Self-represented litigants in America’s federal courts are filing more cases than at any time in recent memory, and a new working paper from MIT and USC points directly at ChatGPT as the accelerant. The paper documents a sharp break from a long-stable baseline of non-prisoner pro se civil filings—a trend that hardly budged for years until late 2022, when OpenAI’s conversational AI became broadly available on Windows desktops, browsers, and mobile devices. The researchers connect the surge to the ease with which people can now generate legal-sounding documents, but they also uncover a darker side: a flood of hallucinated citations that threaten to clog the judicial system.

The Numbers Behind the Surge

The MIT-USC study, a preprint released through the Social Science Research Network, analyzes federal district court dockets from 2010 through early 2025. For years, non-prisoner pro se civil case initiations hovered between 15,000 and 18,000 annually, with no statistically significant upward or downward drift. That pattern shattered in the fourth quarter of 2022. By mid-2023, monthly pro se filings had jumped roughly 30 percent compared with the pre-ChatGPT average, and the elevated rate has persisted through the first quarter of 2025, the last month with complete data.

Gains were not uniform. The sharpest increases appeared in districts with higher-than-average broadband adoption and in cases involving employment discrimination, civil rights, and contract disputes—the kinds of matters where a layperson might feel emboldened to try self-help with an AI assistant. The paper uses a difference-in-differences framework, comparing growth in pro se filings to changes in represented filings for the same claim types, to isolate the ChatGPT effect from pandemic backlogs or other court-scheduling anomalies.

When AI Dreams Up Case Law

Even more troubling than the volume spike is what the filings contain. The researchers manually reviewed a random sample of 1,200 pro se complaints filed after November 2022 and flagged whether any cited authority—statute, regulation, or precedent—appeared to be fabricated. They then verified each suspect citation against Westlaw, Lexis, and official court databases. The result: nearly one in six post-ChatGPT pro se complaints included at least one hallucinated case name or statute, a problem that was virtually nonexistent in the pre-ChatGPT control sample.

Hallucinated citations are not the typical layperson’s mistake of misremembering a case name. ChatGPT’s large language model is prone to assembling plausible-sounding legal citations from fragments of real cases, combining a real party name with a fictitious volume number, or inventing an entire appellate history. The paper reproduces several examples: a plaintiff citing “Thompson v. U.S. Dept. of Housing, 12 F.4th 407 (9th Cir. 2023)”—a decision that never existed—or referencing a nonexistent section of the Federal Rules of Civil Procedure that seemed to perfectly support their argument. These fictions are particularly dangerous because they sound authoritative to a non-lawyer and, if not carefully scrutinized, can slip past busy judges or law clerks.

For the Windows enthusiast, this story hits close to home. ChatGPT is not some exotic technology; it’s a browser bookmark on millions of Windows 11 and Windows 10 machines. The Microsoft Edge sidebar even offers one-click access to Bing Chat (now Copilot), which rests on the same underlying model as ChatGPT. Windows users can also install the official ChatGPT app from the Microsoft Store, pin it to the taskbar, and summon it with a keyboard shortcut. This frictionless availability means that someone facing a landlord-tenant dispute or a workplace grievance can, in a matter of seconds, generate a complaint that looks and reads like a lawyer’s work product—without ever consulting an actual attorney.

The MIT-USC paper does not call out Windows specifically, but its data on broadband adoption and computing-device ownership aligns tightly with the demographic profile of Windows’ installed base. More than 1.4 billion monthly active devices run Windows, many in American households with incomes where hiring a lawyer is a stretch. As AI assistants become even more deeply integrated into the operating system—Copilot is now a dedicated key on new laptops—the barrier to generating legal documents will drop further.

Courts Scramble to Respond

Federal judges have already taken notice. In 2023, a Manhattan federal judge sanctioned two lawyers for submitting a brief written with ChatGPT that cited six non-existent cases. That incident, though involving attorneys, sounded an alarm that reached the Judicial Conference of the United States. By early 2024, several district courts had issued standing orders requiring pro se litigants to disclose whether they used AI to draft their filings and to verify all citations manually. The Eastern District of Texas, a patent-heavy docket, now appends a “Notice to Pro Se Litigants Regarding Generative AI” to every case-opening packet, explicitly warning that reliance on AI does not excuse a failure to cite valid authority.

Yet education and disclosure rules can only do so much. The MIT-USC paper finds that in districts with a formal AI-disclosure order, the rate of hallucinated citations actually rose by two percentage points—possibly because litigants who took the time to read the order were already more cautious, but the core problem of unchecked AI output remained. Some judges have responded by assigning additional law-clerk time to verify every citation in pro se filings, effectively taxing the public fisc to clean up after an AI.

The Bigger Picture: AI Literacy and the Access-to-Justice Gap

Behind the data lies a painful paradox. Pro se filers often turn to the courts because they cannot afford representation; AI promises to bridge the access-to-justice gap by giving them a tool that demystifies legal language and procedure. In theory, that should be a good thing. But the MIT-USC analysis suggests that for every litigant who uses ChatGPT to articulate a meritorious claim more clearly, another is derailed by phantom law that scuttles an otherwise viable case—or worse, prompts sanctions that chill future self-advocacy.

The paper’s authors do not argue for banning AI from the courthouse. Instead, they call for technical interventions that the Windows ecosystem is uniquely positioned to deliver. Imagine a Copilot that, when asked to draft a legal complaint, automatically cross-references its output against a vetted database of primary law and flags any assertion that cannot be sourced. Microsoft has already experimented with grounding Copilot in specific data repositories for enterprise customers; extending that capability to consumer-facing legal aid could reduce hallucinations dramatically. Open-source models running locally on Windows, such as those from Hugging Face, could also be fine-tuned on clean legal corpora and distributed through curated Store packages.

Community Voices from the Windows Forum

In a discussion thread on WindowsForum.com, members have begun trading tips—and warnings—about using ChatGPT for legal tasks. One user recounted drafting a successful small-claims complaint with the AI’s help but admitted they had a lawyer friend review it before filing. Another shared a story of a relative who almost submitted a federal employment-discrimination charge that cited a made-up HR guideline, only to be saved by a court clerk’s skeptical phone call. “I told her, you wouldn’t run a system command you found on the internet without checking it,” the forum member wrote. “Why would you file a court document the same way?”

These anecdotes underscore a broader AI-literacy lesson for Windows users. The operating system’s power-user culture, long accustomed to verifying scripts and commands before executing them, can model a healthy skepticism toward AI output. Microsoft’s own safety documentation for Copilot stresses that users should “always review generated content for accuracy.” Applying that same principle to legal, medical, or financial advice obtained through AI could prevent a world of trouble.

What Comes Next

The MIT-USC working paper has yet to undergo peer review, but its early findings are already feeding into policy discussions. The Administrative Office of the U.S. Courts is examining whether changes to the federal rules of civil procedure might be needed to address AI-generated filings. Meanwhile, the American Bar Association’s Task Force on Law and Artificial Intelligence has released draft guidelines urging lawyers to maintain “competence” in AI tools—a standard that, if applied to pro se litigants, would be impossible to meet without widespread public education.

For Windows enthusiasts watching this story unfold, the takeaway is clear. The same machine learning breakthroughs that power next-generation features in Windows 11—Copilot, real-time captions, smart suggestions—are spilling into domains where the stakes are far higher than a misformatted email. The courts are only the first institutional stress test. Other government agencies, from the IRS to the Social Security Administration, are seeing AI-generated correspondence, and the pattern will repeat wherever complex rules meet a public eager to save on professional fees.

The paper closes with a recommendation that should resonate with anyone who has ever debugged a system crash: build guardrails that catch errors before they cause damage. Whether those guardrails are technical, educational, or regulatory, they will need to evolve as quickly as the AI models themselves. In the meantime, the best advice for a Windows user tempted to let ChatGPT write a federal complaint? Trust, but verify—and maybe keep a real lawyer on speed dial.