AI Legal Research and Lawyers: What the Court Records Show

What firms actually run

The firm-wide deployments are real, and the firms themselves confirm them. In February 2023 Allen & Overy said it had built Harvey, "built on a version of Open AI's latest models enhanced for legal work", into its global practice for more than 3,500 lawyers across 43 offices. During a beta that began in November 2022, around 3,500 of its lawyers had asked Harvey around 40,000 queries for their day-to-day client work. The firm added that "the output needs careful review by an A&O lawyer."² In December 2023 it launched ContractMatrix, its own contract drafting and negotiation tool built with Microsoft and Harvey. More than 1,000 of its lawyers were already using it internally, and the firm said it grounds the AI's output in banks of "gold standard precedents" to reduce hallucinations.⁴

Canada's Borden Ladner Gervais adopted Thomson Reuters' CoCounsel firmwide in December 2025, after piloting it "through multiple versions". The firm says the assistant works alongside the Westlaw and Practical Law databases for research, drafting and document analysis, and that lawyers, law clerks, paralegals and business services staff can all use it.³ Paralegals and legal assistants (SOC 23-2011.00) are a separate occupation in O*NET, but they do much of the same research and drafting.⁹ These announcements tell us what the firms run and for whom. None of them reports how often a reviewing lawyer catches an error.

What the court records count

The errors that do reach a judge are now tracked. The AI Hallucination Cases Database, compiled by the researcher Damien Charlotin and published under a CC BY 4.0 license, records legal decisions in cases where generative AI (software that writes text from a prompt) produced hallucinated content: "typically fake citations, but also other types of AI-generated arguments." It listed 2,184 cases at its update of October 10, 2026.⁷

Its scope is narrower than "every fake citation". It covers decisions where a court or tribunal explicitly found, or implied, that a party relied on hallucinated material, and leaves out mere allegations. As an exception, it also includes some decisions where AI use was alleged but not confirmed, which the compiler calls "a judgment call on my part". It does not track "the (necessarily wider) universe of all fake citations or use of AI in court filings."⁷

By the database's own count, decisions rose from 6 in the second quarter of 2023 to 124 in the second quarter of 2025. They peaked at 454 in the first quarter of 2026, and 356 were recorded for the third quarter. The United States accounts for 1,492 of the cases, followed by Canada (233), Australia (112) and the United Kingdom (76).⁷

Who files the fabricated citations

Most of the filings didn't come from lawyers. The database records 1,259 decisions involving pro se litigants (people representing themselves) and 864 involving lawyers. Judges account for 32, experts for 15 and prosecutors for 5. In every full quarter from late 2024 to the third quarter of 2026, pro se litigants filed between 50% and 60% of the material.⁷ In the court records, the problem is as much about people representing themselves as about professional practice.

Most decisions don't say which tool produced the error. The database identifies one in 237 decisions: ChatGPT or another OpenAI product in 139, Westlaw or CoCounsel in 18, Claude in 18, Google Gemini or Bard in 17, Microsoft Copilot in 14 and LexisNexis in 13. Legal research products appear on that list as well as general chatbots.⁷ Of 2,162 decisions, leaving out 22 where AI use was only alleged, it tags 380 with a monetary sanction and 163 with a referral to a disciplinary body.⁷ We haven't opened each order, so we don't total or compare the amounts. The database's download file also uses placeholder amounts in some rows.

Mata v. Avianca: the case that set the pattern

The best-known early case is a passenger's injury claim against an airline in the Southern District of New York. The court's sanctions opinion begins: "there is nothing inherently improper about using a reliable artificial intelligence tool for assistance. But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings." It found that two lawyers and their firm "submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT, then continued to stand by the fake opinions after judicial orders called their existence into question."⁶

Under Rule 11, or alternatively its inherent authority, the court ordered them to send letters to their client and to each real judge named as the author of the six fake opinions. It also ruled that "a penalty of $5,000 is jointly and severally imposed", payable to the court.⁶

Why research tools built for law still err

Legal research vendors use retrieval-augmented generation (RAG): the model first retrieves documents from a legal database, then writes its answer from them. Some vendors marketed this as "eliminating" hallucinations or as giving "hallucination-free" citations. A preregistered Stanford evaluation of more than 200 queries found the claims overstated. Lexis+ AI, Westlaw AI-Assisted Research and Ask Practical Law AI each hallucinated between 17% and 33% of the time, less often than the general chatbot GPT-4. Lexis+ AI answered 65% of the queries accurately and Westlaw's tool 42%.¹¹ Those rates describe the 2024 versions the researchers tested, and the products have changed since.

The same tools can still speed up the work. In a randomized trial, upper-level law students completed six legal tasks with Vincent AI (a RAG tool), with OpenAI's o1-preview reasoning model, or with no AI. Both tools improved the quality of the work and raised productivity by between 50% and 130% on five of the six tasks. The authors report that Vincent AI "avoids introducing more hallucinations", while o1-preview improved analytical depth.¹⁰ Two law firms and two law schools funded the study, and the vendors gave participants free access.¹⁰ Taken together, the two studies show gains in speed and quality alongside hallucination rates that weren't zero, and every firm deployment above keeps a lawyer checking the output.

What courts and bars now require

Courts have taken different approaches. The Northern District of Texas adopted a local rule, effective September 3, 2024, under which a brief prepared using generative AI "must disclose this fact on the first page". A brief without that disclosure certifies that no part of it was prepared with generative AI.¹² One of the court's judges, who had required his own AI certificate, says he no longer does because of the new rule.¹³ Illinois went the other way. Its Supreme Court's policy, effective January 1, 2025, says AI use "should not be discouraged" and that "Disclosure of AI use should not be required in a pleading", but that "All users must thoroughly review AI-generated content" before submitting it.¹⁴

The American Bar Association's Formal Opinion 512, issued July 29, 2024, applies the profession's existing duties to generative AI: competence, confidentiality, communication, supervision, meritorious claims, candor toward the tribunal and reasonable fees. It warns that "uncritical reliance on content created by a GAI tool can result in inaccurate legal advice to clients or misleading representations to courts". Before filing, it says, lawyers must review AI outputs, "including analysis and citations to authority". A lawyer who bills by the hour may charge only for the time actually spent.¹⁵ In England and Wales, the High Court's Divisional Court said in June 2025 that freely available generative AI tools "are not capable of conducting reliable legal research". It held that lawyers who use them have a professional duty to check the research against authoritative sources.¹⁶

What the records don't show

Court records capture failures that reached a judge. They don't capture AI-assisted research that was checked and was correct, and the firms that use these tools publish adoption figures rather than error rates. The two core tasks where AI would act in the lawyer's place, advising clients and representing them, have no deployed case. The one company that marketed a "robot lawyer" faced an FTC order that requires $193,000 in monetary relief. The order bars it from claiming that its service performs like a real lawyer unless it has evidence to back that up.⁸ For now, the evidence puts AI at the research and drafting desk, assisting lawyers and paralegals who sign off on its work.