MMOOZZGG.com

AI for lawyers

AI hallucinations in court: a neural network invented case law in 2,190 cases

7 min read15 sources

AI hallucinations in court are no longer a curiosity: as of 15 October 2026, the global database kept by Damien Charlotin lists 2,190 cases where filings contained decisions, citations or quotes invented by a neural network. The US has the most such cases, then Canada and Australia, and in 2026 Russian arbitration courts issued their first fines too. Below: the numbers by country, the best-known cases and a citation check process that removes this risk.

On this page · 5 sections

The database numbers by country

The database of AI hallucination cases in courts is kept by researcher Damien Charlotin. As of 15 October 2026 it holds 2,190 cases, and more than two thirds of them come from the US.

The split by country on that date:

1. US: 1,492 cases.

2. Canada: 232.

3. Australia: 112.

4. United Kingdom: 76.

5. Israel: 58.

6. Russia: 1.

The Russian row lags behind practice: in 2026, the Arbitration Court of the West Siberian District, the Intellectual Property Rights Court and the Arbitration Court of the Vologda Region all reported fines for invented case law. The count keeps growing, so we update the figures on this page once a month.

Well-known cases

The first case the whole market heard about was Mata v. Avianca. On 22 June 2023 Judge P. Kevin Castel (Southern District of New York) fined two lawyers and their firm 5,000 dollars for a claim citing six non-existent decisions produced by ChatGPT.

The fine was small, but the court ordered the lawyers to write to each judge whom ChatGPT had named as the author of an invented decision. The lawyers began to admit the problem only on 25 May, after several requests from the court, and that is what the judge held against them.

In December 2023 a similar story happened to Michael Cohen. He gave his lawyer citations to non-existent decisions found through Google Bard, and a motion with them reached Judge Jesse Furman.

In June 2025 the High Court of England heard two such cases together. In Ayinde v Haringey, lawyers cited 5 non-existent sources; in Al-Haroun v Qatar National Bank, 18 fictional cases.

Dame Victoria Sharp, President of the King’s Bench Division, reminded in this judgment that the lawyer is responsible for the accuracy of citations. The software the lawyer used does not take that duty away.

All three stories repeat one pattern: the AI gave plausible case details, and the document went to court without anyone opening the decisions behind those details.

Russia: two fines in 2026

In May 2026 the Arbitration Court of the West Siberian District fined a company $600 for citing, in a cassation appeal, non-existent court acts created by a neural network (case No. A27–7831/2025). According to media reports, this is the first such fine in Russia.

The court said directly that preparing an appeal with AI is no excuse. Under Part 2, Article 9 of Russia’s Arbitration Procedure Code, the party to the case is responsible for the accuracy of generated text.

In October 2026 the Arbitration Court of the Vologda Region imposed a court fine of $350 on a plaintiff who cited four rulings of the Arbitration Court of the North-Western District generated by a neural network (case No. A13–14310/2025). The court treated this as contempt of court.

The plaintiff explained that he had copied the case law from open sources on the internet but could not name the source. The court answered that parties may not knowingly cite non-existent guidance or give knowingly incorrect quotes from court acts.

These are two fines on companies. Between them, according to Pravo.ru of 28 August 2026, the Intellectual Property Rights Court fined an individual entrepreneur $60 for 12 false citations in a cassation appeal (case No. A71–11377/2025).

In ruling No. 301-ES21–23756, the Supreme Court of Russia set the general rule: a party prepares a position made with software at its own risk and must check the result. The reasoning of all the rulings is collected on the page about fines by Russian courts.

Why AI invents things

A public chatbot does not look up the answer in a database of court acts. The guidance for judges in England and Wales explains it this way: the model generates the most likely combination of words from the request and its training data, and this is not necessarily an accurate answer.

That is why invented details look convincing: the right number format, a familiar court name, a smooth quote. The same guidance lists directly that AI may invent cases, citations, quotes and non-existent laws.

Hence the split of tasks in the guidance. It calls summarizing large texts and writing emails potentially useful, and it does not recommend public chatbots for searching new information that cannot be checked, or for legal analysis.

Specialized legal services make fewer mistakes, but they still make them. In May 2024 Stanford RegLab and HAI found errors in more than 17% of answers from Lexis+ AI and Ask Practical Law AI and in more than 34% of answers from Westlaw AI-Assisted Research.

In an earlier study, general chatbots were wrong on 58–82% of legal questions. The difference comes from access to a database: a service that relies on real court acts invents less, but without a human check the error still reaches the court.

In its concept paper, the Supreme Court of Russia rated the likelihood of model hallucinations as high and their impact as very high. The measures for courts: an answer with a link to the source, a check of the result by the judge, and a ban on public AI services.

Checking citations: the process

What catches an invented citation is not attention but a routine the lawyer follows before every filing. The Vologda case shows why checking the number is not enough: the case numbers in the plaintiff’s citations were real, but they belonged to other disputes.

1. Open each act by its details in the primary source: the arbitration case files database, the court’s website, the official publication of the law.

2. Check that the number belongs to the same dispute, the same parties and the same court level as stated in the document.

3. Find the quote word for word in the text of the act. If it isn’t there, state the conclusion in your own words with a reference to the paragraph of the reasoning, or remove it.

4. Check that a higher court has not overturned the act and that the rule is in force in the right version.

5. Record in a check log who opened each citation and when, and only then sign the document.

Courts already ask about checks directly. Since 30 May 2023, Judge Brantley Starr in Texas has required a certificate that a person checked any text prepared by generative AI. The guidance for judges in England and Wales allows asking for confirmation that the lawyer checked the citations personally.

Courts are building the same routine for themselves. The risk matrix in the Supreme Court’s concept paper includes automatic checks of citations to court acts and laws, mandatory disclosure of AI use when filing documents, and procedural sanctions.

The first two steps can be given to an AI agent inside the firm’s closed network: it checks each citation against the database of acts and flags those it can’t find, while the decision and the signature stay with the lawyer. How this works is covered on the page about case law search.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. AI Hallucination Cases Database, 15.10.2026 (damiencharlotin.com)
  2. ABA Journal, 22.06.2023 (abajournal.com)
  3. NPR, 30.12.2023 (npr.org)
  4. Legal Cheek, 09.06.2025 (legalcheek.com)
  5. Peterburg2 citing Rossiyskaya Gazeta, 19 May 2026 (peterburg2.ru)
  6. RAPSI, 5 October 2026 (rapsinews.ru)
  7. Stanford HAI, 23.05.2024 (hai.stanford.edu)
  8. Rossiyskaya Gazeta, 18 May 2026 (rg.ru)
  9. Arbitration Court of the Vologda Region, 5 October 2026 (vologda.arbitr.ru)
  10. Pravo.ru, 5 October 2026 (pravo.ru)
  11. Pravo.ru, 28 August 2026 (pravo.ru)
  12. RAPSI, 9 September 2026 (rapsinews.ru)
  13. Courts and Tribunals Judiciary, AI Judicial Guidance, 31.10.2025 (judiciary.uk)
  14. Garant, text of the order, 3 September 2026 (base.garant.ru)
  15. The Volokh Conspiracy (Reason), 30.05.2023 (reason.com)

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