The Courts Don’t Care If You Used AI. They Care If It’s Real.

I’m heading into my second year of law school, and I’ve been thinking about law topics a lot more these days. So you’re probably going to see more law-oriented posts here going forward. This one’s a legal op-ed about two federal cases from this January, both about people representing themselves in court who used AI to write their filings. One walked away clean. One lost his case over it. I dug into both, built a free tool along the way, and I’m writing about it because in a few years I’m going to be the lawyer standing in that exact gap, and right now nobody’s closing it for the people who can’t afford one. Stick with it even if it gets wonky.

Here is what gave Solomon Jones away: his brief was too good.

Jones is a pro se plaintiff (representing himself, without a lawyer) who sued Kankakee County, Illinois entities under 42 U.S.C. § 1983 over a string of 2023 arrests for trespassing and disorderly conduct. His appeal reached the Seventh Circuit on procedural grounds, a denied recusal motion (a request that the judge step aside from the case) and a stay under Younger abstention (a doctrine letting a federal court pause a case so a related state court proceeding can finish first), not on the merits of his underlying claims. The AI question surfaced almost as a footnote to that.

When the Seventh Circuit examined the appellate brief he filed for himself in January 2026, what drew its attention wasn’t sloppiness. It was polish. The brief was “meticulous in its presentation—very cleanly and professionally formatted, employing prose and citation formats we rarely see from pro se litigants.” It also attributed quotations to two cases that didn’t contain them. Jones denied using AI. The court didn’t believe him, and said so, while noting that it usually sees inaccurate filings from self-represented litigants and simply “move[s] past the misstatements.” The signal wasn’t error. It was competence beyond what someone in his position should be able to produce.

That leaves a self-represented person using AI in a bind with no exit. Use it badly, and hallucinated citations get you sanctioned. Use it well enough to close the presentation gap that has always marked pro se filings, and the quality itself becomes the evidence against you.

Nine days later

On January 30, the Eastern District of Missouri dismissed Theoda Mills’s case with prejudice (permanently, meaning he can’t refile it) as a Rule 11 sanction (a penalty for filing dishonest or frivolous claims with the court). Mills, also pro se, had sued the City of St. Louis and two of its employees, alleging age and sex discrimination and retaliation. Same category of error as Jones: nonexistent cases and fabricated quotes. Mills said he’d used “electronic drafting tools that paraphrase legal principles” and had “mistakenly believed” the summaries were accurate.

It would be easy to call that a coin flip. It isn’t. Jones was a case of first impression. The Seventh Circuit said so, declining to sanction partly because “before today’s decision we have not supplied any guidance on the use of AI by pro se litigants.” Mills, by contrast, had already received a show cause order (a court order demanding he explain why he shouldn’t be sanctioned) cataloguing his misstatements, and then made a brand-new false claim about a different case in a filing that referenced the court’s “Rule 11(b) concerns” by name. He holds a doctorate, which the court weighed. He also never opposed the motion to dismiss, letting the deadline pass by forty-five days, an independent ground for dismissal.

Mills doesn’t ignore Jones. It cites it, quoting the instruction that “all litigants—represented and unrepresented—must read their filings and take reasonable care.” Within nine days, without anyone designing it, two courts assembled a rule that mostly works: notice, persistence, capacity. First offense with no guidance and no means to check, you get a warning. Warned, and you do it again, you’re done.

The question neither court thinks is interesting

Now read what both courts said about AI itself. The Seventh Circuit, after three paragraphs explaining why it believed Jones used a chatbot: “Whether we are right or Jones is right need not detain us.” The Missouri court, on the same question: whether Mills used AI is “largely beside the point,” because “the fact remains that he provided the Court with multiple filings that contained fictitious and misleading information.”

Two courts, reasoning independently, shrug at the tool and ask only whether the content is real, which should trouble the profession, since the tool is nearly its entire visible response. Courts have issued a wave of standing orders requiring AI disclosure; bar guidance follows the same contour. But a litigant who discloses AI use and files fake cases has satisfied that rule and violated Rule 11 anyway. One who uses AI, checks every citation, and files accurate work has done nothing wrong, and under some orders, still has to announce it. Disclosure regulates provenance. Rule 11 asks about accuracy. Only one of those is the actual concern.

The two courts do agree on something else: Rule 11 binds unrepresented parties, full stop. “There is no pro se exception,” the Missouri court wrote, because carving one out “would effectively place all unrepresented parties beyond the reach of Rule 11.” Correct, and also true that a person with no training and no database subscription cannot verify what a drafting tool hands them. What neither court can do from the bench is close that distance. Every order in this line says take reasonable care. Not one says with what. The Missouri court even considered ordering “participation in seminars or other educational programs” as a sanction, and rejected the idea as “neither adequate nor appropriate.” A judge reaching for something that teaches instead of punishes found the shelf bare.

Now read the order the way Mills would have to

The proposition that ended his case, “There is no pro se exception to Rule 11(b),” comes from an unpublished decision cited only as 2025 WL 3052638. The accord citation backing it isn’t in CourtListener, the largest free repository of American case law, at all. The line calling his conduct “fundamentally incompatible with our justice system” is quoted from a law review article behind a subscription. The sanctioning-discretion discussion leans on Wright & Miller, a treatise sold by the volume.

By my count, ten of the thirty-one authorities in the order are out of reach: five exist publicly only as Westlaw numbers, two are missing from CourtListener, one is a paywalled law review, one a paid treatise, one a PACER docket billed by the page. And six pincites, “at *1,” “at *3,” point to Westlaw’s internal pagination, which exists nowhere else. Even the free copy of Jones is cited here that way, because the official reporter pages hadn’t been assigned yet.

So the document explaining to Theoda Mills that he must verify his sources is, in substantial part, a document he cannot verify. This isn’t a criticism of the judge. He cited controlling authority in the format the profession requires, for the audience the citation system was built to serve. That’s exactly the problem. The infrastructure of legal reasoning presumes a paying subscriber on the other end. When the person on the other end isn’t one, “take reasonable care” arrives in a language he has no way to look up.

Whose job is this?

The Seventh Circuit has a view: responding “at the broader level of court rules and policies is a challenge fully upon the judiciary, federal and state.” The organized bar has a different one, expressed by omission. ABA Formal Opinion 512 (the American Bar Association’s official ethics guidance, issued in 2024, on how lawyers may use generative AI) is careful and thorough: competence, confidentiality, communication, candor, supervision, fees. Every obligation in it runs from a lawyer to a client, a court, or a firm. The population receiving no or inadequate help for ninety-two percent of its substantial civil legal problems doesn’t appear in it.

The market has already answered whether a retrieval-only tool can solve the hallucination problem: a tool that only points to opinions that actually exist, instead of generating new text, cannot invent one. The technology isn’t the obstacle. Making it free for the person who needs it most is.

“Many important initiatives are underway,” the Seventh Circuit notes. Legal aid technologists and self-help centers are doing real work, without the profession’s expertise, standards, or money behind them. That’s why a federal judge weighing sanctions couldn’t find one educational program to point a litigant toward. Courts claimed the rulemaking. The bar addressed its own members. The person standing between them gets a disclosure requirement and a warning.

What would actually help

Say the framework out loud. Jones and Mills each landed on it independently, with no court calling it a rule. Write it down, so the next court doesn’t have to reinvent it, and two people who behaved identically get the same outcome regardless of which courtroom they end up in. Aim rules at accuracy, not provenance. Rule 11(b) already asks whether an inquiry was “reasonable under the circumstances,” language that plainly contemplates different circumstances for different filers; that standard exists, it just isn’t being used. And pair every disclosure requirement with a resource to actually meet it.

That last one is cheap. While researching this piece, I built a free citation verifier: paste a citation, and it confirms the case exists and that a quoted passage actually appears in the opinion, checking a local database and then CourtListener’s API. No account, no fee. It took one law student and open data. A profession that produced Opinion 512 could do considerably better.

But that tool can only check what’s already free. It can’t get you a Westlaw-only unpublished opinion, a paywalled law review, or a treatise sold by the volume. Those need a real answer: the industry needs to build a public, reliable resource with the complete record — every opinion, published or not. The pagination problem specifically isn’t even hypothetical to fix: Ohio and about a dozen other states already cite their own courts by paragraph number instead of Westlaw’s page numbers, a vendor-neutral format that’s been in place since 2002. Federal courts, including the Seventh Circuit that decided Jones, could adopt the same thing tomorrow. They haven’t.

One more thing

An earlier draft of this piece said the Missouri court fined Mills $5,000. It didn’t. The court weighed a monetary sanction, noted Mills had asked for a non-monetary one, and dismissed the case instead. The false detail came from a roundup covering several cases at once; a fact from a different case attached itself to this one. It was plausible and wrong, and reading the eight-page opinion caught it.

That’s the argument in miniature: the failure isn’t a technology, it’s the gap between a claim and the source it’s drawn from, and that gap doesn’t care whether the claim came from a chatbot or a secondary source. Mills’s mistake and mine were the same mistake. He couldn’t check the source. I could.

Notice, too, what actually separated Jones from Mills. Not need. The Seventh Circuit found Jones lacked the resources to verify, and nothing suggests Mills had them either. What separated them was notice: one had been told, the other hadn’t yet. Ask yourself honestly which side of that line you’d have landed on, alone, with a filing deadline, and no lawyer to tell you what you were allowed to trust. That’s a defensible way to allocate blame. It’s a poor way to allocate justice, and it’s the only one available when the duty is universal and the means are priced by income.

The Model Rules say a lawyer is “a public citizen having special responsibility for the quality of justice,” obligated to “ensure equal access to our system of justice for all those who because of economic or social barriers cannot afford or secure adequate legal counsel.” Courts are now publishing opinions that describe exactly where that access fails and what would fix it. The profession has the expertise, the standing, and the stated obligation to build it. So far it has built the rules for itself and left the rest to the judges.

Theoda Mills lost his case over this in January. Somewhere right now, someone else is drafting a filing with the same tool, the same hope, and no one standing between them and the same mistake.

---

Sage Lewis is a student at the University of Akron School of Law and an advocate for homeless and low-income communities. This piece was drafted with the assistance of generative AI.