MMOOZZGG.com

AI for lawyers

AI hallucinated case law: how to search court practice and check every citation

5 min read7 sources

AI can find court decisions by the meaning of a dispute, not by matching words, but it can also invent a case with a believable number. This is AI hallucinated case law. By 15 October 2026, Damien Charlotin’s database listed 2,190 court cases about hallucinated AI citations. So check every case number against the primary source before it enters a document.

On this page · 5 sections

Search by meaning

Classic search looks for words: “penalty”, “reduction”, “clear disproportion”. If a court wrote “decrease of the sanction”, that decision does not appear in the results.

AI searches by meaning. The lawyer describes the situation in their own words: a supplier missed a deadline because a subcontractor was late, the buyer demands a penalty, and the supplier claims force majeure. The model finds decisions where courts assessed similar facts, even if the text uses other words.

A useful search result has:

1. A list of decisions with the case number, court, date and instance.

2. A quote from the reasoning part where the court answers your question.

3. The outcome: in whose favor the issue was decided and whether the decision held up in higher courts.

4. Grouping by position: which arguments courts accept and which they reject.

The search runs over the database of rulings the firm connected and over the archive of its own cases. The model does not answer from memory: the answer is built on the texts it found, and each conclusion has a link to a document.

Why AI invents cases

A language model predicts believable text. For it, a case number, a court name and a quote are just words. If no suitable decision exists, the model may put one together from similar ones.

Numbers from independent tests: in May 2024, Stanford RegLab and HAI checked legal research services. They found errors in more than 17% of answers, and in more than 34% for one of the services. In an earlier study, general chatbots got 58–82% of legal questions wrong.

The textbook example: on 22 June 2023, Judge P. Kevin Castel fined two lawyers and their firm $5,000 in Mata v. Avianca. Their filing cited six non-existent decisions that ChatGPT had produced.

The risk goes down not through a ban but through the process: answers only from found texts and a separate check of each citation. Cases from different countries are collected in our review of AI hallucinations in courts.

Checking the case number and the text of the decision

The check happens before a citation enters a document, and it has three steps.

1. Case number: a decision with this number exists in the court’s records, and the court and date match.

2. Text: the quote appears word for word in the decision and is not assembled from several paragraphs or from another case.

3. Status: a higher court has not reversed or changed the decision in the part you cite.

If even one step fails, the citation gets a flag and does not enter the draft. The lawyer sees the reason: number not found, quote differs from the text, decision reversed.

References to laws are checked the same way: the article exists in the current version and says what the document claims. AI shows the version date next to the quote, so the lawyer can see whether the rule changed after the decision.

Responsibility still stays with a person. The lawyer opens the decision by the link and reads the whole paragraph before signing the document: the machine finds, and the lawyer answers for what goes to court.

Court fines in Russia

Russian courts already fine parties for invented citations. In May 2026, the Arbitration Court of the West Siberian District fined a company $600. Its cassation appeal cited non-existent rulings created by a neural network. According to media reports, this was the first such fine in Russia.

In October 2026, the Arbitration Court of the Vologda Region imposed a court fine of $350 on a claimant. The claimant had cited four district court rulings generated by a neural network.

Part 1 of Article 119 of Russia’s Arbitration Procedure Code sets the limits of a court fine. They are up to $60 for individuals, up to $350 for officials and up to $1,200 for organizations. The fine is smaller than the loss a client suffers when the court stops trusting the other arguments of the appeal.

In Damien Charlotin’s global database, as of 15 October 2026, Russia has one case and the United States 1,492. New Russian cases with case numbers are collected on our page about court fines for AI.

How our check works

In our projects, the citation check is built into the process and does not depend on a lawyer’s attention at the end of the working day. The model searches only the connected sources: the database of rulings the firm chose and the archive of its cases. The answer is built from the texts found, not from the model’s memory.

A separate step checks each case number and each quote against the primary source. An unconfirmed citation does not enter the draft, and the lawyer sees which citations are checked and which are not.

AI actions are logged: the firm sees who searched, what the model found and which citations the lawyer kept in the document. This helps if a client or partner asks how the document was prepared.

A pilot starts with one category of disputes, for example debt recovery under supply contracts. It measures two numbers: the time to find case law and the share of citations that failed the check.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. AI Hallucination Cases Database, 15.10.2026 (damiencharlotin.com)
  2. Stanford HAI, 23.05.2024 (hai.stanford.edu)
  3. ABA Journal, 22.06.2023 (abajournal.com)
  4. Peterburg2, citing Rossiyskaya Gazeta, 19 May 2026 (peterburg2.ru)
  5. RAPSI, 5 October 2026 (rapsinews.ru)
  6. ConsultantPlus, Arbitration Procedure Code of the Russian Federation, Article 119, 24 July 2002 (consultant.ru)
  7. Data security when a company adopts AI

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us