Modern justice systems still ration meaningful legal participation through historical scarcities in legal labor and judicial processing capacity. This paper argues that artificial intelligence makes some of those scarcities technologically unnecessary, and that courts should therefore distinguish the human judgment that justice requires from the human clerical, analytical, and procedural labor that justice has historically depended upon only because no alternative existed. The claim is not that AI makes courts faster; it is that AI changes which forms of legal scarcity are actually necessary.
The paper separates two distinct failures: a court-capacity problem (backlogs in small claims, administrative law, and criminal dockets) and a representation problem it names the Commercial Viability Filter — the mechanism by which a legally cognizable claim disappears not because a court found it meritless, but because counsel found it commercially unattractive to pursue. It proposes a governing doctrine, Human Sovereignty / Machine Accessibility, under which AI may expand the ability to understand, organize, and present a legal position, and may assist judicial actors in analyzing it, while sovereign coercive authority remains human and institutionally accountable.
Within that doctrine, the paper develops five bounded mechanisms: AI-assisted small-claims intake and triage; AI-assisted self-representation in civil and administrative proceedings, with a deliberately narrow, force-multiplier-only role in criminal indigent defense; a Justice Preservation Filing that tolls a statute of limitations when a plausible right cannot be presented for want of counsel, modeled on the existing American Pipe tolling doctrine and explicitly bounded by its limits; an AI-Assisted Putative Class Preservation Petition that preserves a potential systemic claim without ever permitting an unqualified filer to represent absent class members; and a Judicial Dispute Mapping Standard that asks AI to continuously identify what remains disputed in a case rather than to predict who should prevail. It proposes federal and state rule and statutory changes, including a model Unauthorized Practice of Law safe harbor, and closes with a safeguards framework against what it names Manufactured Adjudication — collusive, sham, or preclusion-engineering uses of the same tools, addressed through an Adverseness and Authenticity Gate that flags, but never itself adjudicates, suspected abuse.
Two distinct failures are commonly discussed as one problem. The first is a capacity failure: courts cannot process the volume of matters before them quickly enough. The second is a representation failure: many legally cognizable claims and defenses never receive competent presentation at all. They compound each other, but they are not the same failure, and conflating them produces reforms that speed up processing without closing the representation gap, or that expand representation without relieving backlog.
The capacity failure is well documented at every level of the system this paper examines. Social Security disability hearings had roughly 360,000–361,000 cases pending as of mid-2026, up from about 274,000–280,000 a year earlier, even as the average wait per case improved modestly to just under nine months — a pattern in which faster individual processing and a larger backlog are simultaneously true, because faster upstream stages simply funnel more claims into the hearing queue than administrative law judges can clear [1]. The federal immigration courts had 3,141,306 active cases pending as of the end of July 2026 by the Transactional Records Access Clearinghouse's count, of which 2,293,984 were already-filed asylum applications awaiting hearings or decisions [2]. Indigent criminal defense operates under caseloads that a 2023 national workload study — conducted by RAND, the National Center for State Courts, and the American Bar Association's Standing Committee on Legal Aid and Indigent Defense — found routinely exceed defensible standards by multiples: public defenders in St. Clair County, Missouri handled roughly 350 felony cases per lawyer in a single year, and Luzerne County, Pennsylvania public defenders handled more than 300, against a newly recommended standard of 35 hours of attorney time per felony case [3][4].
The representation failure is separate and, this paper argues, less well understood. The Legal Services Corporation's 2022 Justice Gap Study — still the most recent full edition as of this writing, with a more comprehensive 2027 study currently in the field — found that low-income Americans received no or inadequate legal help for 92% of the civil legal problems that substantially affected them, that 74% of low-income households experienced at least one civil legal problem in the prior year, and that cost concerns (46%) and doubt about finding an affordable lawyer (53%) were the leading reasons people did not seek help at all [5]. Stanford Law research cited by the American Bar Association found at least one party lacks representation in three-quarters of civil cases nationally — roughly 15 million cases a year [6]. In immigration court, TRAC's own data shows only 22.8% of immigrants had an attorney at the point a removal order was issued in July 2026 [2]. A 2024 LSC-commissioned Harris Poll found more than half of Americans mistakenly believe they are entitled to a free lawyer in any civil matter — evidence that the representation gap is largely invisible to the public it affects [7].
Neither figure above fully captures a third, distinct mechanism this paper treats as a central contribution rather than a footnote. A legally cognizable claim can disappear not because a court determined it lacks merit, and not only because the claimant could not afford any lawyer, but because the lawyers who could take it performed a cost-benefit analysis and declined. Contingency-fee economics, statutory fee caps, litigation risk, and reputational considerations routinely screen out claims that are legally sound but commercially unattractive — low individual damages, a well-resourced or sympathetic defendant, an evidentiary picture that will require expensive development before it is provable. The claim is not adjudicated and rejected; it is never brought, and the system records nothing, because nothing was filed.
The Commercial Viability Filter: the point at which a private attorney's economic triage — not a court's merits determination — silently decides whether a legally cognizable claim proceeds at all. The justice system effectively outsources justiciability screening to profit motive, and the result is described, inaccurately, as "the market working."
This is the paper's organizing diagnostic. Sections 6 through 8 propose mechanisms aimed specifically at this filter — not at making lawyers faster, but at preventing claims from dying before anyone with authority to judge them ever sees them.
The representation failure interacts with ordinary procedural machinery to produce a further, sharper problem: a substantive right can expire for a purely procedural reason that has nothing to do with its merits. A person may have a plausible claim, be unable to secure counsel — because of cost, because the claim is commercially unattractive under the Commercial Viability Filter, or because the relevant proceeding requires counsel by rule — and watch a limitations period or a procedural deadline run out while the search for representation continues. The claim does not lose on the facts or the law. It loses to a clock that was never designed to track whether representation was actually obtainable.
Existing law already recognizes a version of this problem and has already built one tool to solve it, in a narrower context: class-action tolling. In American Pipe & Construction Co. v. Utah, 414 U.S. 538 (1974), the Supreme Court held that the filing of a class action tolls the statute of limitations for all putative class members, treating the class complaint as a "pre-filing" of every claim it covers and putting the defendant on notice of the claims that may later be asserted individually [8]. Crown, Cork & Seal Co. v. Parker, 462 U.S. 345 (1983), extended that protection to separately filed individual suits, not only motions to intervene [9]. The doctrine's purpose, in the Supreme Court's own words, is to spare absent class members from having to file protective individual suits merely to avoid losing their claims while a class proceeding is pending.
That is precisely the logic this paper argues should be extended, deliberately and narrowly, to the person who has no class action to rely on at all — the person whose only problem is that they cannot find or afford counsel before their clock runs out. Section 6 develops this extension as the Justice Preservation Filing.
Every mechanism this paper proposes is checked against a single governing line, stated here once so that later sections can be read against it rather than re-argued each time.
Machines may expand the ability to understand, organize, research, prepare, verify, and present a legal position, and may assist judicial actors in analyzing that position. Sovereign coercive authority remains human and institutionally accountable.
Concretely: AI can organize evidence, identify precedent, expose inconsistencies between a party's filings, calculate deadlines, maintain a continuously updated map of what is and is not disputed in a case, recommend questions for a human decision-maker to ask, and flag likely procedural defects before they become fatal. AI should not, acting autonomously, imprison a person, terminate parental rights, impose criminal punishment, hold a person in contempt, issue an irreversible final judgment, secretly determine a witness's credibility, or generate an undisclosed risk score that determines a person's liberty.
This is not a novel proposition in the abstract — it restates, for the litigant-facing context, a distinction the federal judiciary itself is already drawing for judge-facing AI use. The Administrative Office of the U.S. Courts distributed interim AI guidance to federal courts on July 31, 2025, explicitly framed as allowing "experimentation with guardrails" while preserving judicial independence and the integrity of the process, pending more permanent guidance [10]. A Northwestern University survey of federal judges, published through the Sedona Conference in March 2026, found more than 60% of responding judges already use at least one AI tool in chambers, mostly for legal research and document review — while nearly half reported receiving no formal training on it [11]. The Judicial Conference's Committee on Rules of Practice and Procedure has separately proposed Federal Rule of Evidence 707, which would require a reliability hearing before AI-generated evidence can be offered at trial without a corresponding human expert [11]. The judiciary, in other words, is already drawing exactly the accessibility/sovereignty line this paper proposes to extend outward to litigants — this paper's contribution is to state that line as an explicit doctrine and apply it consistently to every proposal that follows, rather than leave it implicit and case-by-case.
Small claims is the least constitutionally fraught venue for AI assistance, because self-representation is already the procedural norm there, not an exception requiring justification.
This model extends, rather than invents, existing online dispute resolution (ODR) infrastructure already operating in several state small-claims systems and in British Columbia's Civil Resolution Tribunal — the extension is substituting large-language-model reasoning for the rule-based, form-driven logic those systems currently use, under the same human sign-off requirements those systems already impose.
Document review, discovery-response drafting, motion-practice templates keyed to jurisdiction-specific rules, and real-time plain-language explanation of procedural posture — assistance functionally similar to what a paralegal or staff attorney already provides, made available to a pro se litigant who has neither.
Social Security disability, unemployment insurance, and immigration proceedings combine the highest pro se rates in the system with some of the best-documented representation effects on outcome, and already tolerate lower procedural formality than an Article III courtroom — the natural first venue for AI-assisted self-representation to be tested at scale, ahead of ordinary civil litigation.
The Sixth Amendment guarantees the assistance of counsel, not merely access to legal information, and this paper does not argue that distinction away. It resolves the criminal proposal narrowly: AI as a public-defender force multiplier, addressing a caseload crisis that is independently and rigorously documented (§8), not as any form of self-representation substitute. No proposal in this paper authorizes AI to replace appointed counsel in a matter where liberty is at stake, at any phase of deployment (§14). This restraint is stated here explicitly rather than left implicit, because the civil and administrative proposals in this section rest on much easier constitutional ground, and blurring the two would weaken both.
Section 2 identified the structural failure this proposal answers: a plausible right expiring because the person asserting it could not obtain the professional intermediary required to present it, not because a court ever passed on the merits.
A substantive legal right should not expire solely because the person asserting it was unable to obtain the professional intermediary required to present it.
Mechanism: a verified, structured filing — made through a court-supervised AI-assisted intake system — establishing enough information to identify the claim, the defendant, the injury, and the relevant time period. Where the matter appears potentially cognizable but cannot yet proceed without counsel or further factual development, the filing tolls the applicable statute of limitations for a defined, capped period.
This is not invented from nothing. American Pipe already establishes that a structured filing short of a fully litigation-ready complaint can suspend a limitations period on the theory that it constitutes adequate notice to the defendant of the claims that may follow [8]. That is the load-bearing analogy. But the doctrine's own subsequent limits define exactly how far this proposal can go without inviting the objections that have already been raised — and largely accepted — against overextending it:
A defendant has a due-process interest in adequate notice that they are being pursued. American Pipe answers this for the class-action version of tolling because the class filing already discloses the general contours of the claims that may follow. The Justice Preservation Filing needs an equivalent answer: the filing must specify enough — claim type, alleged injury, defendant, and time period — that the notice rationale actually transfers from the class-action context to the individual one. This specificity requirement is not a formality; it is the entire basis on which a court or legislature could accept this as an incremental extension of recognized doctrine rather than reject it as inventing a new, unbounded substantive right. No existing pilot or program implementing this exact mechanism was found in the research for this paper; the case for it rests on the justice-gap data in §1 and the doctrinal analogy above, not on outcome data from an operating system, and the paper states that candidly rather than implying otherwise.
Because a low-threshold preservation filing is also the proposal most exposed to bad-faith use — a placeholder filing is a cheap, high-leverage move for a party trying to manufacture a limitations advantage — this mechanism should never be deployed without the Adverseness and Authenticity Gate described in §13, which is written with this proposal specifically in mind.
Some case types are reserved to licensed counsel by rule, most prominently class actions under Federal Rule of Civil Procedure 23. This paper does not propose routing around that reservation. Rule 23's adequacy-of-representation and fiduciary requirements exist to protect absent class members from inadequate representation, and an AI-equipped pro se litigant is not a substitute for counsel bound by those fiduciary duties. The problem worth solving is narrower and different: a systemic wrong can go entirely unaddressed not because Rule 23 correctly requires counsel to litigate a class action, but because no counsel was willing to take on the cost of finding out whether a class action was even viable in the first place — the Commercial Viability Filter operating at class scale.
This proposal does not certify a class and does not authorize an individual to represent absent class members at any point. It allows a person alleging a systemic wrong to submit a structured petition — proposed class characteristics, common alleged conduct, known affected people or transactions, common legal questions, estimated damages, documentary evidence, and limitations concerns — for a court to evaluate whether the allegation merits preservation or preliminary development.
Possible court-ordered outcomes: dismissal; permission for limited discovery; referral to legal aid; solicitation or appointment of qualified counsel; interim counsel; consolidation with related petitions; or temporary tolling of the underlying claims pending one of the above.
Attorney representation may remain necessary to control a class action, without attorney availability becoming a prerequisite to preserving the existence of the alleged class claim.
No direct doctrinal analogue for this exact mechanism was found in this paper's research — it should be presented candidly as novel, modeled on American Pipe's policy logic (claims should not die for want of a procedural vehicle) rather than on its doctrine, which governs a different situation. The mechanism's entire claim to legitimacy rests on one repeated, non-negotiable constraint: it must create no preclusive effect on anyone but the petitioner, and it must never itself certify a class or bind an absent member. Every description of this mechanism in any future drafting, filing, or advocacy built from this paper should restate that constraint explicitly — it is the whole basis on which this proposal avoids the objection Rule 23 exists to prevent.
The same logic — lowering the cost of preserving a claim without lowering the bar for who may litigate it — extends to other categories a state reserves to counsel by rule, without requiring that the underlying reservation itself be relaxed.
Two separate populations of counsel benefit from the same underlying capability: private attorneys deciding whether a marginal case clears the Commercial Viability Filter, and public defenders operating under caseloads that are independently documented as unconstitutional in practice.
The 2023 National Public Defense Workload Study — RAND, the National Center for State Courts, and the ABA's Standing Committee on Legal Aid and Indigent Defense, replacing the 1973 National Advisory Commission standards that never anticipated body-camera footage, cell-phone records, or social-media evidence — recommends 35 hours of attorney time per felony case and 22.3 hours per misdemeanor [3]. Against that standard, the same study documents public defenders in St. Clair County, Missouri handling roughly 350 felony cases per lawyer in a year, and Luzerne County, Pennsylvania handling more than 300 — both more than double the outdated 1973 guideline, and far beyond the 2023 standard [4]. An Oregon appellate court has already vacated a lower-court order requiring an overloaded public defender to take on a case, a live judicial recognition that excessive caseloads can void the practical adequacy of appointed counsel.
AI-drafted complaints, damages models, discovery review, and — for indigent defense specifically — AI-assisted review of voluminous digital evidence directly targets the caseload arithmetic above. For private counsel, the same tools lower the marginal cost of taking a case, which is the direct, targeted answer to the Commercial Viability Filter named in §1: fewer meritorious claims are screened out by cost-benefit math when the cost side of that calculation falls.
Every developed case eventually has an implicit structured object: which facts are admitted, denied, or partially admitted; which legal issues are genuinely disputed; what evidence supports each side on each disputed point. Judges currently reconstruct this picture from pleadings, exhibits, discovery responses, motions, oppositions, supplemental filings, and transcripts — a reconstruction task that consumes real judicial time on every case, independent of its ultimate difficulty.
AI can maintain this structure continuously as a case develops, rather than requiring it to be reconstructed at each decision point. The reframing this produces is deliberate: the tool is never asked "who should win?" It is asked, narrower and safer, "what exactly remains for the judge to decide?" That narrower question keeps this proposal cleanly on the accessibility side of the §3 doctrine — it organizes and displays what the parties have already put in the record; it does not weigh that record or recommend an outcome.
This is the same underlying capability as §4's small-claims triage, applied court-side rather than litigant-side, and extends naturally to precedent verification and docket management once the dispute-object infrastructure exists.
Unauthorized Practice of Law (UPL) statutes are the central state-level obstacle to nearly every proposal in this paper. As currently written in most states, a court-supervised AI assistance tool used by a pro se litigant would likely be treated as unauthorized practice if a nonlawyer entity is deemed to be "practicing" — regardless of how carefully the tool is scoped or supervised.
Utah's Supreme Court launched a seven-year legal-services regulatory sandbox in 2020, authorized through August 14, 2027, permitting nontraditional providers — including software- and AI-based services and, initially, nonlawyer-owned firms — to operate under direct court supervision. By 2023, nearly 50 entities had been authorized, collectively delivering more than 40,000 legal services to roughly 24,000 people at an average cost of about $160 per service [15]. That is the version of the Utah story most often cited, and it is incomplete.
In late 2024 and through 2025, the Utah Supreme Court substantially narrowed the sandbox — closing the alternative-business-structure-only track at the end of 2024, discontinuing new applications from for-profit immigration-services entities, and imposing a new "Utah Innovation Requirement" that entities must demonstrate they reach consumers currently underserved by the legal market [16]. As of March 2025, roughly 27 entities faced possibly leaving the program under the tightened rules, with several relocating to Arizona's comparable program; a Utah advocacy organization is now pushing legislation to restore broader participation [17].
The lesson for this paper's UPL proposal is not that Utah's model failed — it is that a successful pilot remains one docket order away from reversal absent legislative or rule-based permanence. A model UPL safe harbor for court-supervised AI tools should therefore build durability into its own design from the outset — a statutory or constitutionally-anchored rule, not solely a revocable supreme court standing order — rather than treat initial authorization as the finish line.
Collusive suits, sham litigation, and manufactured preclusion predate AI by centuries. What AI changes is the cost and scalability of attempting them, not their basic legal character. This paper names the broader category deliberately, rather than treating it as an undifferentiated "AI abuse" concern:
Manufactured Adjudication: attempts to manufacture favorable defaults, collusive settlements, weak adversaries, sham lawsuits, preclusive judgments, manufactured standing, fabricated evidence, fraudulent class settlements, strategic losses intended to impair later litigants, or — where legally possible — sham criminal proceedings engineered to obtain an acquittal that forecloses a later, legitimate prosecution.
Existing law already gives courts inherent authority to vacate a judgment obtained through fraud on the court, without any time limit, under the saving-clause authority preserved in Federal Rule of Civil Procedure 60(b) and its state analogues [18]. But this paper's own research found that the doctrinal bar for "fraud on the court" is set deliberately high: courts describe it as requiring "the most egregious misconduct directed to the court itself," such as bribery of a judge or jury, or fabrication of evidence by counsel [19]. A merely weak or quietly collusive adversarial process — two parties agreeing not to contest a claim seriously, without anything rising to bribery or outright fabrication — likely does not meet that bar on its own.
This is not a reason to abandon the safeguard; it is the reason this paper treats a new layer as necessary rather than redundant. Existing doctrine is a strong tool for the most extreme cases and a weak one for the subtler collusive scenario — including a self-directed filing engineered to obtain a preclusive result, such as a defendant arranging or encouraging a weak prosecution against themselves in the hope that an acquittal will bar a later, legitimate one. That gap is precisely what the mechanism below is built to close.
Before any AI-assisted proceeding can produce an unusually significant preclusive consequence — a result that would bind absent parties, toll or extinguish a limitations period under §6, or carry res judicata or collateral estoppel effect — the system flags, but does not itself decide, indicators including: shared addresses, ownership, or counsel between nominally adverse parties; suspicious filing or settlement timing; extraordinarily weak or absent opposition; coordinated filings across matters; undisclosed relationships between parties; and anomalous settlement terms. A human court determines whether collusion in fact exists. The Gate's entire function is detection at scale — exactly the kind of large-caseload pattern-matching task AI is well suited to and courts currently cannot perform systematically — never adjudication.
This produces the paper's most balanced conclusion, stated directly rather than left implicit: AI can simultaneously create new abuse capacity and make previously undetectable abuse patterns easier for courts to discover. Both are true at once, and neither is the paper's final word on the other.
The strongest single objection to the paper as a whole is that its five mechanisms sit at very different levels of doctrinal readiness — from Section 4's small-claims triage, which extends operating infrastructure already in use, to Section 7's Class Preservation Petition, which has no operating precedent at all. The paper's phased-deployment sequencing (§14) is a direct response to that unevenness: it deploys the best-grounded mechanisms first and treats the least-grounded ones as pilots requiring their own evidence before wider adoption is proposed.
The paper's governing thesis, restated directly: modern justice systems still ration meaningful legal participation through historical scarcities in legal labor and judicial processing capacity. Artificial intelligence makes some of those scarcities technologically unnecessary. Courts should therefore distinguish the human judgment that justice requires from the human clerical, analytical, and procedural labor that justice has historically depended upon only because no alternative existed.
This is not the claim that AI makes courts faster. It is the claim that AI changes which forms of legal scarcity are actually necessary — and that a system which keeps rationing access on those grounds after the technical justification for doing so has weakened is choosing an outcome, not following an unavoidable constraint.
The right to petition a court is of limited practical value when understanding the law, navigating procedure, preserving a claim, and competently presenting the facts all depend on purchasing scarce professional labor. Artificial intelligence does not eliminate the need for lawyers or judges. It does, however, challenge the assumption that access to legal competence must remain scarce simply because it historically was.
Every mechanism this paper proposes is bounded by the same line: AI may expand the ability to understand, organize, and present a legal position; sovereign coercive authority remains human. Within that boundary, the paper argues there is real, currently unexploited room to close the Commercial Viability Filter, to stop procedural clocks from silently converting representation scarcity into substantive denial, and to give courts tools for detecting the abuse this same technology could otherwise make easier — without asking any machine to decide who is free, who is guilty, or who wins.
[1] Social Security Administration disability hearing backlog data, as reported in: "SSDI Wait Times in 2026: Faster Decisions, But a Growing Hearing Backlog Expected in 2027," Saving to Invest, 2026; and "Average Wait Time for a Social Security Disability Hearing in 2026," DisabilitySecrets, 2026. Figures cited (roughly 274,000–280,000 pending mid-2025 rising to roughly 360,000–361,000 pending mid-2026; average hearing wait ~267–274 days) are drawn from these secondary aggregations of SSA data and should be verified against SSA's own published performance and budget reports before final citation.
[2] Transactional Records Access Clearinghouse (TRAC), Syracuse University, "Immigration Court Quick Facts," tracreports.org/immigration/quickfacts/eoir.html (data as of end of July 2026): 3,141,306 active cases pending; 2,293,984 already-filed asylum applications pending hearing or decision; 22.8% of immigrants represented by counsel at removal order in July 2026.
[3] Pace, N. M., Brink, M. N., Lee, C. G., & Hanlon, S. F. (2023). National Public Defense Workload Study. RAND Corporation, RR-A2559-1, conducted with the National Center for State Courts and the American Bar Association Standing Committee on Legal Aid and Indigent Defense. https://www.rand.org/pubs/research_reports/RRA2559-1.html
[4] "Groundbreaking Report Finds Many Public Defenders Are Dangerously Overworked," Arnold Ventures, September 2023 (citing St. Clair County, Missouri and Luzerne County, Pennsylvania caseload figures from the National Public Defense Workload Study); "In 'watershed moment,' report recommends new guidelines for public defender caseloads," ABA Journal, September 13, 2023.
[5] Legal Services Corporation (2022). The Justice Gap: The Unmet Civil Legal Needs of Low-Income Americans. justicegap.lsc.gov. 92% of substantially-impactful civil legal problems received no or inadequate help; 74% of low-income households experienced at least one civil legal problem in the prior year; 46% cited cost as a reason for not seeking help; 53% doubted they could find an affordable lawyer.
[6] "Millions of Americans continue to lack meaningful access to justice. What can be done about it?", American Bar Association, Washington Letter, April 2026 (citing Stanford Law School research on representation rates in civil cases nationally).
[7] Legal Services Corporation, 50th Annual Report (2024), citing a 2024 LSC-commissioned Harris Poll on public understanding of the civil justice gap. lsc.gov/50th-annual-report
[8] American Pipe & Construction Co. v. Utah, 414 U.S. 538 (1974).
[9] Crown, Cork & Seal Co. v. Parker, 462 U.S. 345 (1983).
[10] "Interim AI guidance for US courts aims for experimentation with guardrails," FedScoop, October 2025, reporting on the Administrative Office of the U.S. Courts' interim AI guidance distributed to federal courts July 31, 2025, per correspondence from AO Director Judge Robert J. Conrad to Sen. Chuck Grassley.
[11] "AI & the Courts: Recent Developments," American Bar Association, Washington Letter, April 2026, reporting the Sedona Conference/Northwestern University survey of federal judges (March 2026; 502 judges surveyed; over 60% use at least one AI tool in chambers) and the status of proposed Federal Rule of Evidence 707 (public comment closed February 16, 2026; Advisory Committee on Evidence Rules vote scheduled May 7, 2026).
[12] California Public Employees' Retirement System v. ANZ Securities, Inc., 137 S. Ct. 2042 (2017).
[13] China Agritech, Inc. v. Resh, 584 U.S. ___, 138 S. Ct. 1800 (2018).
[14] Sen. Chuck Grassley, "Grassley Calls on the Federal Judiciary to Formally Regulate AI Use," remarks and press release, U.S. Senate, October 2025. grassley.senate.gov
[15] "Millions of Americans continue to lack meaningful access to justice. What can be done about it?", American Bar Association, Washington Letter, April 2026 (Utah Sandbox Phase 1 figures: ~50 entities authorized by 2023; 40,000+ services delivered to ~24,000 people; average cost ~$160/service).
[16] Utah Office of Legal Services Innovation, "Sandbox Phase 2," utahinnovationoffice.org/sandbox-phase-2/ (narrowed eligibility, Utah Innovation Requirement, closure of ABS-only track effective end of 2024, discontinuation of for-profit immigration-services applications); Utah State Bar, "Utah Legal Services Innovation Office," utahbar.org/legal-services-innovation/ (Sandbox authorized through August 14, 2027).
[17] "Nearly 30 legal entities may leave Utah's regulatory sandbox program after state tightens rules," ABA Journal, March 2025; "Increase Access to Affordable Legal Services," Libertas Institute, 2025 (proposed legislation to restore broader Utah sandbox participation).
[18] Fed. R. Civ. P. 60(b), (d)(3); Richmond, D. R., "Critical Contours of Fraud on the Court," SSRN, 2018 (no time limit for vacating a judgment for fraud on the court, distinct from the one-year limit under Rule 60(b)(3)).
[19] Landscape Props., Inc. v. Vogel, 46 F.3d 1416, 1422 (8th Cir. 1995) ("fraud on the court" requires "the most egregious misconduct directed to the court itself, such as bribery of a judge or jury or fabrication of evidence by counsel"), as cited in U.S. Supreme Court certiorari-stage filings collecting circuit authority on the standard.
[20] "AI Hallucination Cases: The 1,598-Case Sanctions Tracker," HAQQ, June 2026, reporting on the Damien Charlotin AI hallucination case database (~1,600 documented court proceedings worldwide as of June 2026, most in the United States).
[21] Mata v. Avianca, Inc., S.D.N.Y. 2023 (Schwartz/LoDuca sanctions, $5,000); Couvrette v. Wisnovsky-related 2025–2026 sanctions reporting ($95,000–$110,000 range across sources), as reported in "Beyond the Mirage: Beware of Generative AI and Hallucinations," New York State Bar Association, June 2026, and "AI Hallucination Sanctions: A Lawyer's Guide for 2026," July 2026.
[22] American Bar Association, Formal Opinion 512 (July 2024), on the application of the duties of competence, confidentiality, and candor to generative AI use by attorneys, as referenced in secondary 2026 legal-ethics reporting; primary opinion text should be cited directly in any final draft.