The Cited and the Uncited: How AI Research Is Rewiring the Courthouse

A judge reads a brief at 11 p.m. and stops on a citation that reads beautifully, the right court, the right year, a quotation that lands precisely on the point in dispute. The only problem is that the case does not exist. That single experience, repeated in courthouses around the world, has done more to define this technological moment for the judiciary than any vendor demo or conference keynote. The story of artificial intelligence in legal research and drafting is, at its core, a story about trust: who is accountable for an answer, how it was grounded, and whether the institution that issues it can stand behind every word.

The courts occupy a singular position in that story. They are simultaneously the regulators of how AI-assisted work may enter the record, the consumers of research tools for their own opinion drafting and clerk workflows, and the last line of defense for the millions of people who walk in without a lawyer at all. Each of those roles is being reshaped at once. To understand where this is heading, it helps to start with how research and drafting actually worked before any model could write a sentence.

The Old Way: Research by Hand, Drafting by Memory

For most of the modern era, legal research was a craft of physical labor and disciplined doubt. A clerk or associate began with a treatise, traced a doctrine through reporter volumes, and then "Shepardized" each authority, checking, citation by citation, whether a case had been overruled, distinguished, or quietly eroded. Drafting an opinion or an order meant assembling those verified strands into prose by hand, with every quotation pulled from a source the writer had physically opened. The system was slow, expensive, and unevenly distributed, but it had one defining virtue: a human being had read every case they cited.

That model strained under volume long before AI arrived. Federal and state dockets grew faster than the bench, and self-represented litigants, people navigating eviction, custody, debt, and benefits without counsel, became the majority of participants in entire categories of civil court. Studies have found that roughly 63 percent of individuals in civil matters appear without a lawyer, a share that climbs far higher in housing and family courts (Legal Aid Resources). For these litigants, the research-and-drafting craft was simply inaccessible: there was no associate to Shepardize, no treatise to consult, only a clerk's window and a form.

The scale of that exclusion is staggering. The Legal Services Corporation's most comprehensive measurement found that low-income Americans receive no or inadequate legal help for 92 percent of the civil legal problems that substantially affect them, with 74 percent of low-income households experiencing at least one such problem in a single year and 46 percent of those who skipped seeking help citing cost as the reason (Legal Services Corporation). Globally, the World Justice Project estimates that roughly 1.4 billion people have unmet civil or administrative justice needs, part of a justice gap that touches some 5.1 billion people (World Justice Project).

The Shift: Grounded Answers Meet a Hard Lesson

The arrival of large language models promised to collapse the research-and-drafting bottleneck. A tool could read a thousand-page record in seconds, surface the controlling authority, and draft a passable first cut of an order. Adoption moved quickly: industry surveys from the Thomson Reuters Institute found that the share of legal organizations actively using generative AI nearly doubled in a year, from 14 percent to 26 percent, with 95 percent of professionals expecting the technology to become central to their workflow within five years and roughly 72 percent of current users engaging with it at least weekly (Thomson Reuters Corporation).

Then came the hard lesson. Early adopters discovered that a fluent model, asked for case law, would sometimes invent it, producing confident, well-formatted citations to opinions that had never been written. A running database maintained by a legal data researcher now documents more than 1,641 rulings worldwide in which AI generated fabricated citations or fictitious authority, spanning courts in the United States, Canada, Israel, Italy, India, France, and beyond (Damien Charlotin, AI Hallucination Cases Database). The growth curve has been steep: independent reporting drawing on the same data counted roughly 10 such cases in 2023, 37 in 2024, and 73 in just the first five months of 2025 (Business Insider).

The judiciary's response has been swift and, characteristically, decentralized. Courts treated fabricated authority not as a novel offense but as an old one, a breach of the duty of candor and the obligation to make a reasonable inquiry before signing a filing. Sanctions followed. In a closely watched state appellate decision, a court imposed a $10,000 penalty on counsel whose briefs were "replete with fabricated legal authority," noting that 21 of 23 citations were misquoted or nonexistent, and published the opinion expressly "as a warning" (CalMatters). Other tribunals went further, with one Illinois matter producing a combined $60,000 in sanctions against a firm and an attorney, and a federal court electing to disqualify counsel rather than fine them (Klemchuk).

No brief, motion, or pleading should contain citations, whether from AI or any other source, that the responsible attorney has not personally read and verified.

The pattern that emerged is instructive. Courts consistently held that the existence of an AI tool changes nothing about the underlying duty: the human who signs the document is accountable for every authority in it, a principle traceable to the first major federal sanctions order on the subject (Legal AI Governance). And the problem was not confined to litigants. Reviewers have noted instances in which judges' own chambers issued opinions later found to contain AI-generated errors, a reminder that the verification burden runs in every direction (Sterne Kessler).

What It Looks Like Now: A Patchwork of Rules and a Risk Ladder

Out of that turbulence, a working governance model has taken shape. The dominant instrument is the standing order, a judge-by-judge directive requiring litigants to disclose AI use, certify that every citation has been verified, or both. By early 2026, more than 300 federal judges had issued such orders, up from a single pioneering order in May 2023, and an openly accessible tracker indexed well over 500 court directives across federal and state benches (AI Vortex). The result is a genuine patchwork: requirements that differ courthouse to courthouse, with no uniform federal rule binding them together (Hintyr).

State court systems are now layering structure on top. California became one of the first large systems to adopt a formal framework, requiring every superior, appellate, and supreme court that does not ban generative AI to adopt a written use policy by December 15, 2025, a rule reaching roughly 65 courts and some 1,800 judges (Judicial Council of California). Florida moved in a different direction, adopting a statewide citation-certification rule effective June 2026 that requires every attorney signing a filing to certify that the cited authorities exist and are accurately quoted (Law Tech AI).

Underpinning much of this is a shared conceptual tool: a tiered risk ladder. Guidance developed by the National Center for State Courts and the Thomson Reuters Institute's policy consortium sorts AI uses into minimal, moderate, and high risk, and assigns a matching level of human oversight to each. Summarizing a routine document is minimal risk and needs only a "human on the loop." Drafting an opinion or running research on a reliable platform is moderate risk and demands a "human in the loop" who verifies every citation. Any use that could affect a person's rights, predicting recidivism, informing sentencing, is high risk and requires the closest human control (NCSC / Thomson Reuters Institute).

Crucially, the same guidance separates the use of AI in non-adjudicative work, drafting an internal memo, summarizing intake forms, from its use in the act of judging itself. California's framework, for instance, explicitly bars judges from using AI to make judicial decisions, even as it permits administrative uses (Morgan Lewis). That line, between assisting the work and substituting for the judgment, is the philosophical spine of nearly every court policy now in force.

Beneath the policy debate, courts are quietly running practical experiments. Trends reporting from the state court community describes AI being deployed to help residents navigate housing issues, to screen guardianship and conservatorship reports for timeliness, and to flag potential errors for human review, concrete, bounded uses that keep a person at the center of the process (National Center for State Courts). To help laggard courts catch up, the NCSC released an AI Readiness framework with a self-assessment tool and a twelve-month governance roadmap (National Center for State Courts).

The Next Few Years: Grounding, Plain Language, and the Accountability Question

Where does the craft go from here? The clearest near-term trajectory is the shift from open-ended generation toward grounded, retrieval-based answering, systems that cite only authority drawn from a verified corpus and surface the underlying source for human inspection. The mechanical mitigation that courts have converged on is unforgiving but simple: every cited authority must be checked against a primary source before filing, a step that produces an auditable verification trail (Legal AI Governance). Expect that verification log to become a standard artifact of court practice, the way a certificate of service is today.

The second trajectory is plain language. The deepest promise of AI drafting for the courts is not faster opinions for lawyers but readable ones for everyone else, orders that a self-represented litigant can actually understand, generated and then reviewed by a human. With self-representation already the norm in much of civil court and the justice gap measured in the billions, even modest gains in comprehension and navigation could move more people through the system with their rights intact (Legal Services Corporation).

The third, and thorniest, trajectory is the accountability question turned inward. A 2026 random-sample survey reported that more than 60 percent of responding federal judges had used at least one AI tool in their judicial work, creating an unavoidable dynamic: the bench is regulating attorney AI use while increasingly relying on AI itself (PlatinumIDS). The next few years will test whether courts can hold themselves to the same verification standard they impose on the lawyers before them.

Conclusion: The Verified Future

The arc from hand-Shepardized briefs to machine-grounded answers is not really about replacing the lawyer or the judge. It is about relocating the one thing the old craft did well, the discipline of reading every authority before relying on it, into a faster, broader, and far more accessible system. The courts have learned, expensively, that fluency is not the same as truth. The institutions now building verification into the workflow, extending readable orders to the unrepresented, and drawing a bright line around the act of judgment are the ones turning a risky novelty into durable infrastructure. The cited and the uncited will always coexist; the work ahead is making sure the court can tell them apart.

Sources

1. Legal Services Corporation, The Justice Gap Study (2022)
2. Legal Services Corporation, The Justice Gap Report
3. World Justice Project, Global Insights on Access to Justice 2019
4. World Justice Project, Global Justice Gap report (PDF)
5. Legal Aid Resources, Key Studies on Self-Represented Litigants (PDF)
6. Damien Charlotin, AI Hallucination Cases Database
7. Business Insider, Fake Legal Citations on the Rise as Lawyers Misuse AI Tools
8. CalMatters, California Issues Historic Fine Over Lawyer's AI Fabrications
9. Klemchuk, Sanctions for Misuse of AI in Court Submissions Increasing Sharply
10. Legal AI Governance, AI Hallucinated Citations: Sanctions Case Tracker
11. Sterne Kessler, AI Hallucinations in Court Filings and Orders: A 2025 Review
12. Thomson Reuters, Generative AI Adoption Nearly Doubles (2025 Report)
13. NCSC / Thomson Reuters Institute, Principles and Practices for AI in State Courts (PDF)
14. National Center for State Courts, Trends in State Courts 2025 (PDF)
15. National Center for State Courts, AI Readiness for the State Courts
16. Judicial Council of California, Rule of Court 10.430
17. Morgan Lewis, California Courts Adopt Rule Governing Generative AI Use
18. Law Tech AI, AI Policy Tracker (Florida Rule 2.515)
19. AI Vortex, How Many Federal Judges Require AI Disclosure
20. Hintyr, AI Disclosure in Court: Rules You Need to Track
21. PlatinumIDS, Inside the AI Hallucination Crisis Hitting Courts Worldwide