Since 2022, the share of self-represented cases in U.S. federal courts has climbed from a stable 11 percent to 16.8 percent. That 5.8 percentage point jump represents roughly 53 percent more pro se litigants than the historical baseline—and according to a new pre-print research paper, large language models are a significant driver.
The paper, "Access to Justice in the Age of AI: Evidence from U.S. Federal Courts" by Anand Shah and Joshua Levy, analyzed more than 4.5 million non-prisoner civil court cases and 46 million PACER docket entries spanning 2005 to 2026. The data shows that pro se filings held steady at 11 percent for nearly two decades, then began climbing sharply in 2022—coinciding with the public release of capable LLMs like ChatGPT and intensifying after GPT-4 launched in March 2023.
The researchers' core argument is not that AI improves access to justice. It is that AI makes legal self-help possible while making court administration harder. Before LLMs, filing a federal civil complaint required identifying the correct jurisdictional basis, pleading facts sufficient to survive a motion to dismiss, and navigating procedural requirements that varied by case type. These barriers kept the pro se rate stable. Large language models have eliminated that friction. Any person with an internet connection can now generate passable legal documents, draft complaints, and receive case-specific guidance without a law degree—at de minimis cost.
But the paper found that these AI-assisted cases are "heavier": each one generates more motions and demands more work from judges and court staff. The volume is rising alongside the complexity. This is not access to justice expanding; it is a new category of case materializing—one that clogs dockets without reducing adjudication burden.
The paper is careful to note it establishes correlation, not individual causation. "We do not claim to identify a causal effect of GPT-4 on pro se filing," the authors write, "only that the observed time series is difficult to rationalize without generative AI playing a role." That caveat matters. But the timing is difficult to dismiss.
Courts were already strained. Legal aid organizations have long argued that civil litigants without counsel face structural disadvantages—听不懂程序, cannot respond effectively to motions, and often generate work for judges who must essentially guide both sides. AI tools do not fix this imbalance. They shift it. The person who could not afford a lawyer now has a machine that can produce a well-formatted complaint. The judge still must adjudicate it.
The researchers' warning is stark: "If generative AI dramatically lowers the cost of self-represented litigation, the resulting surge in filings could overwhelm a system that depends on human judgment at every stage of adjudication." The question is not whether AI will reshape legal self-help—it already has. The question is whether courts can adapt to AI-generated filings at scale when the underlying procedures have not changed. The answer, at least for now, is no.