In Access to Justice in the Age of AI: Evidence from U.S. Federal Courts, Anand V. Shah and Joshua Y. Levy, document what may be a tectonic shift in the civil justice system.  According to Shah and Levy, Generative AI is lowering the barriers to filing suit—and, as a result, more self-represented litigants are initiating claims.

Analyzing 46 million docket entries, the authors uncover a surge in non-prisoner pro se filings.  Federal pro se filings have risen in percentage terms—from a nearly 20-year steady state of roughly 11%, up to nearly 17% in fiscal year 2025.<a href=”https://legalpro.jotwell.com/opening-the-courthouse-door-or-just-lowering-the-threshold/#easy-footnote-bottom-1-2234″ title=”As Shah and Levy recognize, prior to the recent uptick, pro se litigation rates in federal courts stayed remarkably steady, at around 11%.  See Mark D. Gough &amp; Emily Taylor Poppe, (Un)Changing Rates of Pro Se Litigation in Federal Court, 45 L. &amp; Soc. Inquiry 567, 574 (2020).  The story has been different in state courts.  See Nora Freeman Engstrom &amp; David Freeman Engstrom, The Making of the A2J Crisis, 74 Stan. L. Rev. Online  146, 150-51 (2024) (explaining that, in recent years, pro se representation rates in state courts have risen sharply, although quantifying the exact rise is difficult).” target=”_blank”>1 And, they have also jumped on a numbers-basis.  Self-represented individuals initiated just 23,000 federal cases in 2022, and nearly twice that—41,000—in 2025.<a href=”https://legalpro.jotwell.com/opening-the-courthouse-door-or-just-lowering-the-threshold/#easy-footnote-bottom-2-2234″ title=”For these figures, the authors pull data from the Federal Judicial Center Integrated Database. Filings by represented plaintiffs also increased over this time, but pro se filings increased faster.” target=”_blank”>2

Meanwhile, deploying sophisticated AI-detection tools to analyze a sample of 1,600 complaints filed over an eight-year span, Shah and Levy also identify a growing share of complaints that appear to contain AI-generated text.

Put the two pieces together, and the authors paint a fairly positive portrait.  They suggest that, armed with new technology, the courthouse door, long closed to many self-represented litigants, may be creaking open.  In the authors’ words: “[F]or most people, self-representation is prohibitively hard. . . . The widespread, public diffusion of capable LLMs changes that calculus.” (P. 3.)

That certainly sounds like a striking—and welcome—development.

For years, scholars have lamented that the American civil justice gap is (as one of us has put it) “sizable and scandalous.”<a href=”https://legalpro.jotwell.com/opening-the-courthouse-door-or-just-lowering-the-threshold/#easy-footnote-bottom-3-2234″ title=”See Nora Freeman Engstrom, Legal Insurance and Its Limits, 124 Mich. L. Rev. 1, 6 (2025).” target=”_blank”>3 A big chunk of the gap is attributable to the fact that, whether owing to high cost or bewildering complexity, most people with valid legal entitlements—who have, perhaps, been sexually harassed by an employer, ripped off by a lender, or wrongfully denied coverage by an insurer—lump it.  They take no legal action to protect their interests.<a href=”https://legalpro.jotwell.com/opening-the-courthouse-door-or-just-lowering-the-threshold/#easy-footnote-bottom-4-2234″ title=”See Nora Freeman Engstrom, She Stood Up: The Life and Legacy of Deborah L. Rhode, 74 Stan. L. Rev. Online 1, 8 (2021).” target=”_blank”>4

Further, there’s a strong scholarly consensus that this inaction is problematic.  Individuals’ persistent failure to vindicate their legal rights mutes deterrence signals, deprives victims of compensation, and can consign individuals to ongoing illegality, insecurity, and exploitation.

If generative AI can help these lumpers identify valid legal theories and initiate filings, it may chip away at that long-standing access-to-justice deficit.  And, that, of course, is what Shah and Levy seem to be suggesting: that sophisticated and widely-available AI tools are changing the “calculus” to make it so self-representation is no longer “prohibitively hard.”  (P. 3.)

But here is the rub: While Shah and Levy show that there’s a surge in AI use, and while they also show that more pro se litigants are initiating claims, they do not show that these new pro se litigants are faring any better than they did in the analog era.  Nor do they shed light on the legitimacy of these new pro se filings.  Nor even—and critically—do they show that more AI use and more pro se litigation are two litigation trends that are actually connected, rather than operating in parallel. They do take an important step, however, in setting the stage for the careful analysis that must be ongoing as AI technology continues to evolve and the federal courts attempt to keep pace.

Take the first point first.  Yes, more pro se folks are initiating claims in federal court.  But, on the paper’s own terms, case outcomes look stubbornly familiar.  Dismissals remain the norm. Settlements are not obviously rising as a share of dispositions.  And victories for self-represented plaintiffs continue to be few and far between.5 In other words, more self-represented litigants appear to be getting into court—but it isn’t clear that they are prevailing, settling on better terms, or avoiding adverse outcomes.

Now, consider the case quality question:  Empowered by AI, are today’s self-represented litigants simply bringing marginal cases that (in 2022, say) any lawyer worth her salt would have screened out and refused to file?  Or are AI tools lowering barriers so that some poor and middle-class folks with valid legal entitlements can at least make it into court?  The answer to this question matters enormously.  In the first scenario, AI is—on balance—degrading the quality of claims and eroding the integrity of our courts.  In the second, the picture is arguably a bit rosier, and, as the tools themselves improve, it’s apt to become brighter still.

That brings us to the connection question.  Shah and Levy seem to assume that AI use is powering the rise in self-representation.  They state, for instance: “Generative AI has lowered the cost of producing passable legal documents.  Americans are noticing, and they are walking into federal civil court in much larger numbers to act on it.”  (P. 33.)

We agree that’s probably right.  But Shah and Levy stop short of proving it.

True, Shah and Levy document a rising share of complaints flagged as AI-generated—from just 1% in 2023 to 18% today.   Yet their dataset—drawn from RECAP—likely overrepresents cases involving lawyers and institutional actors, who are more likely to upload filings.<a href=”https://legalpro.jotwell.com/opening-the-courthouse-door-or-just-lowering-the-threshold/#easy-footnote-bottom-6-2234″ title=”RECAP is an archive attached to CourtListener in which users can upload case documents.  The authors note that pro se filings are less likely to end up in this sample since uploaders to RECAP are “disproportionately attorneys, journalists, and legal researchers.” (P. 29.).” target=”_blank”>6 As the authors acknowledge, this creates a selection problem—and it means that the evidence they assemble stops far short of establishing that pro se litigants are the primary (or even significant) drivers of AI uptake.

Thus, we’re left with several critical questions—the answers to which might chart the next chapter in civil justice.  Are AI tools helping self-represented litigants identify and pursue meritorious claims, or are they simply making it easier to file frivolous ones?  If the former, will the AI tools improve so that pro se litigants can, in short order, not only make it into court but also overcome procedural barriers to obtain meaningful relief?  Or, will tool development stall (or be matched, or surpassed, by defense-side innovations), such that AI-assisted pro se plaintiffs will be able to file claims, but not actually vindicate rights?  Are we witnessing a true expansion of self-representation, or more of the same, where the “haves” harness ever-better-technology—and the “have-nots” fall ever further behind?<a href=”https://legalpro.jotwell.com/opening-the-courthouse-door-or-just-lowering-the-threshold/#easy-footnote-bottom-7-2234″ title=”See David Freeman Engstrom &amp; Nora Freeman Engstrom, Legal Tech and the Litigation Playing Field, in Legal Tech and the Future of Civil Justice 133, 154 (David Freeman Engstrom ed., 2023).” target=”_blank”>7 And, the $64,000 question: How do these forces net out for courts and society?  These are the questions that legal scholars, empiricists, and court administrators should be collaborating to answer.

None of these open questions diminish the paper’s core contribution.  In their piece, Shah and Levy distill valuable data for a wide range of observers, who are starved for a picture of how courts are changing at scale.  They provide some of the first large-scale empirical evidence that (1) case filing figures are in flux, and (2) generative AI is reshaping participation in the federal courts.  They also sharpen the distinction that we cannot lose sight of: getting new plaintiffs past the courthouse door does not mean we are helping them achieve the goals of our civil justice system.  The civil justice gap is filled, not by filings, but by outcomes that make—or fail to make—legal rights real.

Download PDF
  1. As Shah and Levy recognize, prior to the recent uptick, pro se litigation rates in federal courts stayed remarkably steady, at around 11%.  See Mark D. Gough & Emily Taylor Poppe, (Un)Changing Rates of Pro Se Litigation in Federal Court, 45 L. & Soc. Inquiry 567, 574 (2020).  The story has been different in state courts.  See Nora Freeman Engstrom & David Freeman Engstrom, The Making of the A2J Crisis, 74 Stan. L. Rev. Online  146, 150-51 (2024) (explaining that, in recent years, pro se representation rates in state courts have risen sharply, although quantifying the exact rise is difficult).
  2. For these figures, the authors pull data from the Federal Judicial Center Integrated Database. Filings by represented plaintiffs also increased over this time, but pro se filings increased faster.
  3. See Nora Freeman Engstrom, Legal Insurance and Its Limits, 124 Mich. L. Rev. 1, 6 (2025).
  4. See Nora Freeman Engstrom, She Stood Up: The Life and Legacy of Deborah L. Rhode, 74 Stan. L. Rev. Online 1, 8 (2021).
  5. All told: “Judicial dismissals account for approximately 60% of resolved pro se cases in the pre-AI period and 63% in the post-AI period.  Settlements account for roughly 20% pre-AI and 23% post-AI.  Pro se losses fall modestly from 19% to 13%, and pro se wins remain rare throughout (under 1%).” (P. 26.) Of course, stable outcome proportions with increased filings could mean that some of the new filers are surviving dismissals. It is also possible that litigants who would have filed without AI are marginally improving the quality of their claims (potentially with the new tools), and all the new filers are having their cases dismissed. Without an evaluation of the marginal filers, we do not know how much justice they are truly achieving.
  6. RECAP is an archive attached to CourtListener in which users can upload case documents.  The authors note that pro se filings are less likely to end up in this sample since uploaders to RECAP are “disproportionately attorneys, journalists, and legal researchers.” (P. 29.).
  7. See David Freeman Engstrom & Nora Freeman Engstrom, Legal Tech and the Litigation Playing Field, in Legal Tech and the Future of Civil Justice 133, 154 (David Freeman Engstrom ed., 2023).