Courts, lawyers, and litigants will be better served by policies and habits that treat AI output as a draft to be verified, not a research result to be repeated.
For years, bad citations in a brief usually meant a rushed deadline, a sloppy edit, or a misunderstanding of precedent. Now, courts are confronting a different problem: filings that cite cases that never existed, quote language no judge ever wrote, or describe holdings that real opinions do not support.
New research suggests this isn’t a niche issue anymore. It’s becoming routine enough that judges, clerks, and opposing counsel are spending real time separating legitimate legal argument from AI-generated fiction—and doing so without consistent guidance on what the consequences should be.
New data shows the problem is accelerating in U.S. courts
According to the AI Hallucination Cases Database maintained by legal researcher Damien Charlotin, there were 574 U.S. court cases involving AI-generated errors in the first half of 2026. That already surpasses the 526 documented across all of 2025.
A separate analysis by LawDistrict breaks down what these errors look like in practice, why they’re spreading, and how unevenly courts are responding. The full findings and methodology are available here: AI Hallucinated Court Cases.
At the current pace, LawDistrict projects roughly 1,200–1,300 documented cases by the end of 2026—numbers that would have sounded implausible even a year ago.
What the errors actually are: not just “typos,” but invented authority
“AI errors” can sound abstract until you see the patterns. The LawDistrict analysis reviewed 3,609 documented errors and found that most were not minor misstatements. They were failures that go to the heart of legal credibility: whether a court can trust what’s being cited.
- 53% were fabricated citations—cases that do not exist.
- 26% misrepresented real cases—an opinion exists, but the description of it does not match.
- 20% involved false quotations—language attributed to a judge that cannot be found in the opinion.
These categories matter because they often require different fixes. A mistaken pin cite may be corrected with a revised filing. A made-up case can trigger deeper concerns about diligence, competence, and candor to the tribunal.
Who is most at risk: self-represented litigants, but not only them
One of the clearest signals in the data is who is behind many of the problematic filings. LawDistrict found that 60% of the U.S. cases involved self-represented litigants.
That isn’t particularly surprising. People who can’t afford counsel may turn to general-purpose AI tools for help drafting motions or responding to deadlines. But it does sharpen a practical question courts have to answer: how do you deter misuse without punishing pro se parties who may not understand that an AI model can generate plausible-sounding—but nonexistent—authority?
It also underscores why reliable public-facing guidance is important. The Federal Judicial Center, for example, has tracked and discussed emerging issues courts face as technology changes litigation practice, including questions tied to AI use.
Sanctions are rising, but court responses remain inconsistent
There’s a tendency to assume the outcome is predictable: submit fake citations, get sanctioned. The data paints a messier picture.
LawDistrict reports that recorded monetary sanctions tied to AI-generated errors have exceeded $1.15 million. Yet the analysis also highlights inconsistent court responses—similar errors can lead to very different consequences, including:
- warnings or orders to show cause
- mandatory correction or refiling
- substantial fines
- suspension
- referral to a state Bar
That inconsistency matters for two reasons. First, lawyers and litigants need clearer signals about what courts expect when AI tools are used. Second, uneven enforcement can encourage risk-taking: if the worst outcome seems like a slap on the wrist, some parties may gamble.
For baseline expectations, courts and practitioners frequently look to professional responsibility frameworks like the American Bar Association’s model rules and related commentary, as well as state-specific guidance.
Practical takeaways for litigants, lawyers, and court systems
Courts don’t need to ban AI to address this. But they do need a shared floor for what “reasonable verification” looks like—especially when citations and quotations are involved. Based on the trends reflected in the database and LawDistrict’s breakdown, a few steps are becoming hard to argue with.
For legal professionals
- Verify every citation in a primary source database before filing—especially anything produced by a generative tool.
- Quote-check like you mean it: confirm the exact language appears in the opinion and matches the context.
- Keep an audit trail of research steps when using AI-assisted workflows, so diligence can be shown if challenged.
For self-represented litigants
- Do not assume an AI tool is a legal research platform.
- Cross-check case names and citations on free, reputable sources where possible.
- Ask the clerk’s office about self-help resources and local rules before filing.
For courts and administrators
- Clarify expectations through standing orders or local rules focused on citation verification and disclosure where appropriate.
- Standardize responses so that consequences are not wildly different for similar conduct.
- Invest in education for pro se communities, where the data suggests the risk is concentrated.
As a general matter, the National Center for State Courts continues to publish research and practical resources on how state systems can adapt to technology-driven shifts in litigation.
Closing: the actionable next step is simple—treat AI output as unverified draft material
The headline number—574 U.S. cases with AI-generated errors in just the first half of 2026—tells you the issue is no longer hypothetical. And the details matter: fabricated citations are the majority, self-represented litigants are heavily represented, and real money is already being assessed in sanctions.
If there’s one practical conclusion that follows from the data, it’s this: AI can help with structure and language, but it cannot be trusted as a source of authority. Courts, lawyers, and litigants will be better served by policies and habits that treat AI output as a draft to be verified, not a research result to be repeated.
Join the conversation!