194, or 55.4%.
Only three judges separated support for a legal basis for fines from support for disclosure of AI use, 215 compared with 212. The close figures suggest respondents favored several ways of addressing the problem rather than concentrating on a single response.
Each proposal addresses a different stage of preparing and submitting a court document. Verification would require checking the legal material. Disclosure would tell the court that AI had been used. Fines would create a financial consequence under the proposed legal framework, while disciplinary referrals would bring a lawyer's conduct before the relevant professional process.
Repeated false citations were among the conduct identified for possible disciplinary referrals. The administration's proposals combine responsibility for the accuracy of a filing with consequences when inaccurate material reaches court.
Legal Changes and Online Guidance Come Next
The National Court Administration plans to discuss amendments to the Civil Procedure Act and Criminal Procedure Act with relevant agencies, including the Ministry of Justice, and submit its views to the National Assembly. The proposed amendments would provide legal grounds for measures such as fines.
The sequence described by the administration is:
- October 6 to 16: Judges across courts nationwide took part in the survey.
- November 2: The survey findings were released.
- Next steps: Consultations with relevant agencies and submission of the administration's views to the National Assembly on legislative changes.
No timetable was given for those consultations, a legislative vote or the introduction of fines. The amounts of any proposed fines and the detailed conditions for imposing them also remain undecided in the announced plans.
The administration is considering publishing guidance about AI use on the electronic litigation website. That would place advice where lawyers and litigants already carry out court business, with the aim of helping them check the accuracy of material prepared using AI.
The survey findings point to a practical priority for that guidance: checking both existence and meaning. A real case number helps locate a decision, while reading the decision establishes what the court actually said. The examples judges encountered show why both checks matter.
Key Points
- The Supreme Court's first survey on AI related false citations drew responses from 360 judges.
- Among civil, family and administrative judges answering the question, 80.7% had encountered citations of nonexistent precedents.
- Judges described fabricated Supreme Court judgments, unsearchable case numbers, altered rulings and AI conversation text left in filings.
- False material came from both litigants themselves and lawyers representing parties.
- 91.7% of respondents supported institutional action, with majorities favoring fines, verification duties, disclosure and lawyer sanctions.
- The National Court Administration plans consultations on procedural law changes and is considering guidance on the electronic litigation website.
