AI Fake Legal Citations: 8 in 10 Korean Judges Report Them

South Korea’s Supreme Court said on Monday, November 2, that about 80 percent of the civil, family and administrative judges in a survey have run into court filings that cite nonexistent or false rulings. The court’s administration office ran the survey from October 6 to 16, and 360 judges responded, according to Kyodo News and the Korea Times, both relaying Yonhap. Newspim, a Korean outlet, says the results came in a press release from the court administration.

SBS News says this is the first court survey of judges on the question.

What the survey found

Among judges handling civil, family and administrative cases, 196 said they had seen citations of nonexistent rulings. That was 80.7 percent, per the Korea Times. Another 159 (65.4 percent) had seen real-looking rulings that were misquoted or false. Fabricated laws or regulations turned up for 41 judges (16.9 percent), and invented papers or publications for 10 (4.1 percent).

The Seoul Economic Daily added examples of what judges reported. In one case a party used AI to generate a fake judgment that copied the format of a Supreme Court decision. Some filings listed Supreme Court case numbers next to legal reasoning the judges had never seen. In a few, transcripts of the person’s chat with the AI, or notes on litigation strategy, had been left in the document.

Nearly 92 percent of respondents (91.7 percent, per the Seoul Economic Daily) said the courts need an institutional response. Top of the list was a legal basis for fining people who file fake citations, at 61.4 percent, per the Seoul Economic Daily. Requiring parties to verify what they cite got 61.1 percent, requiring disclosure of AI use got 60.6 percent, and disciplinary requests against lawyers got 55.4 percent. All of these survey figures rest on news reports of the court’s release.

Who cited the fakes

The reports part ways here.

Outlet	What it says about the 164 and 136 figures
Korea Times	164 judges “said they themselves have cited false precedents”; 136 (57.6 percent) said lawyers did
SBS News	164 respondents (69.5 percent) pointed to litigants themselves; 136 (57.6 percent) pointed to attorneys
Seoul Economic Daily	69.5 percent of judges pointed to the parties, 57.6 percent to attorneys

My read: the Korea Times line looks like a slip. Judges aren’t the ones filing arguments, and 164 and 136 both turn into percentages over the same base. Dividing them by their stated shares gives about 236 judges, so both numbers describe who the judges saw cite the fakes, not what the judges did. SBS and the Seoul Economic Daily read it that way. I couldn’t open the court’s own release, so I can’t rule out a wording quirk in the original.

Same arithmetic, second catch: the 80.7 percent is measured against roughly 243 judges, not all 360. That’s my calculation from the published percentages, and no outlet states it. It would mean “8 in 10” applies to the civil, family and administrative judges who answered that question, which is how the Korea Times phrases it, not to every judge in the country.

The figures are also self-reported by judges, and none of the reports I read says how many filings that adds up to. My inference: a judge who saw one fake citation and one who saw fifty probably both count once.

Why a judge’s time is the cost

An official at the National Court Administration said, as quoted by the Seoul Economic Daily, that fabricated citations force the other side and the court to spend extra time checking whether a cited ruling exists, and that this can slow hearings. It is also reviewing a plan to post guidance on AI use on its electronic litigation website.

My read: that is a modest step next to what judges asked for, since a posted warning doesn’t fine anyone. The court says it is also working on the bigger one. It judges that fines need changes to the Civil Procedure Act and the Criminal Procedure Act, and it plans to consult the Ministry of Justice and submit its views to the National Assembly, the Seoul Economic Daily reported. This isn’t new work, either. Chosun Biz reported in March that a court task force had already proposed fines for citing false statutes and a duty to tell the court when AI is used.

My read

I don’t think this is mainly a Korean story. Chatbots producing a convincing case number for a ruling that never happened is a known failure. Models still get it wrong. Artificial Analysis reports an 11 percent hallucination rate for Google’s Gemini 4 Carbon on its AA-Omniscience v1.1 benchmark, up from 4 percent for the previous model but still the second lowest it has measured. That’s a test of hard knowledge questions, not legal citations, so it only shows the failure is measurable. I covered it in the Gemini 4 Carbon hallucination post. What the Korean survey adds is a count from the receiving end, from the people who have to look each citation up.

Boring fixes won: check what you cite, and pay a price if you don’t. I’d expect other courts to land in the same place.

Note on sources

I could not find or read the National Court Administration’s release itself, which is in Korean, so there is no direct link to it. The survey figures come from the news outlets listed below (the Seoul Economic Daily’s English edition is machine-translated from Korean), and I’ve named the outlet next to each one. The 11 percent figure is from Artificial Analysis, and the March task force detail is from Chosun Biz. The 236 and 243 base sizes are my arithmetic from the published percentages and aren’t in any report.

About 8 in 10 South Korean civil-case judges have seen AI-invented rulings cited in filings