Legal Insights
You Know It When You See It: AI-Drafted Pro Se Pleadings in Federal Court
A law professor of mine once asked the class, half rhetorically, what pro se means. His answer: it means you lose. Cynical, but it stuck with me.
It came back to mind recently, though not in the way I expected. The complaint in front of me did not look thin or informal. It looked polished. Densely cited. Confident. Somewhere around paragraph three, it felt as though it had help.
What began as a personal injury claim had evolved into a multi-count action against our client, an insurance broker, complete with civil RICO allegations, mail and wire fraud predicates, and a demand for treble damages. The plaintiff was proceeding pro se. She also appeared, by every visible marker, to be working with an AI drafting tool.
The interesting part was not that she used one. It was how you could tell and what those clues reveal about what these tools are actually good at.
The Tells
Everything Was Catastrophic; Nothing Was Just an Injury
The complaint did not describe an eye irritation allegedly caused by lash-extension glue or even an eye injury more generally. Instead, it repeatedly referred to a “catastrophic bilateral orbital toxic exposure.” That phrase, or a close variation of it, appeared more than a dozen times across twenty-two pages and survived largely unchanged through three successive versions of the complaint.
AI-assisted drafting tends to reach for the most dramatic available phrasing and then reuse it almost verbatim. The result often reads less like a person describing an injury in her own words and more like a stylistic pattern generated by a machine.
The Citations Are Dense, Confident, and Mostly Beside the Point
The original complaint cited more than a dozen federal and state statutes, including the FTC Act, TSCA, OSHA, the FDCA, the Consumer Product Safety Act, ERISA, RICO, I think we ran out of the alphabet, and a state insurance producer statute. These authorities were arranged in neat tables that mapped each statute to a corresponding “duty” and “breach.”
Beneath that elaborate facade, however, the case remained what was essentially a negligence claim asserted against the wrong defendant.
The complaint also sought $3 billion in damages and requested that the court refer the matter to the Kansas Attorney General.
A recent working paper from the University of Miami Law & AI Lab, The New Pro Se: Generative AI and the Surge in Federal Civil Self-Representation by Cohen-Sasson, analyzed approximately 2.8 million federal filings. The paper found that AI-flagged pro se complaints are roughly 50% more citation-dense than pre-ChatGPT complaints yet are dismissed more frequently, not less frequently (61.1% versus 53.6%). More formal presentation, worse outcomes. The paper describes the phenomenon as a “litigation-efficacy paradox.”
It Didn’t Just Imitate a Brief. It Answered Ours, Draft by Draft.
This was the tell I found most interesting. The issue was not an AI tool generating generic responses to hypothetical defense arguments. Instead, there were three complaints filed over roughly six months, each arriving after we had moved to dismiss the previous version.
The original filing was the roughest. It included statutory tables, a multibillion-dollar damages demand, and lengthy digressions into unrelated grievances. After we filed our motions to dismiss, much of that disappeared in the next amendment. The list of defendants expanded to include more than a dozen affiliates of our client around the world, while the damages demand was reduced to $2 billion. By the third amended complaint, specific dollar figures had disappeared altogether.
More notably, the complaint had begun arguing with itself. Mid-paragraph hedges and caveats appeared in places where none had existed before, often tracking our motions to dismiss almost point for point. In other words, it appeared that the plaintiff was feeding our briefs into the AI system and instructing it to revise the allegations to circumvent the legal arguments we had raised.
The Geography Is Not a Coincidence
The Cohen-Sasson paper also broke its results down by federal circuit. According to the study, the Tenth Circuit had the highest rate of AI-flagged pro se complaints in the country at 33.3%, substantially higher than the Ninth Circuit (14.9%) or the Sixth Circuit (15.5%).
If you defend pro se litigation in the Tenth Circuit, you are statistically operating in the center of this trend, not at its margins.
What It Doesn’t Change
None of this changes how courts must treat a pleading while it is pending. Liberal construction still applies. Leave to amend still applies. Each iteration still deserves genuine engagement.
Courts perform their role by reading past the drafting style and focusing on what is actually alleged. They do that amendment after amendment, regardless of how the pleading was produced or how its language evolves over time. Style may tell you what tool helped create a document. It does not tell you whether the claim survives a motion to dismiss.
Recognizing the tells took about a paragraph. Winning required working through all ninety-nine pages, examining every allegation and every count, and asking whether a legally cognizable claim remained once the citations and adjectives were stripped away.
The Third Amended Complaint portrayed itself as a coordinated federal enterprise. Examined paragraph by paragraph, however, it remained a negligence case against the wrong defendant. Every count unraveled for the same reason, one allegation at a time.
The district court ultimately agreed and granted our motion to dismiss in favor of our insurance-broker client.
Sources: AO Table C-13 (pro se filings by district, FY2025); Cohen-Sasson, “The New Pro Se: Generative AI and the Surge in Federal Civil Self-Representation”(Univ. of Miami Law & AI Lab working paper, 2026); Gough & Taylor Poppe, “(Un)Changing Rates of Pro Se Litigation in Federal Court,”Law & Social Inquiry 45(3) (2020).
