How AI Is Flooding Federal Courts: The Pro Se Filing Surge and Legal Automation

• Executive summary
• Self‑represented (pro se) federal filings rose from about 11% to 16.8% in FY2025—roughly 41,490 cases—after browser‑accessible large language models (LLMs) became mainstream.
• Automated screening (Pangram) flags for AI‑generated text climbed from ~1% in 2023 to ~18% in early 2026—about one in five sampled complaints contains AI text.
• The surge is concentrated in template‑friendly matters (civil rights, consumer credit, foreclosures) and has multiplied docket activity per case, producing heavy operational burdens for courts and agencies.
• Leaders must combine triage, verification, targeted capacity, and measured rule‑making to preserve access while protecting court capacity.

A single laptop, a Copilot prompt, and a crowded docket

Picture a litigant at a kitchen table opening Microsoft Copilot or ChatGPT, asking for a writ of mandamus template, tweaking a few facts, and hitting file in a friendly district. That modest workflow — free or inexpensive, fast, and good enough for routine pleadings — is what researchers point to when they say AI is reshaping who sues and how cases begin.

“Pro se” (self‑represented) filings in federal courts held near 11% for about two decades, then jumped to 16.8% in FY2025—about 41,490 filings—according to a large analysis of 4.5 million civil lawsuits and 46 million PACER (federal docket) entries. The rise isn’t uniform: it gathers in formulaic case types where an LLM can produce a procedurally adequate complaint quickly. The practical effect: more filings plus more follow‑up activity per filing, which strains fixed judicial staffing and agencies already stretched thin.

The data: a sudden shift

Key empirical highlights:

• Researchers analyzed 4.5 million civil lawsuits (FY2005–2026) and 46 million PACER entries.
• Self‑represented filings accounted for about 59% of civil lawsuit growth over the study period.
• Pangram (an AI‑detection tool) flagged AI‑origin text in sampled federal complaints at rates of ~1.0% (2023), 3.5% (2024), 10.5% (2025), and 18.0% (early 2026).
• Docket activity in the first 180 days of pro se cases is now about 158% higher than pre‑AI averages; attorney‑represented cases show ~23% more entries per case.
• Geographic spread is broad: 44 of 50 states saw increases; Vermont’s district went from ~45 pro se filings a year to over 1,100 in FY2024, largely mandamus petitions against USCIS (a writ of mandamus is an extraordinary court order compelling a government official to act).

How LLMs lower the marginal cost of filing

Large language models provide ready‑made text, procedural language, and citation templates. For straight‑forward claims—consumer credit disputes, certain civil rights claims, foreclosures—an LLM plus a quick human polish yields a filing that clears the initial procedural bar. That lowers the marginal cost of filing from something only a retained attorney would typically do to something any motivated person with internet access can attempt.

Cheap drafting also encourages forum shopping and repeat filing. Public guides and gig services have emerged: Reddit threads recommend Microsoft Copilot, and some users pay around $150 for a cursory lawyer review on marketplaces like Fiverr. The result: strategic filings in fast or permissive districts, multiplying administrative work for clerks and defendants.

Operational consequences: more work per case

More complaints is just the start. Courts report higher volumes of follow‑ups: motions, status updates, service issues, and clerical corrections. The study found that pro se cases generate far more docket entries early on—158% above pre‑AI averages over the first 180 days—while attorney files also saw a 23% increase in entries per case. Every extra entry requires human review: docketing, calendaring, and occasional judicial attention.

Back‑of‑envelope scale: if an average docket entry consumes 15–30 minutes of staff time (clerk review, indexing, and basic processing), tens of thousands of additional filings translate to hundreds of thousands of additional staff hours annually. That’s a labor uplift measured in the low millions of dollars across the federal system, absorbed without corresponding staffing increases.

Quality control and the hallucination problem

LLMs can hallucinate—fabricate facts, invent cases, or misattribute authority. A public database lists 129 documented AI‑fabrication incidents in court filings across multiple countries, and even a major model once generated a made‑up source in litigation. That risk creates downstream costs: opposing counsel or courts must flag, correct, and sometimes sanction fabrication, wasting scarce adjudicative time.

“One in five complaints now contains AI‑generated text.”

Judge Patrick J. Schiltz ordered future filings from a repeat filer to be “shredded without any additional notice,” calling the trend “an existential threat to the federal courts.”

But the picture isn’t only bleak. Judge Michael Y. Scudder emphasized that AI offers “great promise for enhancing access to justice for those without the resources to retain counsel or to represent themselves effectively.” Legal aid voices note the core problem: why do litigants feel their only path is a DIY filing powered by AI rather than supported, affordable counsel?

Detection tools and their limits

Pangram and similar detectors provide probabilistic signals, not certainties. Detection tools have false positives and false negatives, and models evolve. Relying on detection alone risks misclassification. Proper use is as a triage signal that triggers human review, not as a final adjudication about authorship.

Methodological caveats: the dataset focuses on federal filings (where fees are ≈ $405), so state courts—often with lower fees and different rules—may experience even larger shifts. Detection rates capture a snapshot; they’ll change as model behavior and detector techniques evolve.

Policy options: tradeoffs and practical pilots

Policymakers face a narrow set of real choices, each with tradeoffs between access and capacity:

• Do nothing and staff up: Accept higher workload and budget more clerks and magistrate time. Pros: preserves access. Cons: expensive, reactive.
• Raise procedural friction (fees or pre‑filing reviews): Could deter low‑value filings but risks blocking legitimate, low‑income litigants.
• Triage lanes and magistrate streams: Route formulaic matters to fast tracks or magistrate judges for streamlined processing. Pros: matches resource intensity to complexity. Cons: requires local rule‑making and staff redesign.
• Allow limited judicial AI use under strict safeguards: Judges and clerks could use vetted AI tools with audit trails, human verification, and disclosure rules to regain parity. Pros: efficiency. Cons: legal and ethical safeguards needed to prevent hallucination in opinions.
• Pilot a lower‑tier AI‑assisted adjudication stream: Process high‑volume, low‑complexity disputes with integrated automation and human oversight. Pros: scalable. Cons: significant legal, political, and fairness hurdles.

Practical checklist: what leaders should do now

For court administrators

• Implement triage lanes for template‑friendly claims and pilot magistrate‑led streams.
• Deploy AI‑detection tools as a flagging mechanism with audit trails and human review.
• Track KPIs: entries per case, time to disposition, detection precision/recall, fabricated citation rate, and litigant outcomes.
• Budget temporary staffing for high‑impact districts and invest in clerk training for AI signals handling.

For judges

• Require attestations about AI use in filings where appropriate and consider local rules for repeat offenders.
• Allow limited, documented AI assistance for judicial research under verification protocols.
• Favor proportional sanctions and swift triage rather than reflexive exclusion of pro se filings.

For government agencies (defendants)

• Prepare surge plans for litigation intake, and centralize responses for repetitive mass filings.
• Coordinate with court administrators to flag abusive patterns and reduce redundant processes.

For legal aid organizations and policymakers

• Expand non‑AI access channels (hotlines, guided intake, limited‑scope counsel) to reduce reliance on DIY AI workflows.
• Subsidize counsel for high‑volume, low‑complexity matters to prevent unfair triage outcomes.

For law firms and educators

• Publish and enforce AI use policies, emphasize verification of citations, and train staff to catch hallucinations.
• Teach students responsible AI practices: research vs. drafting, and mandatory fact‑checking routines.

Priority pilot metrics

• Entries per case (first 180 days) — baseline and post‑intervention.
• Time to disposition for triaged vs. standard tracks.
• Fabricated citation incidents detected per 1,000 filings.
• Detection tool precision and recall, audited quarterly.
• User satisfaction for litigants using AI‑assisted intake vs. traditional intake.

Questions leaders are asking now

Will AI continue to drive more pro se filings?

Yes. LLMs lower the marginal cost of drafting routine pleadings, so unless courts redesign triage, reduce friction in other ways, or expand affordable legal help, volume is likely to grow—particularly in template‑friendly case types.

Are detectors like Pangram reliable enough for courts?

Detectors are useful triage tools but imperfect. They provide probabilistic signals that require human audit; courts should avoid treating detector flags as definitive proof of authorship.

Should judges be allowed to use AI?

Limited judicial use can help courts keep pace, but it must be accompanied by transparency, human verification, and documentation to prevent hallucination from seeping into opinions or orders.

Is a lower‑tier “AI‑handled” court feasible?

Conceptually promising for high‑volume, low‑complexity matters, but legally and politically challenging. Any pilot must ensure due process, equitable access, and robust human oversight.

Final tradeoff

LLMs have opened a door to wider access to legal drafting tools—an advance with real democratic value. But when access arrives as a flood of templated filings without matching process redesign or verification, courts face a paperwork tsunami. Leaders who balance measured triage, targeted capacity, transparent AI handling, and expanded non‑AI access channels can preserve both access to justice and the operational health of the judiciary. That balance is the immediate business and governance problem AI has handed to courts—and it’s fixable with deliberate policy and a few carefully run pilots.

“The real problem is: how come these people don’t have another way, other than using AI.” — Sateesh Nori