Pro Se Litigants and AI Fake Citations: What Courts Have Done

This guide covers what the data shows, how courts have responded in recent decisions, and what a self-represented party should do before filing. Every case below links to the written decision or to its tracker entry. The underlying database is Damien Charlotin's AI Hallucination Cases Database, and we credit it for the case collection.

What the data shows

The counts below come from the tracker dataset at aivortex.io/legal-ai-risk, which is built on the Charlotin database. The snapshot has 2,202 decisions dated through October 15, 2026. A decision counts as pro se when the tracker carries its `pro-se` tag, the label for self-represented parties.

Two cautions. The database grows as people find and submit decisions, so recent months are still filling in and the yearly counts are not an incidence rate. And the tag reflects how the decision was classified, not a count of everyone who ever filed an AI-generated citation. Our statistics page explains how we read the dataset.

What courts did to pro se filers

Each decision in the tracker has one or more outcome labels. Among the 1,268 pro se decisions, the labels break down like this. A decision can carry more than one label, so the rows add up to more than 1,268.

A warning is the most common result. That fits the pattern in the cases below: courts often explain the problem, say that Rule 11 or its state equivalent applies, and hold the harsher step in reserve for the next filing.

Why "I'm not a lawyer" doesn't help

Federal Rule of Civil Procedure 11 applies to "an attorney or unrepresented party" who signs a filing, and it requires that the legal contentions be warranted by existing law. You can read the rule text at Cornell's Legal Information Institute. Courts do give self-represented parties some leeway on form and procedure. The recent decisions below show that the leeway stops at citations that don't exist.

The practical point is that the person who signs answers for the citation. The tool that produced it does not.

Recent pro se cases

These are decisions listed in the tracker with October 2026 dates. The links go to the written decision hosted by the Charlotin database and to the tracker entry. Where we've written a summary, it's linked too. The outcome column is the tracker's label for the result.

What the recent cases show

Our own summaries of these decisions, each linked above, point to a few patterns.

• The first response is usually a warning. In Waushara County, the Wisconsin Court of Appeals found nonexistent and unrelated citations in a traffic appeal, declined to sanction, and warned about generative AI. Sabatini and Thornton ended the same way: a warning, not a penalty.
• Costs come when the problem persists. In Davis, the Indiana Court of Appeals found that the briefs rested on nonexistent cases and miscited law, and remanded to set appellate fees against the self-represented appellant.
• Some courts add a process requirement. In King v. AT&T, the judge found Rule 11 violations, skipped fees and dismissal, and ordered a sworn certification about AI use on every filing.
• A bad citation can lose the argument even without a sanction. In Deen, the court flagged a nonexistent case, rejected the argument that relied on it, and affirmed.

None of this is a forecast. Each judge decides on the facts, and the tracker labels are a summary, so read the decision itself before you rely on any of it.

Where Mata v. Avianca fits

Mata v. Avianca is the 2023 case most people cite first, and it involved lawyers, not self-represented parties. It matters here because it set the template: a court found citations to cases that didn't exist and put the signer's duty to check at the center. The pro se decisions apply the same logic to people without a license. The Mata page now carries a short list of the recent pro se decisions as well.

A pre-filing check for self-represented parties

If you're filing without a lawyer and you used any AI tool to find or write citations, this is the minimum.

1. Open every case in a primary source. Use the court's own website, a government site or a free case law database. If you can't find the case there, don't cite it.
2. Read the part you cite. A real case can still say something different from your sentence. Find the passage and compare.
3. Check every quotation word for word. Altered or invented quotes get flagged separately from invented cases.
4. Write down what you checked. A short list with the date and source for each cite is enough, and our citation verification log gives a format.
5. Read your court's rules on AI. Some courts require a disclosure or a certification. What courts require on AI disclosure lists examples, and the rule in your court can differ.
6. If you find a bad citation after filing, fix it quickly. Tell the court and the other side and withdraw or correct the filing. Mata shows that delay made the outcome worse.

For a longer walk-through, see how to verify AI legal citations and what happens if AI hallucinates in a court filing.

Sources

• Damien Charlotin, AI Hallucination Cases Database (Case collection behind the tracker)
• AI Vortex legal AI risk tracker, dataset API (Counts read October 21, 2026)
• Federal Rule of Civil Procedure 11, Cornell LII

Operational information, not legal advice. Counts are a snapshot of the tracker dataset read October 21, 2026 (decisions dated through October 15, 2026); the database keeps growing. Case descriptions follow the linked decisions and our case summaries; read the decision itself before relying on one.