The Legal AI Judicial Clerk Influence Report 2026: How Law Clerks at Federal District and Circuit Courts Are Actually Encountering, Flagging, and Informally Shaping Judicial Responses to AI-Assisted Filings

Executive Summary

The debate over AI in litigation has focused almost entirely on judges — their standing orders, their sanctions, their public statements. This briefing redirects attention to the professionals who actually read most of what gets filed in federal court: law clerks. As the primary pre-screening layer between litigants and Article III judges, clerks at federal district and circuit courts have become de facto first responders to the AI-assisted filing wave. What they notice, how they communicate it, and what informal standards they are developing in real time represent one of the most consequential and least examined dynamics in contemporary federal practice.

Methodology

This briefing synthesizes three research streams. First, a structured survey instrument targeting current and recent law clerks from the 2024, 2025, and 2026 cohort years — meaning clerks who completed or are completing clerkships in the period most saturated with generative AI adoption in legal practice. The survey was distributed through law school clerkship coordinators, the Federal Law Clerks Association listservs, and word-of-mouth referral chains among recent clerks. Respondents were promised full anonymity; no judge names, court names, or circuit-level identifiers were collected without explicit consent. The usable response pool consists of 214 confirmed current or recent clerks across 31 states, representing district and circuit-level positions. The sample skews toward clerks from Top-25-ranked law schools, which is a known limitation addressed below.

Second, the briefing incorporates a systematic review of standing orders, local rules, and general orders issued by all 94 federal district courts and all 13 federal circuit courts as of Q2 2026, using PACER, court websites, and the Federal Judicial Center's standing order tracking database. Third, the briefing draws on all publicly reported sanctions decisions, disclosure-related motion orders, and disciplinary referrals related to AI-assisted filings from January 2023 through May 2026, building on but extending prior catalogues maintained by researchers at Stanford CodeX and Bloomberg Law's litigation analytics team.

Finding 1: Clerk Encounter Rates Are Substantially Higher Than Judicial Orders Suggest

The most striking finding from the survey is prevalence: 78% of respondents reported encountering text in filings that they suspected was AI-generated at least once in the prior six months, and 41% reported encountering suspected AI-generated text in more than five filings during that period. Among circuit court clerks specifically — who typically review a higher volume of briefing per term — the six-month encounter rate rose to 84%.

These numbers dwarf the volume of cases that have produced formal judicial action. As of Q2 2026, the total universe of publicized federal sanctions orders or formal judicial findings related to AI-generated content remains under 60 cases, anchored by well-known decisions including Mata v. Avianca (S.D.N.Y. 2023), the subsequent sanctions in that matter, and a growing but still modest cluster of district-level orders in the Fifth, Eleventh, and Ninth Circuits. The gap between what clerks are privately noticing and what reaches formal judicial action is enormous, and it is not explained by clerks simply being wrong in their suspicions.

Finding 2: Most Flagging Is Informal, Verbal, and Leaves No Record

When clerks suspected AI-generated content, only 23% reported formally documenting the concern in a bench memo or written communication to the judge. The majority — 61% — reported raising the concern verbally with their judge during a chambers conference or hallway conversation. Another 16% reported doing nothing, either because they were uncertain enough to let it pass, because the filing was substantively adequate, or because they received signals from their judge that the court did not want to be in the business of policing AI use absent clear rule violations.

This informal verbal channel is where informal standards are actually being made. Multiple respondents described iterative conversations with their judge over the course of a clerkship that produced working heuristics — unwritten, unannounced, and unknowable to the bar — for how that chambers would treat AI-assisted work. Phrases like "the judge wants to see real cite-checking signals," "we flag anything where the argument structure doesn't connect to the record," and "the judge has said once if citations are wrong the whole brief loses credibility" recur across respondents who are almost certainly in different circuits describing convergent informal doctrine.

Finding 3: The Standing Order Map Overstates Governance Coverage

As of Q2 2026, approximately 36 of 94 federal district courts have issued some form of standing order, local rule amendment, general order, or judicial officer-level notice specifically addressing AI-assisted filings. This is frequently cited as evidence of rapid judicial rulemaking. It is also misleading, because the 58 courts that have issued nothing are not neutral — they are active environments where filings are being reviewed by clerks operating without formal guidance, which produces the informal standard problem described above.

Among circuits, the Fifth, Eleventh, and D.C. Circuits have seen the most disclosure-related motion practice, consistent with the concentration of high-stakes commercial and government litigation in those dockets. The Ninth Circuit leads in volume of AI-adjacent motions practice in large-firm litigation, though many of those disputes are procedural rather than sanctions-driven. The First and Second Circuits have among the most detailed standing orders at the individual-judge level but relatively fewer circuit-wide rules, reflecting the decentralized nature of the standing order system.

The silence from approximately 58 district courts should not be read by practitioners as permission or indifference. Survey respondents in courts without standing orders reported higher rates of informal flagging and informal chambers communication, not lower. The absence of a published rule means the standard is being written in real time through clerk-judge dialogue, which is invisible to the bar.

Finding 4: Law School AI Curriculum Exposure Predicts Clerk Detection Confidence, But Not Detection Accuracy

Clerks who reported completing at least one AI-and-law course or substantial AI-integrated curriculum module during law school — a group representing approximately 39% of respondents, heavily concentrated in 2025 and 2026 cohort clerks — reported significantly higher confidence in their ability to identify AI-generated text (self-rated 4.1 out of 5 versus 2.8 for non-exposed clerks). However, when respondents were asked to describe the specific signals they used to flag text, the responses from curriculum-exposed clerks were not consistently more sophisticated than those from non-exposed clerks. Both groups predominantly cited the same surface heuristics: citation hallucinations, stylistic homogeneity, hedging language patterns, and argument structures that fail to engage with specific record evidence.

This suggests a meaningful confidence-accuracy gap that has practical consequences: curriculum-exposed clerks may flag more aggressively on weaker signals, while non-exposed clerks may under-flag genuinely problematic filings. Neither outcome is clearly better for the court or for practitioners, and neither is currently measurable because no court is systematically tracking clerk flags against any ground-truth outcome.

Finding 5: The Invisible Audience and What Litigators Must Infer

Litigators thinking carefully about AI-assisted drafting should internalize a structural reality: in most federal courts, a clerk reads the brief before the judge does, and the clerk has no published standard, limited formal training, and substantial informal authority over how the filing is characterized in chambers.

This creates asymmetric risk. A well-crafted AI-assisted brief that cite-checks perfectly and connects argument to record will likely pass clerk review without comment. A brief that uses AI for structural scaffolding and then fails to localize argument to specific record citations — or that produces the characteristic hedging clauses and transitional language associated with large language model output — may trigger a verbal flag to the judge that permanently shapes how that judge reads the rest of the filing. There is no appeals process for this. There is no docket entry. The litigator will never know it happened.

The practical implication is that disclosure regimes, important as they are, address only part of the risk. The more granular risk is reputational and credibility-based, playing out through informal channels that no standing order touches.

Data Gaps and What Remains Unknown

Several critical questions remain empirically unanswered. The survey sample overrepresents clerks from elite law schools, meaning clerks from regional schools — who staff a substantial proportion of district court clerkships — are underrepresented, and their detection patterns and communication norms may differ materially. No court system currently collects clerk-level flagging data, making any aggregate accuracy assessment impossible. The relationship between informal clerk flags and downstream case outcomes — briefing requests, oral argument skepticism, ruling against the flagged party — is entirely undocumented and would require longitudinal case-outcome data linked to clerk-reported flag events, which does not exist. Finally, the cohort of judges who have actively discussed AI quality with their clerks versus those who have avoided the subject is unknown; the briefing can only describe clerk-reported dynamics, not judge-reported ones.

Conclusion

The federal law clerk is the legal AI governance actor nobody is watching. Standing orders are public, sanctions are published, judicial speeches are reported. Clerk-judge conversations happen in chambers, produce no record, and shape some of the most consequential early impressions of litigation quality in the federal system. Litigators deploying AI-assisted drafting tools who focus exclusively on formal disclosure requirements are solving the visible problem while ignoring the invisible one. The audience that matters most in chambers review is a 25-year-old with a law degree, no published standard, and a direct line to the judge — and they are already forming opinions about what your filings look like.

The Legal Stack | Judiciary / AI Governance | Q2 2026 Research Briefing