A pro se litigant can walk into court with an AI-drafted motion that looks better than anything they could have written alone. The timeline is cleaner. The legal issue has a name. The paragraphs sound like they belong in a filing. Then a judge or opposing lawyer checks the cases and finds that one of them does not exist, or that a quoted passage says something different from what the motion claims.
That is where the safety question around Legal AI for pro se filings becomes concrete. The risk is not that artificial intelligence writes in an unfamiliar style. The risk is that a self-represented person signs and files a document containing legal authority they have not independently verified.
In January 2026, the U.S. Court of Appeals for the Seventh Circuit put the point plainly in Jones v. Kankakee County: “pro se litigants shoulder responsibility too” for AI-generated filing errors.[1] That sentence matters because it removes the most tempting defense. A party may be unrepresented. They may have used a tool because the system is hard and counsel is expensive. But once the filing reaches the court, the signature still carries responsibility for the authorities cited inside it.
The problem has moved from warning to consequence
For several years, the public conversation treated AI errors in court filings as a lawyer problem: embarrassed firms, sanctioned attorneys, and fake cases exposed in federal litigation. By 2026, the pro se side is no longer marginal. An ArentFox Schiff discussion citing a New York Times analysis reported that pro se complaints bearing AI markers rose from near zero in 2019 to more than 18% of all pro se complaints by 2026.[2]
Those filings are not all defective. AI markers do not prove hallucinated law, and volume is not the same thing as misconduct. But volume changes what courts and opposing parties have to inspect. A self-help desk can help someone understand formatting, deadlines, and service. It cannot certify that every case in an AI-generated brief exists.
The Damien Charlotin AI Hallucination Cases Database, last updated July 3, 2026, tracked 1,696 decisions globally involving AI hallucinations in legal matters; 991 involved pro se litigants and 663 involved lawyers.[3] The database changes often, so those numbers should be read as a mid-2026 snapshot rather than a final count. Still, the direction is hard to miss. DISCO’s analysis of the database found that in 2025, pro se litigants accounted for 39% more hallucination incidents than licensed attorneys worldwide, with 304 incidents for pro se litigants compared with 219 for lawyers.[4]
The consequences are no longer limited to stern footnotes. Bloomberg Law, relying on Charlotin data, reported that at least 24 U.S. pro se litigants had received monetary sanctions since the second half of 2023, with more than half imposed since December 2025. Reported examples included a $10,000 Missouri fine and a California fee award exceeding $66,000, described as the largest recorded award against a pro se litigant specifically for AI-generated content.[5]
Those amounts are not paperwork inconveniences. For a person already priced out of counsel, a sanction can become the most lasting result of the case. A bad citation may also consume the court’s patience before the underlying claim is ever heard on the merits.
The comparison that matters is not brand against brand
Most pro se users do not begin with a procurement chart. They begin with a practical problem: they need a complaint, motion, opposition, declaration, or appeal brief, and they do not know how to structure it. Some use general chatbots such as ChatGPT or Gemini. Some use purpose-built platforms marketed to self-represented litigants, including tools such as Courtroom5, Cetient, or Prosei AI. Some combine both with search results, courthouse forms, and whatever free materials they can find.
That difference in tool type does matter, but not in the way marketing copy often implies. Stanford HAI benchmarking found that purpose-built legal AI systems hallucinated on 17% to 34% of legal queries, while general-purpose chatbots hallucinated on 58% to 82%.[6] Purpose-built legal AI performed materially better. It also still produced wrong legal information often enough that filing its output without checking would be reckless.
| Source of draft text | What it may help with | What remains unsafe without verification |
|---|---|---|
| General chatbot | Turning a messy story into a timeline, organizing a first draft, explaining common legal vocabulary | Case citations, quotations, jurisdiction-specific rules, procedural deadlines, claims about what a court has held |
| Purpose-built legal AI | More legally oriented drafting, issue spotting, form-like organization, more relevant legal language | Authorities that may still be nonexistent, outdated, misquoted, from the wrong jurisdiction, or used for a broader proposition than they support |
| Independently verified legal research | Confirming that a cited authority exists, is still good law, and says what the filing claims | Nothing is automatically safe; the filer still must match the authority to the facts, court, rule, and requested relief |
A 17% hallucination rate is not a small defect when the unit of error may be a case citation in a court filing. If a draft contains several authorities, the relevant question is not whether the tool is generally better than a public chatbot. The question is whether each authority in this filing has been checked outside the tool.
What courts are checking
Courts are not asking whether a self-represented party used a fashionable tool. They are asking whether the filing contains accurate legal authority. The basic checks are familiar: does the case exist, does the citation identify the real case, does the quoted language appear in the opinion, does the case stand for the proposition asserted, and does the authority apply in this court?
That standard is hard on pro se litigants, but it is not mysterious. If a motion says a federal appellate court adopted a rule, the judge should not have to determine whether the cited case was invented. If a brief quotes a sentence, the opposing party should be able to find the sentence. If an appeal relies on authority, the panel should not have to rebuild the research from scratch because an AI system supplied a confident falsehood.
The same-standard principle is now appearing in the pro se context. Commentary discussing Wilcox v. Gingrinch and the Seventh Circuit’s Jones ruling describes courts holding self-represented litigants to the same citation accuracy standards as attorneys when AI-generated authorities appear in filings.[8] That does not mean every procedural rule is applied identically to a lawyer and a nonlawyer in every setting. It does mean self-representation is a poor shield for fabricated law.
Verification is the real safety feature
A useful AI draft can still be an unsafe filing. The safety feature is not the polish of the prose or the legal vocabulary. It is the verification work that happens after the draft appears.
For a pro se litigant, independent verification starts with a simple rule: do not file a cited case, statute, rule, quotation, or legal proposition unless it has been confirmed somewhere other than the AI output. That outside source may be an official court opinion, a court rule page, a public legal database, a courthouse law library resource, or another reliable source the litigant can identify and revisit. The point is not to perform lawyerly theater. The point is to avoid asking the judge to rely on authority that may be imaginary.
- Confirm existence: search the case name and citation outside the AI tool, and make sure the court, date, and parties match.
- Confirm the quotation: compare any quoted sentence against the original opinion or rule text.
- Confirm the proposition: read enough surrounding text to know whether the case actually supports the statement in the filing.
- Confirm jurisdiction: check whether the authority binds the court, merely persuades it, or comes from the wrong court system.
- Confirm current status: look for signs that the case, rule, or statute has been reversed, amended, superseded, or limited.
That workflow is slower than pasting facts into a chatbot and signing the result. It is also the part that separates a draft from a filing. A more detailed human-in-the-loop process is available in the AI legal research workflow, but the basic discipline is the same for pro se use: the person signing the document must be able to show where each legal claim came from.
A hypothetical filing problem
Suppose a tenant uses an AI tool to draft an emergency motion. The draft cites a case for the rule that a court must hold a hearing within a short period after the motion is filed. The motion sounds urgent and organized. Before filing, the tenant searches the citation and finds no matching case. That is not a minor formatting problem. It means the tenant cannot rely on that authority unless they find a real source for the rule.
The safer next step is not to ask the AI to “fix the citation” and trust the replacement. The safer next step is to locate the actual rule, statute, local procedure, or case that governs emergency motions in that court. If no verified source supports the deadline, the deadline should not appear as a legal command in the filing.
Disclosure orders are court-specific, not a universal script
A second safety question is disclosure. Some courts now require lawyers, and in some settings parties, to disclose or certify the use of generative AI. Tracelaw’s tracker listed standing orders in courts including the Northern District of Texas, Eastern District of Pennsylvania, Northern District of Illinois, District of Massachusetts, the Fifth Circuit, Florida state courts, and others as of March 2026.[7]
That does not support a blanket statement that every pro se litigant in every court must file an AI disclosure. The scope of these orders varies, and many are written with attorneys in mind. A self-represented party still has to check the judge’s standing orders, the court’s local rules, and any case-specific order. If an order applies to “parties,” “filers,” or all submissions, a pro se litigant should not assume they are outside it simply because they do not have a bar number.
Disclosure also does not cure false authority. Telling the court that AI helped draft a motion is not a license to include nonexistent cases. The court may care both that the use was disclosed when required and that the filing itself is accurate.
Where AI can still help a self-represented litigant
None of this means pro se litigants should be told to avoid AI entirely. That advice is unlikely to be followed, and it misses the genuine benefit. A person with a pile of emails, notices, medical bills, photographs, and dates may need help turning chaos into a chronology. AI can help identify missing facts, separate background from requested relief, translate legal jargon into ordinary language, and make a draft easier for a human reviewer to read.
The danger begins when the tool moves from organization into authority. A chronology can be checked against the litigant’s own documents. A legal citation requires outside legal research. A summary of what happened is different from a claim that a particular court held a particular rule on a particular date.
This is why the rise of free or low-cost AI lawyer apps has changed the pressure on federal litigation without changing the filing party’s obligations. The broader pro se surge is discussed in more detail in the AI-fueled pro se surge, but the courthouse-level problem is immediate: more filings arrive with legal-sounding language, and someone must determine whether the law inside them is real.
A practical risk ranking for 2026
The safest comparison is not between product names. It is between levels of verification.
| Approach | Risk level | Why |
|---|---|---|
| Filing a general chatbot draft without checking citations | Very high | General chatbots have shown far higher hallucination rates on legal queries, and the filer has no independent basis for the authorities used. |
| Filing a purpose-built legal AI draft without checking citations | High | Legal-specific tools may perform better, but benchmarked hallucination rates remain too high for unverified court filings. |
| Using AI for structure, then independently verifying every authority | Lower | The tool assists with drafting, while the filer confirms the law before signing. |
| Using verified court forms, rules, and confirmed authorities with limited AI drafting assistance | Lower still | The filing relies on sources the court can inspect, not on the AI system’s assertion that the law exists. |
Even the lower-risk approaches are not legal advice and do not guarantee a good filing. A verified citation can still be irrelevant. A real rule can still be misunderstood. A deadline can still be missed. But the first and most avoidable failure is filing fake or misrepresented authority.
What a filer should be able to answer before signing
Before a pro se litigant signs an AI-assisted filing, they should be able to answer a few plain questions without relying on the AI chat window as the source.
- For every case citation: where did I confirm that this case exists?
- For every quotation: did I compare the quote against the original source?
- For every legal rule: is this rule from a binding source, a persuasive source, or only a general explanation?
- For every court-specific statement: did I check the local rules, judge’s orders, and current docket?
- For AI use itself: does this court or judge require a disclosure or certification?
If the answer to one of those questions is “the AI said so,” the filing is not ready. That is not hostility to technology. It is the minimum documentary discipline courts are now enforcing with real consequences.
In 2026, legal AI can help a pro se litigant draft, sort facts, and understand the shape of a filing. It cannot absorb the responsibility that comes with signing and submitting that filing. The person whose name is on the document remains responsible for every cited authority, every quoted proposition, and every disclosure obligation that applies in that court.
References
- US appeals court warns self-represented litigants over AI errors, Reuters, January 21, 2026
- AI-Powered Pro Se Litigation Flooding Federal Courts: What Consumer Products Companies Need to Know, ArentFox Schiff
- AI Hallucination Cases Database, Damien Charlotin, July 3, 2026
- AI Hallucinations in Legal Decisions: Trends, DISCO
- Big Law Grapples With AI-Fueled Pro Se Surge, Rising Legal Costs, Bloomberg Law
- AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries, Stanford HAI
- Court AI Disclosure Orders, Tracelaw, March 2026
- AI and Pro Se Litigation: How LLMs Are Reshaping Legal Strategy, Stites & Harbison
Comments
Join the discussion with an anonymous comment.