Federal courts are seeing a new kind of AI pressure
A new study from researchers at MIT and the University of Southern California suggests that generative AI is reshaping access to the U.S. legal system in a direct and measurable way. The researchers found that lawsuits filed without a lawyer in U.S. federal civil courts rose sharply after ChatGPT became widely available, with one in five complaints now containing AI-generated text. What looks at first like a democratizing effect for people who cannot afford counsel is also becoming an operational strain for judges and court staff.
The study analyzed 4.5 million civil lawsuits spanning fiscal years 2005 through 2026, along with 46 million entries from PACER, the federal electronic case registry. Its central measure was the pro se rate, or the share of cases filed by people representing themselves. For roughly two decades that rate held steady at about 11 percent of federal civil cases. In fiscal year 2025, it climbed to 16.8 percent, with 41,490 pro se filings that year alone, nearly double the pre-AI average described in the study.
That jump matters because federal court is not the easiest venue for self-represented litigants. Filing fees are high, formal pleading standards are strict, and most U.S. civil litigation happens in state and local courts instead. The researchers argue that if federal filings are showing this level of AI-related growth, the effect in lower courts could be even larger. In other words, the federal numbers may be the visible leading edge of a broader shift.
Where AI is helping, and where it is not
The pattern is not uniform across all types of cases. The growth clusters in areas where a complaint can often be organized around familiar templates and procedural forms, including civil rights claims, consumer credit disputes, and foreclosure-related filings. By contrast, fields that depend on deep specialist knowledge, such as patent law or securities law, show no comparable effect in the study.
That distinction is one of the study’s most important findings. Large language models appear to be lowering one specific barrier: the drafting of procedurally viable documents. They are not flattening the full complexity of the legal system. In simpler or more standardized disputes, AI can help a non-lawyer produce something that looks enough like a proper complaint to get into court. In highly technical litigation, that advantage appears limited.
The increase is also concentrated on the plaintiff side and appears across 44 of 50 states at once, a distribution the researchers treat as evidence against local explanations. This broad geographic spread strengthens the case that the shift is connected to a national technological change rather than a few regional legal trends.
The workload problem may be larger than the filing problem
The study does not describe a collapse in court outcomes. Case durations and outcome distributions are said to be largely unchanged. But that does not mean the system is unaffected. The more immediate stress shows up in the volume of activity within each case.
According to the researchers, the number of docket entries per court from pro se plaintiffs in the first 180 days of litigation stood 158 percent above the pre-AI average as of the second quarter of 2025. Each one of those entries requires attention. Motions, responses, orders, procedural notices, and corrections all consume staff time. Even when a filing does not materially advance a case, someone still has to process it.
The study also reports that represented cases are generating more entries too, rising 23 percent per case, which suggests lawyers and firms may be using large language models as well. That detail is easy to miss, but it points to a broader administrative change. AI may not just be bringing new users into the system. It may also be accelerating the pace at which all kinds of litigants produce paperwork.
Access to justice and administrative overload are arriving together
The idea that AI could narrow the justice gap has intuitive appeal. Legal services are expensive, and many people with legitimate grievances never reach a courtroom because the procedural burden is too high. If AI tools can help someone prepare a complaint, organize facts, and navigate formality, that can look like a win for access.
But the study suggests that access and overload are arriving together. Federal judges are reportedly resorting to stronger measures to manage the flood of filings. Even if many complaints are sincere, the system still has to absorb a greater volume of documents that may be formulaic, repetitive, or only superficially compliant with court rules. A court system designed around human-limited drafting capacity is now facing software-assisted abundance.
That creates a policy dilemma. Restricting AI-generated filings too aggressively could shut out people who genuinely need help. Allowing the trend to expand without adjustment could bog down already burdened courts. The challenge is not simply whether AI belongs in legal drafting. It is whether institutions can distinguish between assistance that broadens meaningful access and output that multiplies administrative noise.
What this shift could mean next
The most immediate implication is operational. Courts may need new screening procedures, clearer disclosure expectations, or stronger formatting and verification rules for self-filed complaints. They may also need to invest in staff capacity and digital workflows built for higher filing volume.
The deeper implication is cultural. Generative AI is changing who feels capable of initiating formal legal action. A complaint that once required money, expertise, or both can now be drafted in minutes with a prompt. That does not guarantee a strong case, but it changes the threshold for participation.
The study frames that threshold change as historically significant. For years, the federal pro se rate remained strikingly stable. The post-ChatGPT jump breaks that long pattern. Whether that becomes a durable expansion of legal access or an enduring paperwork crisis may depend less on the models themselves than on how quickly the courts adapt to the new volume of machine-assisted advocacy.
This article is based on reporting by The Decoder. Read the original article.
Originally published on the-decoder.com
