AI Is Already Inside the Federal Courts. Legal Purism Is Not a Governance Strategy

VII. The Training Deficit and the Myth of Responsible Informality

The survey’s training findings expose a problem deeper than uneven adoption. They reveal a judiciary in which use has begun before education has been systematized. According to the paper, 45.5% of responding judges said court administration had not provided training on AI tools, and another 15.7% said they were not sure whether such training had been offered. Among the judges who did report that training had been provided, most attended. That result matters for two reasons at once. It confirms that training remains underprovided, and it suggests that the obstacle is not simply judicial resistance. There appears to be demand for serious instruction. What is missing is a consistent institutional structure for delivering it. A court system that permits meaningful AI use without a training architecture is not practicing controlled modernization. It is improvising under the appearance of professionalism.

Improvisation is especially dangerous in legal institutions because legal culture tends to overestimate the transferability of general analytical skill. Judges and lawyers are trained readers. They are trained skeptics. They are trained source-checkers. From those strengths, it is easy to infer that they can safely experiment with AI so long as they remain cautious. But caution is not a substitute for instruction. It is not a curriculum, not a protocol, and not a method for distinguishing one class of risk from another. General professional competence does not tell a chambers user what kinds of hallucination are most common, how a legal-specific AI tool differs from a general-purpose model, when a summarization feature may flatten adversarial nuance, why a polished paraphrase may silently distort a holding, or how retrieval and prompt structure can affect output. Training addresses those questions. Without it, even highly capable professionals can mistake fluency for accuracy and convenience for competence.

The survey’s own qualitative material makes that danger visible. Judges expressed concern about hallucinations, “zombie cases,” and skill atrophy. One chambers anecdote described a clerk obtaining ten fake cases out of eleven cited in an AI-generated memo. That anecdote matters not because it proves widespread collapse, but because it captures the basic hazard of the technology. AI output often looks professionally usable before it has been professionally verified. It arrives wearing the surface markers of competence: structure, tone, fluency, and confidence. That is precisely why training cannot be treated as an optional supplement to otherwise sound professional judgment. The risk lies in the gap between polish and trustworthiness.

A court system that ignores that gap is likely to fall back on what might be called responsible informality. The idea is simple and superficially appealing: prudent, educated professionals can be trusted to explore cautiously, verify what matters, and develop good habits through experience. This model flatters the institution. It preserves autonomy, avoids bureaucracy, and allows actors to reassure themselves that professionalism alone will absorb the risk. But responsible informality is the wrong governance model for a technology whose attraction lies in reducing friction. AI saves time. It lowers the cost of producing first-pass work. It rewards convenience. In any setting shaped by workload and deadlines, convenience presses against scrutiny. The more a tool seems to accelerate familiar tasks, the greater the temptation to treat verification as selective rather than comprehensive. A training deficit therefore does not leave the institution in a neutral state. It creates the conditions in which convenience is more likely to outrun judgment.

That dynamic is especially acute in chambers. Judicial work is often deadline-driven, research-heavy, and delegation-dependent. Clerks and staff are under pressure to organize records, identify authorities, summarize briefs, and distill large volumes of material into workable form. AI is attractive in precisely those circumstances. It promises compression. It promises orientation. It promises speed. The danger is that users operating without formal instruction may fail to appreciate how easily machine-generated shortcuts can distort the very material they are trying to clarify. A summary may omit the adversarial caveat that changes the weight of a fact. A research-oriented output may surface plausible but secondary authorities while neglecting controlling law. A chronology generated from contested submissions may carry an unexamined narrative bias into later internal analysis. These are not spectacular failures in the sanctions sense. They are workflow distortions, and they are harder to detect precisely because they often appear helpful.

There is also a second category of training need that discussions of AI misuse often understate. The risk in courts is not limited to fabricated legal authority. It includes confidentiality, data handling, record integrity, and confusion about the line between assistance and delegation. A chambers user who pastes sensitive material into an unapproved system may create a problem entirely different from false citations. A clerk who uses AI to generate a chronology from contested facts may not fabricate a single authority and still distort the structure of the case. A judge who uses an AI tool to orient himself to an unfamiliar technical subject may come away with a subtly skewed understanding before ever turning to primary authority. None of these failures fits neatly into the headline category of “hallucinated cases,” but all of them implicate adjudicative integrity. Training is what allows users to see that these are not separate curiosities. They are part of the same institutional risk profile.

This is also why verification language in administrative guidance cannot do all the work by itself. Saying that users must independently review and verify AI-generated content is correct but incomplete. Verification is not a magic word. It is a practice, and competent practice has to be taught. What counts as verification of a legal proposition? What counts as verification of a factual summary? Is checking whether a cited case exists enough, or must the user also verify the quoted proposition, the procedural posture, the factual analogy, and the current validity of the authority? How should a chambers user verify a synthesized statement that compresses multiple authorities into one polished sentence? What about a chronology assembled from mixed record materials? Verification sounds straightforward only when the output is simple. In real workflow, the difficult cases are precisely the ones in which the AI produces something too smooth to force the user to notice what must still be checked.

Training also has an accountability dimension. Courts insist, correctly, that lawyers remain responsible for the filings they submit regardless of how drafts were generated. That principle appears in modern standing orders addressing AI-assisted submissions. But a judiciary that expects external actors to understand verification, candor, and accuracy obligations cannot rely on ad hoc internal learning for its own personnel. Otherwise the institution risks a posture of asymmetric rigor: warning lawyers about a technology the courts themselves have not yet systematically taught their own users to manage. That is not a sustainable position for an institution that claims authority through disciplined process.

The survey result that most judges attended training when it was offered is therefore more important than it first appears. It suggests that the present deficit should not be explained away as a natural byproduct of judicial conservatism. The problem may not be reluctance to learn so much as institutional underprovision. That matters because it changes the governance question. If the federal judiciary already has an audience for serious instruction, then the failure to provide it consistently becomes harder to justify. A judiciary that recognizes AI as significant enough to warrant an advisory task force and interim guidance should also recognize that the efficacy of any such guidance depends on whether users are taught how to operationalize it.

A serious training architecture would have to be judiciary-specific rather than vendor-specific. It would need to cover at least several distinct domains. It would need to address source verification and how to confirm not only the existence of authority but the accuracy of propositions attributed to it. It would need to address hallucination and fabrication risk, including the difference between confidently expressed error and transparent uncertainty. It would need to address confidentiality and data handling, especially for chambers personnel working with sensitive materials. It would need to address task-based boundaries so that users understand not only what they may do, but why certain uses are more dangerous than others. It would need to address chambers supervision and approval structures so that judges, clerks, and staff understand how responsibility travels through collaborative work. And it would need to address disclosure or certification obligations where AI influences material that reaches a court filing or an operative judicial product. A training program that simply repeats “be careful” or “always verify” does not address those institutional needs. It converts responsibility into a slogan.

There is a final point, and it is not merely pedagogical. Courts cannot preserve legitimacy by assuming that professionalism will absorb technological risk on its own. Professionalism is not self-executing. It depends on infrastructure. In this context, that means training. Without training, even a formally cautious judiciary remains vulnerable to the very kind of drift it claims to oppose. Legal purism often presents itself as a defense of standards. But standards that are not taught, specified, and operationalized are not standards in any meaningful sense. They are aspirations, and aspirations do not govern institutions.

VIII. The Purists Are Right About the Risks—and Wrong About the Response

The legal purists deserve a cleaner hearing than the more breathless advocates of legal technology often give them. Their core concern is not irrational, reactionary, or unserious. Courts are not ordinary workplaces. They are institutions that issue authoritative judgments in the name of law. A technology capable of inventing cases, misstating holdings, flattening factual nuance, compressing legal complexity into smooth error, and presenting all of it in the tone of professional competence is not a neutral efficiency aid in that setting. It is a risk multiplier. If the profession is uneasy, it has reason to be uneasy.

The survey itself confirms that the concern is grounded. Judges reported apprehension about hallucinations, “zombie cases,” and skill atrophy. The paper also includes the anecdote of a law clerk who used AI to generate a memo and found that ten of the eleven cited cases were fake. That is not a cosmetic defect. It cuts directly at the point where legal reasoning claims seriousness: fidelity to actual authority. A profession built on precedent, traceable source use, and articulated justification cannot casually absorb a tool that may produce convincing legal form without legal truth. The purists do not need to exaggerate the danger. The record already gives them enough.

The federal courts’ own conduct outside the survey underscores the same reality. Courts have already issued standing orders requiring disclosure of AI use in preparing filings and certification that citations were verified. They have warned that AI-assisted filings can contain fabricated or incorrect legal authority. Federal appellate courts have treated fabricated AI-generated citations as a live sanctions problem rather than a thought experiment. The question is no longer whether AI can corrupt legal work. It plainly can. The question is what follows from that fact.

That is where the purists fail. Their mistake is not in diagnosis. It is in remedy. They often speak as though the existence of serious risk itself proves that the proper response must be categorical resistance, moral suspicion, or professional abstention. That conclusion might once have carried more force—at the threshold moment when AI could still plausibly be treated as a technology outside judicial practice, one that might be stopped at the courthouse door by principled resistance. But that moment has passed. Once AI is already inside chambers, already shaping legal research and document review, already prompting standing orders, already generating sanctions disputes, and already forcing administrative and rulemaking responses, the problem is no longer whether the institution may someday confront AI. It is already confronting it. At that point, denunciation without design is not rigor. It is abdication.

The error is subtle because the purists are often right at the level of principle. They are right that judicial legitimacy depends on human accountability. They are right that citation to nonexistent authority is intolerable. They are right that a legal profession built on source-traceable reasoning cannot safely rely on systems that generate polished synthesis without guaranteeing actual source fidelity. They are right that convenience threatens discipline. But they are wrong to assume that the defense of those principles can be achieved by refusing to articulate a governing framework. Refusal to govern does not stop use. It pushes use into ambiguity, informality, and hidden dependence.

Courts already understand this in every other domain of legal risk. They do not respond to unreliable evidence by abolishing the concept of evidence. They create authentication standards, admissibility rules, burdens of proof, and review mechanisms. They do not respond to discovery abuse by abolishing discovery. They impose preservation duties, certification requirements, and sanctions. They do not confront attorney misconduct by lamenting professional decline in the abstract. They define obligations and enforce them. In other words, courts routinely confront danger not with nostalgia but with procedure. The law’s enduring strength has never been purity. It has been control.

That is why the federal judiciary’s emerging response matters. Interim guidance emphasizing independent verification, cautioning against delegation of core judicial functions, and urging task-based approval is not a concession to technological fashion. It is the beginning of what courts do when they recognize that a risk cannot be wished away. The movement toward governance is not evidence that the purists lost the moral argument. It is evidence that the institution has begun to understand a harder truth: uncontrolled use is a bigger threat than acknowledged and bounded use. The purists see the danger. What they do not offer is a viable method for controlling it once it is already present.

There is also an epistemic weakness in the purist position. It often treats AI as though the principal danger were explicit substitution of machine for lawyer or judge. But the more serious danger may be hidden mediation. A legal professional who knows AI is prohibited for overt drafting may still use it to orient research, organize a chronology, summarize a long brief, or test a line of inquiry informally. Those uses may feel too preliminary to count as real reliance. Yet they can shape what authorities are later pursued, what facts are later foregrounded, and what themes are later treated as central. A system of categorical denunciation does not address that problem because it does not identify where the lines are, how those lines will be supervised, or what obligations attach to staying within them. It condemns the visible breach while leaving the quieter mechanisms of institutional drift underdescribed.

That is why the purists’ position is best understood as incomplete rather than wholly wrong. They correctly perceive that AI presents doctrinal, professional, and legitimacy risks. But they stop at condemnation. That is the point at which they cease to be institutionally useful. The judiciary does not need additional moral theater about why fabricated citations are bad. It needs enforceable distinctions between acceptable assistance and unacceptable delegation, between permissible experimentation and careless contamination, between competent verification and reckless reliance. Fear may be justified. But fear alone does not build a governance architecture.

There is an irony here as well. The purists often present themselves as taking the harder line because they refuse compromise. In institutional terms, theirs may be the easier position. It is always simpler to declare that a dangerous technology should stay out than to do the more difficult work of defining task categories, supervision rules, verification duties, disclosure obligations, training requirements, and sanction structures. It is easier to denounce a system than to regulate one. But institutions are not judged by the purity of their rhetoric. They are judged by whether they can govern reality. On that measure, the purists have not offered enough.

The harder and more serious position is therefore not enthusiasm. It is governance. It is the insistence that if AI is already inside the federal courts, then the institution must say where it may be used, where it may not, who is accountable, how outputs must be checked, and what consequences follow when those rules are breached. That is not compromise with risk. It is the only available alternative to unmanaged drift.

IX. Federal Courts Are Already Building Patchwork Controls

The clearest proof that artificial intelligence has become a governance issue in the federal courts is no longer the survey alone. It is the fact that federal judicial actors have already begun building controls in multiple directions at once. Those controls are not unified. They do not cover the same terrain. They do not reflect a single settled theory. But taken together, they show something important: the federal courts have moved beyond generalized anxiety and into procedural response. That transition matters because institutions do not build controls around unreal problems. They build them when practice has advanced far enough that unmanaged discretion is no longer tolerable.

One of the earliest and clearest examples is Judge Michael Baylson’s June 2023 standing order in the Eastern District of Pennsylvania. Its logic is narrow, direct, and institutionally sound. Any attorney or pro se litigant who uses artificial intelligence in preparing a filing must disclose that fact, and the filing must certify that every citation to law or the record has been verified as accurate. The order does not attempt a broad theory of machine assistance. It does not spend time philosophizing about innovation or warning dramatically about existential danger. It targets the immediate problem a court can least afford to tolerate: submissions containing synthetic authority that no human has actually checked. In doing so, it ties AI use back to familiar professional duties. The lawyer remains responsible. The citation must be real. The record reference must be real. The filing cannot become a vehicle for machine-generated fiction under the cover of legal form.

That order is important not only because of what it requires, but because of what it assumes. It assumes that concealment is part of the problem. AI use “in any way” must be disclosed. That feature can be debated. Reasonable minds can differ about whether that scope is overbroad or whether disclosure should attach only to certain categories of use. But as an institutional move, it captures two enduring governance principles that remain difficult to improve upon: candor and verification. The order recognizes that the problem is not merely false output. It is false output entering adjudication without a clear line of accountability.

The District of Kansas’s January 2026 standing order is broader and, in some respects, more system-conscious. It opens by acknowledging that courts are seeing an increasing number of filings containing false statements of fact or law, including fabricated or incorrect legal authority, and that such filings may have been generated using AI. That framing matters because it ties the technology directly to judicial workload and procedural harm. The order does not treat AI misuse as a novelty embarrassment. It treats it as a threat to the integrity of proceedings and a source of waste for both courts and litigants. Most significantly, it states that lawyers and parties, including pro se litigants, remain responsible for the contents of their filings even when AI generates part or all of them. That is a critical governance move. It refuses the idea that technological mediation dilutes professional responsibility.

The Kansas order also goes beyond simple warning. It identifies a menu of possible judicial responses: striking filings, monetary sanctions, disciplinary referrals, disqualification, filing restrictions, and dismissal. It authorizes case-specific sworn statements requiring disclosure of the extent of AI use, the tool used, the sections generated, and verification efforts. This is not merely a cautionary memo against fake cases. It is an operational framework for policing AI-assisted filing practice through responsibility, verification, and remedial discretion. It does not settle every question, but it reflects a court attempting to move from generalized concern to enforceable process.

Appellate sanctions practice reinforces the same trend from a different direction. Once federal appellate courts begin sanctioning or formally addressing AI-related briefing failures, the issue changes character. It is no longer simply a local concern or an item of professional gossip. It becomes part of federal procedural seriousness. Sanctions at the appellate level communicate that AI misuse is not confined to inexperienced filers, fringe cases, or isolated trial-court embarrassments. It has reached a level at which federal appellate courts view fabricated citations and AI-assisted misrepresentation as conduct warranting formal judicial response. That development is significant because appellate sanctions help convert cultural alarm into legal consequence. They teach the bar that the problem is no longer rhetorical.

But sanctions, by their nature, remain reactive. They address breakdown after the defective filing has already been submitted, the court’s time has already been consumed, and the institutional cost has already been incurred. Sanctions matter. They are indispensable as a back-end control. But they are not governance in the full sense. They tell actors what happens when they fail. They do not fully tell chambers, clerks, judges, and litigants what the institution affirmatively permits, expects, supervises, or forbids before failure occurs. That distinction is crucial. A mature control structure requires more than punishment. It requires ex ante architecture.

That is where the federal judiciary’s administrative response enters the picture. Interim guidance produced through the Administrative Office’s AI Task Force adds a different layer of control. It is not principally a public-facing sanction regime. It is an internal administrative scaffold. It is aimed at shaping conduct before breakdown rather than merely punishing misconduct after the fact. By emphasizing independent verification, cautioning against delegation of core judicial functions, and encouraging courts to define approved tasks and consider disclosure, the guidance reflects a preventive rather than purely punitive mindset. In governance terms, that matters a great deal. A system cannot rely on sanctions alone unless it is content to let institutional learning occur mainly through failure.

The evidentiary rules process adds still another dimension. Proposed responses to machine-generated output and suspected deepfakes show that federal concern about AI is not limited to the drafting of filings or the internal operation of chambers. It extends to the admission, evaluation, and authentication of evidence itself. That broadening of focus is revealing. It means the judiciary is not treating AI as a single narrow problem. It is confronting AI across several fronts at once: filing practice, chambers use, internal administration, sanctions doctrine, and evidentiary integrity. That multidirectional response is itself evidence of institutional seriousness. It signals that the judiciary has begun to understand AI not as one issue, but as a cluster of related procedural and epistemic pressures.

And yet, despite all of this activity, the controls remain unmistakably patchwork. Judge-specific standing orders are not district-wide policy. District-wide standing orders are not national rules. Interim administrative guidance is not the same thing as binding Code-of-Conduct language or formal rule text. Proposed evidentiary responses do not answer the problems of chambers supervision or internal task definition. Appellate sanctions punish some categories of misconduct but do not define affirmative best practices for ordinary use. Each development is rational in its own domain. Each solves or attempts to solve a real problem. But together they still reveal the absence of a single settled governance architecture.

Patchwork has costs, and those costs are not merely aesthetic. It creates uneven expectations across districts. It makes litigant obligations contingent on local assignment. It leaves chambers personnel subject to differing degrees of supervision depending on where they work and under whom they serve. It risks generating the misleading impression that AI governance is mainly a problem for careless lawyers rather than a problem implicating the judiciary’s own internal workflow. Most importantly, patchwork invites adaptation through loopholes. Where one court demands disclosure, another may not. Where one chamber has a written policy, another may operate on informal assumptions. Where one district imposes strong sanctions, another may rely on ambient caution without procedural teeth. Institutions governed through uneven controls tend to produce uneven habits, and uneven habits are exactly what courts ordinarily resist in other contexts.

Still, patchwork should not be dismissed as though it were meaningless. It is often the normal form of early institutional response. The point is not that the federal courts have done nothing. The point is that they have already done enough to prove the issue is real, while not yet enough to produce convergence. That is what makes the present moment significant. The federal courts are no longer deciding whether AI requires procedure. They are deciding what kind of procedure, at what level, and under whose authority. That is a much more serious question, and it is the one the profession has to answer next.

X. From Patchwork to Principle: What Judicial AI Governance Should Require

The present landscape suggests that the federal judiciary is moving toward governance by accretion. Standing orders appear in individual courts. Sanctions emerge in response to defective briefing. Interim guidance develops at the administrative center. Committee work proceeds at the rulemaking level. Chamber-specific norms exist in scattered and uneven forms. That pattern is understandable in a transition period. It is not adequate as a destination. A judiciary that derives legitimacy from disciplined process should not remain indefinitely in a state where AI governance depends on a mixture of local innovation, professional instinct, and institutional improvisation. The question is therefore no longer whether some controls exist. They do. The question is what principles should organize them so that the system can move from reactive patchwork to coherent governance.

The first principle must be task distinction. Artificial intelligence cannot be governed intelligently if all uses are treated as though they carry the same institutional meaning. Legal research triage is not the same as opinion drafting. Administrative correspondence is not the same as summarizing a contested record. Preparing a speech outline is not the same as framing dispositive legal analysis. Chambers need a governance structure that begins with categories of function rather than abstract anxiety. Some tasks are peripheral enough that bounded use may be tolerable under verification rules. Some tasks are close enough to adjudicative judgment that they should be tightly cabined or prohibited. The point is not that every possible use can be reduced to a neat chart in advance. The point is that no serious institution should govern a consequential technology by pretending that all uses pose the same risk.

This principle matters because task confusion is one of the main ways informal dependence takes hold. If a chamber allows “AI for research” without defining what research includes, the label becomes elastic enough to swallow far more than intended. Does research include issue spotting? Summary of a factual record? Drafting a case comparison? Proposing an analytical framework? Organizing deposition excerpts into narrative order? Without task distinction, broad permission quietly becomes uncontrolled drift. Governance begins when the institution stops using large abstractions and starts specifying what conduct actually occurs inside workflow.

The second principle is verification as a non-delegable duty. Every serious federal response to AI so far converges on this point for good reason. Machine assistance may be tolerated in certain settings, but it cannot relieve the human actor of responsibility for source accuracy, factual correctness, procedural truthfulness, or fair representation of authority. Verification is not a best practice in this setting. It is the minimum condition of legitimacy. A court cannot claim fidelity to law if it permits machine-generated assertions to circulate through chambers or filings without disciplined human checking. That principle should be universal, not local. It should not depend on whether a judge happens to issue a standing order or whether a particular district has already encountered an embarrassing failure.

But verification has to be understood properly. It is not enough to check whether a case exists. Competent verification requires more than existence. It requires confirming that the authority says what the AI claims it says, that the quotation is accurate, that the procedural posture is correctly represented, that the case remains good law where that matters, and that the analogy being drawn has not been silently distorted in translation. The same is true for factual material. A polished summary of a record must be checked against the actual record, not merely accepted because it sounds balanced. Verification, in other words, is not a ritual word. It is a structured practice, and governance must treat it that way.

The third principle is supervisory accountability. Governance that focuses only on the individual end-user is inadequate because judicial work is distributed. Chambers are collaborative institutions. Clerks and staff may use AI in ways judges do not fully see, especially at the early stages of research and document handling. A judge therefore cannot meaningfully preserve human judicial accountability while allowing subordinate AI use to develop through private discretion and unwritten assumptions. Chambers need written protocols. Judges need to know whether, when, and how AI is being used by clerks and staff. Approved-task boundaries must be communicated and monitored. A governance regime that does not account for distributed judicial labor is not governing the institution that actually exists.

This is also where the old model of chamber culture becomes insufficient. Trust, informality, and professional instinct are often enough to manage ordinary workflow variation. They are not enough for a technology whose appeal lies precisely in how easily it can disappear into routine work. A clerk using AI to orient a research question or compress a record may believe the use is too preliminary to matter. A judge who does not ask may never know it happened. Supervisory accountability exists to close that gap. Not because clerks are untrustworthy, but because legitimacy cannot depend on invisible technological mediation inside a process that still presents itself as wholly traditional.

The fourth principle is role-based disclosure. Disclosure should not be performative, but neither should it be avoided simply because it is awkward or because the line is difficult to draw. The question is contextual: when does AI involvement become significant enough that candor to the court, the parties, or internal judicial supervisors is required? For litigant filings, the answer is clearest. If AI materially shaped a submission in ways that affect verification burdens, legal accuracy, or the court’s ability to rely on the filing, the court may reasonably require disclosure or certification. For chambers, the issue is more delicate but no less real. A judge may reasonably require that clerks disclose AI use upward within the chamber even if no public disclosure to the parties is required. The point is not maximal transparency at all times. The point is to identify the points at which undisclosed use creates accountability gaps the institution should not tolerate.

Role-based disclosure is also preferable to a crude universal rule because it respects the differences between internal and external use. Not every instance of AI assistance should trigger the same disclosure consequence. A clerk using a tool to draft a speech outline does not present the same concerns as a lawyer using AI to generate legal analysis for a filing. A judge using a tool to test wording in a nonadjudicative context does not raise the same institutional stakes as a chambers user summarizing a contested evidentiary record. Governance must therefore ask not simply whether AI was used, but by whom, for what, and with what effect on the integrity of the material later relied upon.

The fifth principle is mandatory training wherever use is permitted. This follows directly from the survey’s training deficit and from the judiciary’s own emphasis on verification. No court should authorize AI-assisted work without ensuring that users understand hallucination risk, fabricated authority, confidentiality constraints, task boundaries, and the difference between apparently polished output and actually reliable work. Training should be judiciary-specific, recurring, and role-sensitive. It should not be outsourced entirely to vendors, whose incentives do not fully align with institutional caution. And it should distinguish among tool types. The risks presented by AI-assisted legal research are not identical to those presented by transcription tools, document-summarization systems, or open-ended generative models. A serious training regime cannot speak in slogans. It must speak in functions.

Mandatory training also serves a legitimacy function. Courts cannot insist that lawyers remain responsible for AI-assisted submissions while treating their own internal training obligations as optional. If the judiciary expects external actors to understand the duties that survive technological mediation, it must build comparable understanding internally. Otherwise it will create the appearance of asymmetric rigor: a court system warning others about risks it has not systematically taught its own users to manage.

The sixth principle is calibrated enforcement. Courts should punish misconduct, not mere technological association. This distinction matters. An enforcement regime that punishes the label “AI use” without regard to function, transparency, or verification will either chill transparent low-risk uses or drive uses underground. Neither result is desirable. The proper targets of enforcement are fabricated authority, reckless verification failures, concealment where disclosure is required, misuse of sensitive data, task-boundary violations, and other conduct that undermines adjudicative integrity. Enforcement must be principled enough to deter abuse without encouraging secrecy as the easiest route to avoiding consequences.

That requires a mature understanding of deterrence. Institutions deter best when actors know not only that sanctions exist, but what conduct triggers them and why. A lawyer or chambers user who understands that the violation lies in fabricated authority or unverified factual representation is more likely to internalize the correct norm than one who hears only that “AI is dangerous.” Governance succeeds when it teaches actors to connect discipline to institutional principle rather than to the cultural stigma of a tool.

The seventh principle is convergence. Not total uniformity immediately, and not elimination of all local variation. But convergence toward a recognizable federal baseline. The federal courts should not allow the acceptable use of AI to depend too heavily on the happenstance of district assignment, chamber culture, or individual judicial preference. Some variation is inevitable and sometimes healthy. But baseline expectations concerning verification, supervisory responsibility, training, disclosure, and categories of prohibited delegation should not remain radically fragmented. Legitimacy suffers when institutions confronting the same technology operate under irreconcilably different assumptions about what professional responsibility requires.

Convergence matters for another reason as well. Patchwork governance encourages loophole thinking. Users learn that what is unacceptable in one setting may pass unnoticed in another. Institutions with sharply uneven control structures tend to breed sharply uneven habits. A judiciary that values equal process should hesitate before permitting such divergence to become the long-term condition of AI use. The goal is not an instant national code that answers every possible question. The goal is a baseline strong enough that local experimentation occurs within recognizable boundaries rather than in procedural free space.

This is the point at which the reader should resist two opposite temptations. One is the demand for total prohibition. The other is the demand for immediate normalization. Neither is realistic and neither is institutionally intelligent. The better path is principled containment: define the tasks, train the users, require verification, supervise chambers, demand candor where it matters, sanction actual misconduct, and move the system toward convergence. That is not capitulation to technology. It is the ordinary legal response to risk.

The law has never managed danger by purity alone. Its most durable achievement has been something more demanding: it translates principle into procedure. If artificial intelligence is already inside the federal courts, then the question is no longer whether the judiciary can preserve itself by refusing to think about it. The question is whether it can build a control structure worthy of adjudication.

XI. Legitimacy Requires Candor, Not Denial

At bottom, the dispute over AI in the federal courts is not mainly about efficiency. It is about legitimacy. Courts claim a special kind of authority. They do not merely resolve disputes. They purport to do so through reasoned judgment under law. That claim depends on more than correct outcomes. It depends on visible fidelity to process, source integrity, and accountable human agency. Any technology that threatens to blur those elements will naturally provoke anxiety. That anxiety is not technophobic. It is constitutional in spirit. A system that permits coercive state power to rest on synthetic error, hidden machine mediation, or unverifiable legal work risks weakening the very basis on which judicial authority is publicly accepted.

That is why the AI debate cannot be reduced to a dispute about convenience. The real issue is whether the federal courts can preserve confidence in adjudication while integrating a technology associated with fabrication, opacity, and uncertain control. Courts are not technology firms, consulting shops, or ordinary professional offices. They exercise power in the name of law. Their legitimacy therefore depends not only on what they decide, but on how they reach decision. If the process becomes opaque at the very points where authority, fact, and analysis are first assembled, the problem is not merely technical. It is institutional.

But legitimacy is not preserved by pretending that nothing operational has changed. Once artificial intelligence has entered judicial research, document review, filing practice, evidentiary debate, and administrative governance, the idea that courts can protect their authority through silence becomes untenable. Silence is not neutrality. It is institutional opacity. And opacity is itself corrosive when the subject is a technology widely associated with fabricated authority, hidden assistance, and uncertainty about who actually controlled the work. A court system that says little while the technology becomes more integrated does not appear restrained. It appears behind its own reality.

That is why recent federal judicial responses matter even beyond their immediate substance. Task forces, interim guidance, standing orders, sanctions, and rulemaking discussions all reflect a common institutional recognition: AI is no longer a hypothetical issue external to the judiciary. It is already a judicial problem. These measures may be partial and uneven, but they are nonetheless forms of candor. They acknowledge that the institution must respond. And that acknowledgment is itself important, because it aligns the judiciary’s public self-description more closely with its actual operating conditions.

Candor matters for at least three reasons.

First, candor aligns institutional rhetoric with institutional reality. Courts lose credibility when their public posture suggests an unchanged traditional process while their actual workflow has materially shifted. If chambers are using AI in research, summary, or drafting-adjacent tasks, and the judiciary continues to speak as though the problem exists only at the margins or only among careless lawyers, then the institution begins to describe itself inaccurately. That kind of mismatch is dangerous. Public confidence erodes not only when courts make mistakes, but when courts appear unwilling to describe honestly the conditions under which they now work.

Second, candor allows governance to develop before repeated breakdown forces it into existence. Institutions that suppress discussion of operational change often end up regulating only after visible scandal. That is the worst way to govern. It produces reactive policy, moral panic, uneven punishment, and rushed line-drawing under pressure. A candid institution does something different. It identifies the issue before the next failure. It states the risks openly. It defines responsibilities in advance. It tells judges, clerks, lawyers, and litigants what standards will apply before a fabricated citation, distorted summary, or undisclosed use creates a legitimacy event. Candor is therefore not merely transparency for its own sake. It is a condition of timely supervision.

Third, candor enables accountability. A court cannot assess whether standards are being met if it refuses to say when AI use matters, who is responsible for checking it, what uses are acceptable, and what counts as failure. Accountability requires visible norms. Without them, error can always be treated as an isolated lapse rather than the predictable product of underdefined practice. Candor supplies the baseline against which conduct can be judged. It makes responsibility traceable rather than atmospheric.

The alternative is denial, and denial is not a neutral resting state. In institutional life, denial usually produces one of two outcomes. Either the practice continues informally under conditions of uneven supervision, or the practice becomes visible only through scandal when a particularly glaring failure forces acknowledgment. Both outcomes are worse than open governance. Informal continuation breeds drift. It allows actors to adapt quietly, without clear limits, until the institution’s actual practice is farther along than its formal rules admit. Scandal-driven acknowledgment is no better. It invites overcorrection, symbolic punishment, and improvised restrictions that may be more performative than principled. Neither path is consistent with a judiciary committed to reasoned and evenhanded administration.

There is also a democratic dimension to candor that should not be understated. The public is entitled to know, at least in broad terms, how the courts are confronting a technology that can affect filings, evidence, judicial workflow, and the production of authoritative legal texts. This does not mean every internal use of a tool must become a public spectacle. It does mean that the judiciary should not rely on mystique. In modern institutions, trust no longer survives on ceremony alone. It depends increasingly on intelligible safeguards. If courts expect lawyers and litigants to disclose or verify AI-assisted work where appropriate, it is hardly unreasonable to expect the judiciary to articulate its own baseline principles for internal use and control.

This is where the legitimacy question becomes sharper than many purists allow. Yes, artificial intelligence poses risks to reasoned human judgment. But unmanaged or undisclosed AI use poses a second and perhaps greater risk: that the judiciary will appear either unable or unwilling to govern a technology already affecting its work. In that circumstance, the problem is no longer just machine unreliability. It is institutional evasiveness. A judiciary that warns others about technological risk while speaking too softly about its own encounter with that risk invites a different kind of doubt—the doubt that it understands the demands of its own legitimacy.

This is also why legal purism, standing alone, underperforms as a legitimacy theory. It treats judicial authority as something protected by refusal alone. But refusal cannot protect authority once the underlying practice persists. If AI continues to enter chambers, filings, and evidentiary disputes despite rhetorical resistance, then legitimacy no longer depends on whether the institution disapproves. It depends on whether the institution governs honestly. An institution that refuses to name the problem after the problem has already arrived does not preserve integrity. It forfeits control.

The question, then, is no longer whether the federal courts can preserve legitimacy by keeping AI outside the system entirely. That possibility has already been overtaken by events. The real question is whether the judiciary will preserve legitimacy by facing the technology directly—by stating where it is present, where it is permitted, where it is prohibited, who is accountable, and how the integrity of adjudication will be protected despite its presence. That is the work candor makes possible. Without candor, legitimacy becomes an unsupported claim. With candor, it at least has a chance to remain a governed one.

XII. Conclusion: Denial Is Not Discipline

The debate over artificial intelligence in the federal courts has moved beyond its opening stage. The issue is no longer whether AI will matter to federal adjudication. It already does. It is present in judicial work. It is present in chambers. It is present in legal research, document handling, filing practice, sanctions disputes, and the judiciary’s own administrative planning. The question is no longer whether the federal courts may someday confront AI as an institutional problem. They are confronting it now.

What makes the present moment significant is not that the issue has been resolved. It plainly has not. The judiciary is in a transitional condition rather than a settled one. A substantial number of judges still report no use of listed AI tools. Training remains uneven. Formal policy is often absent, incomplete, or chamber-specific. Some courts have begun building procedural controls. Others are still operating through local improvisation, verbal custom, or interim guidance. The national rulemaking process is moving, but it is not complete. The federal courts have therefore entered the AI era without yet completing the work of governing it. That is not unusual in institutional life. Practice often arrives before principle. But in courts the lag matters more, because courts claim legitimacy through disciplined method, not through retrospective rationalization.

The legal purists are right about the stakes. Artificial intelligence can fabricate, distort, flatten, and deceive. It can tempt busy professionals to mistake polished output for reliable analysis. It can encourage a style of legal work in which fluent synthesis arrives before disciplined engagement with actual sources. In adjudicative systems, those are not ordinary efficiency tradeoffs. They are legitimacy risks. A court that relies, even indirectly, on synthetic error does not merely make a technical mistake. It compromises the very conditions under which its judgments claim authority.

But the purists remain wrong about the response. Denial cannot solve a problem already inside the institution. It cannot train judges, clerks, or staff. It cannot define task boundaries. It cannot protect confidentiality. It cannot tell chambers when AI use becomes significant enough to require supervision or disclosure. It cannot distinguish between acceptable assistance and impermissible delegation. It cannot create a verification culture. It cannot build a rule structure. It can only leave those questions unresolved while actual use continues. That is not institutional rigor. It is a refusal to govern.

The proper response is therefore governance. Not rhetorical panic. Not abstract technological enthusiasm. Governance. That means written policies rather than ambient expectations. It means task-based distinctions rather than undifferentiated slogans. It means verification as a non-delegable duty. It means supervisory accountability inside chambers. It means role-sensitive disclosure where candor is necessary. It means judiciary-specific training rather than vague appeals to caution. It means sanctions for actual misconduct rather than symbolic hostility to software labels. It means movement toward baseline federal expectations even if some local variation remains unavoidable.

This is not a concession to technology. It is the ordinary legal response to risk. Law has never preserved its seriousness by merely repeating its values in louder language. It has preserved seriousness by translating values into procedure. When evidence is unreliable, the law builds admissibility rules. When lawyers overreach, the law imposes duties and sanctions. When institutions confront a new source of distortion, the law’s task is not to sentimentalize older methods. It is to decide what structures are now necessary to preserve the integrity of judgment. AI is no exception.

That is why the question before the federal courts is now narrower and harder than many commentators admit. The question is not whether the judiciary can restore a world in which this technology remained wholly outside the courthouse doors. That moment has passed. The question is whether the judiciary can now govern a technology already in its midst with enough seriousness, candor, and procedural discipline to preserve confidence in adjudication. If it can, then AI will become one more risk managed through law’s familiar tools of supervision, verification, and accountability. If it cannot, then the problem will not be technology alone. The problem will be that the courts failed to impose on themselves the same rigor they demand from everyone else.

Artificial intelligence is already inside the federal courts. That is not prophecy. It is the present tense. The real measure of the judiciary will not be whether it managed to keep AI entirely outside the system. The real measure will be whether it can meet the technology with the one thing courts are supposed to know how to provide: disciplined control. Denial is not discipline. In this setting, only governance is.