Wider public access to generative AI coincided with a sharp rise in self-represented federal civil plaintiffs and an estimated 14% of non-form post-GenAI complaints showing AI-consistent drafting; despite denser citations, these complaints are more likely to be dismissed and to exit earlier in the process.

Main Finding

Public diffusion of generative AI (GenAI) after November 2022 is associated with a large, persistent increase in federal civil self-representation and with a detectable subset of pro se complaints that carry text signatures consistent with AI-assisted drafting. The pro se plaintiff rate rose from 11.33% pre-GenAI to 16.94% post-GenAI (+5.61 percentage points, ≈49.5% relative increase). An interpretable, pre‑AI–calibrated AI-consistent drafting score flags a net 13.9% of post-GenAI non-form pro se complaints; those AI-flagged complaints are more citation-dense, disproportionately associated with observed first-time filers, geographically concentrated, and they perform worse on several outcome measures (higher dismissal rates, earlier termination) with no win-rate advantage.

Key Points

• Scope and findings at a glance
  • Dataset: ≈2.8 million federal civil filings (FJC) FY2008–FY2025; complaint-text sample: 12,842 pro se complaints (CourtListener), focused on Civil Rights and Other Statutory categories.
  • All-civil pro se rate: 11.33% pre-GenAI → 16.94% post-GenAI (+5.61 pp; χ² significant).
  • Post-AI pattern appears as an accelerating trend rather than a one-time jump (interrupted time-series: post-AI slope change ≈ +0.729 pp per quarter, p<0.001).
  • Largest category-level increases: Other Statutory (+8.28 pp) and Civil Rights (+7.44 pp).
• AI-consistent drafting measure and prevalence
  • Measure: an interpretable proxy based on three stylometric/textual indicators (pre-AI-calibrated threshold); paper reports the net flagged share as 13.9% of post-GenAI non-form complaints.
  • Distinction: non-form complaints (free prose) vs. form complaints (standard templates). Signal is stronger and more detectable in non-form complaints.
• Composition and outcome differences for AI-flagged complaints
  • Formal features: AI-flagged complaints are more citation-dense.
  • Filer composition: AI-flagged complaints are more likely to be associated with observed first-time filers in the linked sample; geographic concentration across circuits; modest, borderline-significant increase in name-inferred female plaintiffs among AI-flagged non-form complaints.
  • Outcomes: higher dismissal rate for AI-flagged (61.1% vs. 53.6%), more likely to terminate at earlier procedural phases, and no evidence of improved win rates.
• Robustness and limits of inference
  • Results robust to district and case-category controls and COVID-era controls; logistic regression post-AI OR ≈ 1.36 for plaintiff pro se status (95% CI ~[1.35,1.38]).
  • Design is descriptive/associational; the study does not observe individual AI use directly—AI-consistent drafting is a proxy; first-time and gender measures are name-inferred and limited to the sample.

Data & Methods

• Two-stage, nested design:
• Macro filing analysis using Federal Judicial Center (FJC) data (FY2008–FY2025). Pre/post cutoff set at November 30, 2022 (ChatGPT release). All-civil filings excluding prisoner petitions/unknown types (≈2.8M filings).
• Complaint-level analysis using pro se complaints from CourtListener (FY2018–FY2025), filtered and cleaned to 12,842 pro se complaints; linked subset matched back to FJC metadata for outcomes and filer characteristics.
• Focal categories: Civil Rights and Other Statutory (together ≈36.7% of all-civil filings) because they showed the largest pro se increases.
• AI-consistent drafting proxy:
  • Constructed from three textual/stylometric indicators drawn from existing literature and calibrated on the pre-GenAI baseline; the reported signal is the net post-AI increase above baseline.
  • Non-form complaints are the primary analytic focus because free-text drafting more readily reveals AI-consistent signatures.
• Outcome and composition measures:
  • Case trajectories (dismissals, termination phase), win rates.
  • First-time vs. repeat filer inferred from appearance of normalized primary-plaintiff names within the focal sample (not a full-history measure).
  • Name-inferred gender estimated via NamSor with a probability threshold ≥ 0.80 (binary classification).
• Identification and limitations:
  • Interrupted time-series, quarterly trajectories, and covariate-adjusted logistic regressions used to characterize trends and robustness.
  • Explicitly associational: cannot claim GenAI caused the changes; measurement error possible in AI-flagging, name-inference, and first-time indicators.

Implications for AI Economics

• Demand-side effects and access to justice
  • GenAI diffusion appears to lower some barriers to entry (more filings by self-represented plaintiffs), increasing demand for court access even if substantive remedies are not realized. Economically, this is a demand shock to the adjudication system originating from consumer-facing AI tools rather than from professional adoption.
  • Access expansion is uneven (by case category, geography, and observable filer characteristics), raising distributional concerns: AI-enabled entry may benefit some groups/regions more than others, potentially exacerbating inequalities in legal outcomes.
• Productivity vs. effectiveness trade-off
  • AI-consistent drafting is associated with more formally polished complaints (higher citation density) but worse procedural outcomes (higher dismissal rates, earlier termination). This suggests a wedge between formal productivity (ability to produce polished documents cheaply) and substantive effectiveness (successful litigation outcomes). From an economic-welfare perspective, marginally cheaper production of complaints may generate private utility (people can assert grievances) but also social costs (court congestion, screening burdens, wasted processing on claims unlikely to succeed).
• Market adjustments and legal labor
  • No clear evidence of broad lawyer displacement: represented filings did not uniformly decline, and in Civil Rights both represented and pro se filings rose. This points toward complementarity or market segmentation: GenAI may enable more lay filings while sustaining or even increasing demand for lawyers in some categories.
  • Potential market responses: growth in low-cost legal assistance (unbundled services), subscription/AI-assisted legal tools aimed at consumers, triage services, and entrepreneurs providing post‑filing support. Law firms may respond by offering assistance tailored to newly filing pro se litigants or by competing on higher-value tasks less amenable to public GenAI.
• Public-good and administrative externalities
  • Increased filings by pro se litigants—particularly those more likely to be dismissed—impose screening and administrative costs on courts. These are negative externalities not internalized by users of public GenAI, suggesting a role for policy or institution-level responses (e.g., triage, filing fees/waivers, form redesign, or AI-use disclosures).
  • The asymmetric diffusion and geographic concentration imply uneven local impacts on court workloads and legal markets; resource allocation and budgeting for courts may need to be responsive to these demand shifts.
• Research and policy priorities
  • Better measurement of individual AI use (surveys, disclosures) and causal identification are needed to move from association to mechanism.
  • Cost–benefit assessment: quantify social welfare effects of expanded formal access versus increased adjudication costs and potentially lower average case quality.
  • Consider regulatory responses: rules on AI-generated filings, AI-use transparency, court-level technological adaptations (automated triage, standardized templates), and targeted support for litigants whose filings are unlikely to succeed but who lack counsel.

Overall, the paper documents a meaningful institutional effect associated in time with public GenAI diffusion: more people are entering federal civil courts pro se, a nontrivial share of filings carry AI-consistent textual signatures, and these filings generate distinct downstream outcomes and administrative pressures. For AI economics, this is evidence that consumer-facing AI can shift demand in regulated professional markets, produce heterogeneous welfare effects, and trigger both market and public-sector responses.