Executive Summary

A court has counted for itself how often filings arrive carrying case law that AI invented. Korea's National Court Administration put the question to 360 judges in courts across the country between October 6 and 16, and published the results on November 2. Of those who answered, 243 said a filing like that had reached them at least once. That is two in three.

The numbers in this survey carry two denominators. The 80.7% that the headlines carried is not the share of all judges who met fake case law. It is the share, among the 243 who said they had met it, who saw case law that does not exist at all. Read the two as one and the sentence comes out a notch stronger than the survey allows.

Sections 1 through 3 carry what the National Court Administration published, and the first part of Section 4 carries the overseas tally and the sanction it describes. Setting the measures the judges chose beside three ordinary moves in data work, confirming the source, recording the tool and pricing the error, is this article's reading: the end of Section 4 and all of Section 5.

Key Numbers

The first three of these four come from this survey, and the last one shows how far the same thing has piled up elsewhere. Each card counts a different population, so the small print belongs with the figure.

Source: the National Court Administration's publication of its judge survey (November 2, 2026) and the reporting that carried it, and the AI Hallucination Cases database (updated October 19, 2026)

67.5%

Judges who met one at least once

243 of the 360 who answered had received a filing citing case law that was false or did not exist

80.7%

Saw case law that never existed

The denominator here is the 243 who said yes, not the full group of 360

91.7%

Want measures put in place

Counting only the answer "strongly necessary" gives 64.8%

2,202

Rulings worldwide over AI-invented citations

What Charlotin's tracking database had gathered as of October 19, 2026

1

The Court Asks Its Own Judges, for the First Time

This is the first time the court has put the question of AI hallucination to judges directly. The National Court Administration ran the survey from October 6 to 16 among 360 judges handling civil, family and administrative cases in courts at every level, and published the results on November 2. Filings citing case law or statutes that did not exist had been arriving one after another, and the stated purpose was to find out how far it had spread.

Of the 360 who answered, 243 said such a filing had reached them at least once. That is 67.5% of the whole, two in three.

The denominator changes at the question about which types they had seen. From there the base is not the 360 but the 243 who said yes. Of those 243, 196 (80.7%) saw case law that does not exist at all, 159 (65.4%) saw real case law whose content had been altered, 41 (16.9%) saw nonexistent statutes, and 10 (4.1%) saw nonexistent papers or other sources. Multiple answers were allowed, so the four add up to more than 100%.

The denominator under 80.7% is 243, not 360 The 360 judges who answered the survey Met one: 243 (67.5%) Did not: 117 Setting the 243 who met one back at 100% (multiple answers allowed) Case law that never existed 80.7% (196) Altered real case law 65.4% (159) Nonexistent statutes 16.9% (41) Nonexistent papers 4.1% (10) The four values below hold only inside the orange block. Swap the denominator to 360 and every one of them shrinks.
▲ Pebblous original diagram. The respondent counts and the per-type figures are taken from the National Court Administration survey; drawing the two denominators as separate rows is this article's construction.

Collapse the two denominators into one and the sentence becomes "eight in ten judges have met fake case law." What the survey says is that two in three met one at least once, and that eight in ten of those who met one saw the most common type. The published material does not give the per-type shares measured against all 360. So this article does not work those figures backwards either.

2

What Turned Up Inside the Filings

The survey also asked who had filed them. Among the judges who said yes, 69.5% pointed to litigants running their own cases without a lawyer, and 57.6% pointed to attorneys acting as counsel or defense counsel. Multiple answers were allowed here too, so some judges named both sides.

Particular scenes came back with the answers. Filings were common that copied the format of a Supreme Court judgment and produced a fake one with AI. One response said a filing "gave a Supreme Court docket number alongside reasoning I had never seen before, and when I actually searched for it, the case law did not exist." Real rulings came in with their content changed at will, or cited against the meaning of the judgment. In some cases, the conversation held with an AI or notes on litigation strategy had been left sitting inside the filing.

The last type is a different kind of mistake. The others put false grounds into a document. This one leaves the whole working process showing. It is closer to not knowing where the paste should stop.

The National Court Administration noted that "litigants who conduct their own cases without a lawyer may have particular difficulty finding these errors and responding to them." The side that brings in the most fake case law and the side least able to spot it when the other party brings it are the same side.

3

Administrative Fines Are Where the Court Moves First

91.7% of the judges who answered said institutional measures need to be put in place. The answer "strongly necessary" accounts for 64.8%. Asked what should be done, with multiple answers allowed, four options came back at roughly the same height.

Measure Share choosing it What the device actually asks for
A legal basis for administrative fines 61.4% Puts a money cost on whoever filed the false citation
A duty to verify cited case law 61.1% Makes someone confirm that the case law written into a filing is real
A duty to state that AI was used 60.6% Makes the filing say which tools it passed through
Referral for attorney discipline and similar sanctions 55.4% Hands the matter over to professional ethics procedures

The figures come from the National Court Administration survey. The right-hand column, saying what each measure does as a device, is this article's own wording.

The National Court Administration said it will push to amend the Civil Procedure Act and the Criminal Procedure Act so that administrative fines have a legal basis. It plans to consult the Ministry of Justice and other relevant agencies and then submit its view to the National Assembly, and it is also weighing a notice on the electronic litigation site covering what to watch for when using AI.

It gave its reason as well. "Citation of false or nonexistent case law arising from the use of AI burdens both the other party and the bench, and in the end it delays the public's right to a trial." One fake precedent lands and the other side writes a rebuttal while the bench searches docket numbers one at a time. That time comes out of the hearing of some other case.

4

A U.S. Court Sanctions Two Lawyers Without Confirming AI Was Used

Korean courts are not alone in this. The AI Hallucination Cases database run by Damien Charlotin of HEC Paris gathers decisions from courts and tribunals around the world in which citations or content invented by AI became an issue. As of the October 19, 2026 update it holds 2,202. Split by who filed the document, self-represented litigants account for 1,267, lawyers for 872, judges for 32 and experts for 17.

That tally and the National Court Administration's figures mean different things. One counts decisions; the other counts the share of judges who ran into something. The direction still matches, with the self-represented side the thicker one of the two. People who do not know how to work a legal database ask a chatbot first, and on that point the same thing is happening in courtrooms in both countries.

Where the line for sanctions falls comes out more clearly in the American case. On March 13, 2026 the U.S. Court of Appeals for the Sixth Circuit sanctioned two attorneys in Whiting v. City of Athens. Over two dozen fake citations and misrepresentations turned up in their appellate briefs, and the court wrote that this was a conservative estimate. The sanction came to $15,000 from each attorney, full payment of the other side's appellate attorney fees, and double costs.

The part worth watching is not the AI. When the court asked how the briefs had been reviewed and whether generative AI had been used, the two attorneys did not answer and challenged the court's authority to ask. The opinion mentions AI without establishing as fact that AI was used. The ground for the sanction was the duty of candor to the court, which was there long before AI. Without proving that any tool was used, the failure to confirm that the citations written down were real is enough for the sanction to stand.

What the court held in its hand to enforce that duty was not a new rule either. The opinion names Federal Rule of Appellate Procedure 38 and the court's inherent authority: a rule that puts the cost of a groundless appeal on the party that brought it, and the long-standing power to punish a party for carrying litigation in bad faith. Not one provision written with AI in mind was used.

The measures the National Court Administration's judges chose aim at the same spot. None of the four bans the use of AI. They ask whether it was checked (the duty to verify), require that the route be written down (the duty to state AI use), and make the side that filed the false material pay for it (administrative fines). The direction is record and responsibility rather than prohibition.

5

Why Pebblous Is Watching This Survey

The three the court settled on are moves that people who handle data have used for a long time. Confirming that a source is real. Recording which tools and which route something came through. Putting the cost of an error on whoever introduced it. They go under the names provenance, audit trail and accountability. This is not a problem that gets solved by building a bigger model. It is the problem of writing down where the grounds for a single sentence came from.

A courtroom has two readers standing between a filing and the record. Opposing counsel takes the brief apart, and the bench searches the docket number. That is why the fake case law was caught. Somebody actually opened it, and that is how it became clear that the case does not exist.

Reports, proposals and contracts that AI writes inside a company often have neither reader. Nobody is reading from the opposite side, and in most places nobody opens the sources in the footnotes one at a time. That leaves three things to check now.

  • Whether a procedure exists in which somebody actually opens the sources behind the figures and citations in that document. If it does not, the place the court called a duty to verify is standing empty
  • Whether the document itself keeps a record of which paragraph came through which tool. Once it has gone outside and something goes wrong, that record is the only place left to go back to
  • Who bears the cost when false grounds leave the building. If the structure is one where nobody bears it, nobody checks

What the National Court Administration survey showed is not that AI tells lies. It is that the route by which a lie reaches an official document unfiltered is already open. In the courtroom a judge stood at the end of that route, and the lie was caught. It is worth asking who stands at the end of a company document.

Thank you for reading this far. The survey period and the 360 respondents, the 243 (67.5%) who met one and the per-type counts of 196 (80.7%), 159 (65.4%), 41 (16.9%) and 10 (4.1%), the 69.5% and 57.6% naming who filed them, the 91.7% calling for institutional measures and the 64.8% answering "strongly necessary", the shares for the four measures, the fake judgment documents and the AI conversations left inside filings, the push to amend the civil and criminal procedure acts and the electronic litigation notice under consideration, and the two statements from the National Court Administration were all confirmed in the survey results published on November 2 and in the reporting that carried them. The worldwide tally of 2,202 and its breakdown by filer come from the October 19 update of the AI Hallucination Cases database. The date and the sanction in the Sixth Circuit case, the account of the attorneys declining to say whether generative AI was used, and Federal Rule of Appellate Procedure 38 and the court's inherent authority as the grounds cited follow the opinion itself and the legal press analysis of it. Reading the three measures as provenance, audit trail and accountability at the end of Section 4 and through Section 5, and the three questions aimed at company documents, are this article's own. If you check the sources behind AI-written documents inside your own organization, we would be glad to hear how.

Pebblous Data Communication Team
November 3, 2026

R

References

Reporting on the National Court Administration Survey

Overseas Tally and Case Law